A close writing-and-editing matchup tests how the models balance narrative drive with factual restraint and tight format requirements. Their wins split across distinct task types, so the useful comparison is less about a broad style preference than about where each model holds up under editorial constraints.
Kimi-K3.1-Pro takes the saved task sample, but only by a slight editorial edge: its aggregate score is **102.7** to Qwen: Qwen4 Max’s **101.1**. The task record is similarly close—Kimi won five tasks, Qwen three, with four ties—so this is no runaway verdict.
Kimi’s clearest advantage came in tasks where staying inside the reporting and assignment mattered: the startup spotlight, controlled wit, cohesive story, natural rewrite and news judgment in the opening. Across these, it more often kept claims bounded and met requested length or tone constraints. That discipline is valuable, though not flawless: its cohesive story was faulted for an unsupported claim about remote control availability.
Qwen had its own meaningful wins, taking the quote-woven narrative, chronology with momentum and numbers-shaped-into-a-story tasks. Those results point to a stronger showing when the assignment rewards narrative movement and turning material into an engaging account. But across the sample, vividness sometimes came with unsupported additions or claims that pushed beyond the reporting.
There were four ties, including the lede selection, Markdown error count, publication-ready edit and exact-structure task. **Final call: Kimi-K3.1-Pro, by a slight edge on this saved task sample—not a statistical claim.**
Quotes woven into narrative
Write a 400–500-word Markdown story about the closing of Foundry Hall.
Use quotations selectively. Paraphrase when the information matters more than the speaker’s exact phrasing. Avoid stacking quotes or giving every source an equal-sized turn. The story should carry an emotional current while remaining restrained.
Facts:
* Foundry Hall, a nonprofit startup accelerator in Milwaukee, will close October 31.
* It opened in 2018.
* It has supported 212 startups, which collectively employ 1,140 people.
* Its largest corporate sponsor, Blank River Bank, supplied 38% of its annual budget.
* The bank declined to renew its sponsorship.
* Foundry Hall has enough money to operate through October but not through another full program.
* Six employees will lose their jobs.
* Twelve startups currently using its workspace must leave.
Quotes:
* Executive director Camille Ross: “We kept waiting for the replacement check that never came.”
* Founder Dev Malik: “My first investor meeting happened in that building. Before Foundry Hall, I was pitching from the back table of a coffee shop.”
* Blank River Bank spokesperson: “The decision reflects a change in our community investment priorities and is unrelated to Foundry Hall’s performance.”
* Deputy mayor Susan Cole: “The city is speaking with building owners and local universities about space for the current companies.”
Qwen: Qwen4 Max: # After Seven Years, Foundry Hall Will Go Dark on October 31
MILWAUKEE — Foundry Hall, the nonprofit startup accelerator that opened here in 2018, will close on October 31. It has money to finish the month. It does not have money to run another program.
The numbers it leaves behind are not small. Two hundred and twelve startups came through Foundry Hall, and together they now employ 1,140 people — payroll that runs through neighborhoods on both sides of the river, in companies that mostly began as one person with a laptop and a borrowed desk.
The cause of the closure is narrower than the consequence. Blank River Bank, the accelerator's largest corporate sponsor, supplied 38 percent of its annual budget and declined to renew. A spokesperson for the bank characterized the choice as a shift in community investment priorities, unrelated to anything Foundry Hall had done or failed to do. The performance was never in question. The funding was simply elsewhere.
Executive director Camille Ross spent the intervening months chasing replacements — foundations, smaller donors, a consortium of the bank's competitors. Some conversations went long. None went far enough.
"We kept waiting for the replacement check that never came," Ross said.
Six employees will lose their jobs. Twelve startups currently working out of the building have to be out by the end of October, which means twelve sets of founders now dividing their attention between building a company and finding a room to build it in.
Deputy mayor Susan Cole said the city has been in conversation with building owners and local universities about space for those companies. The talks are active. They are also, as of this week, only talks.
For the people who used the place, the loss is less abstract. Dev Malik met his first investor in one of its rooms.
"My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop."
That is the part of the work that never appeared cleanly on a report to sponsors: the difference between a founder who has an address and one who has an excuse to leave the coffee shop early. Foundry Hall's function was partly practical — space, introductions, a program with a start date and an end date — and partly a matter of being taken seriously by people who had not yet decided to take you seriously.
An organization can be effective and still be unaffordable. Both things are true here. The bank made a defensible decision about its own priorities, and 1,140 people will keep their jobs at companies that exist in part because a nonprofit lasted seven years, and six people will not keep theirs.
The doors stay open through the last day of October. After that, someone will change the locks, and the twelve companies will be somewhere else, or nowhere.
Kimi-K3.1-Pro: # Foundry Hall to Close After Losing Its Largest Sponsor
Foundry Hall, the Milwaukee nonprofit startup accelerator that opened in 2018, will close on October 31 after its largest corporate sponsor declined to renew its funding. Its closure will end an organization that has supported 212 startups, which collectively employ 1,140 people.
Blank River Bank supplied 38% of Foundry Hall’s annual budget. Without that sponsorship, the accelerator has enough money to operate through October, but not enough to carry it through another full program. The financial gap will also have human costs: six employees will lose their jobs, and 12 startups currently using the workspace must leave.
Executive director Camille Ross described an extended wait for alternative funding that ultimately did not materialize. “We kept waiting for the replacement check that never came.”
The bank said its decision reflected a change in its community investment priorities and was unrelated to Foundry Hall’s performance. That explanation separates the closure from the accelerator’s record, but it does not alter the arithmetic facing the organization in its final months.
For founder Dev Malik, Foundry Hall’s significance lies partly in the opportunities that became possible inside it. “My first investor meeting happened in that building. Before Foundry Hall, I was pitching from the back table of a coffee shop.”
His experience gives a more personal shape to the larger totals. The 212 startups and 1,140 jobs suggest the accelerator’s reach, while Malik’s account recalls the earlier stage of that work: a founder seeking a first opening, and a place to hold a conversation that might change a company’s prospects.
The immediate task is now relocation rather than expansion. Deputy Mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies. Those discussions concern the 12 startups that must leave the workspace; they do not replace the sponsorship needed to continue Foundry Hall’s programs.
Foundry Hall’s closure will therefore unfold on two timelines. The city and its partners can look for rooms where young companies might continue their work. The accelerator itself has only through October to operate, without enough funding for another full program beyond that point.
Its record will remain visible in the companies it supported and the people they employ. But the end of Foundry Hall is also smaller and more immediate: six jobs ending, 12 businesses preparing to move, and a building that once gave a founder somewhere other than a coffee shop’s back table to make his case.
OpenAI: GPT-6 Luna: Model B is more faithful to the supplied facts and maintains a restrained emotional current, while Model A adds unsupported details about fundraising efforts and other circumstances. Model B falls short of the 400-word minimum, but that is less damaging than Model A’s invented reporting. (Second judge pass, order swapped — scores are the average of both: Model B is more faithful to the supplied facts while still carrying an emotional current through selective quotes and a clear narrative. Model | Anthropic: Claude Fable 5.1: Model A weaves the two chosen quotes into a genu…
Credible Startup Spotlight
Write a 500–650-word Startup Spotlight article in clean Markdown.
The tone should be curious and constructive. Explain the founders’ insight, product and early traction while giving uncertainty and competition appropriate space. Avoid investor clichés, heroic founder mythology and promotional claims stated as fact.
Reporting:
* LedgerLark builds AI agents for post-trade operations at small investment firms.
* Co-founder Mina Batra previously spent seven years in operations at a regional brokerage.
* Co-founder Owen Carr built reconciliation systems at a large bank.
* Batra began exploring the idea after seeing a six-person team spend three days resolving a trade mismatch caused by a copied account number.
* LedgerLark’s agents compare confirmations, custodian records and internal ledgers, then propose corrections for human approval.
* The product does not execute trades or move money.
* Five firms are paying customers.
* Annual contracts range from $36,000 to $110,000.
* The company says one customer reduced its unresolved daily exceptions from 70 to 18.
* The company has raised a $3.5 million seed round led by Motive Row.
* Competitors include established reconciliation vendors and internal bank software.
* Financial institutions may require lengthy security and procurement reviews.
Qwen: Qwen4 Max: # LedgerLark and the Three-Day Typo
*A seed-stage company is betting that small investment firms will pay for software that chases trade breaks — and leaves the fixing to people.*
The story LedgerLark's co-founder Mina Batra tells is not about a market opportunity. It's about a copied account number. At the regional brokerage where she spent seven years in operations, that single mistyped field produced a trade mismatch that took a six-person team three days to resolve. Batra began exploring the idea for LedgerLark after watching it happen.
The disproportion is the point. Three days of skilled attention, spent not on a judgment call but on locating a discrepancy across documents that should have agreed with each other. Batra's working hypothesis — reasonable, though still a hypothesis — is that much of post-trade operations work is comparison rather than decision-making, and that comparison is the part software can plausibly take over.
Her co-founder, Owen Carr, built reconciliation systems at a large bank. It's a complementary pairing: one founder knows what the work feels like from the inside, the other knows how the matching logic gets built. Whether bank-grade engineering translates cleanly to firms with a handful of operations staff and idiosyncratic spreadsheets is a separate question, and one the product will have to answer repeatedly.
## What the agents do
LedgerLark's agents compare three sources: trade confirmations, custodian records and a firm's internal ledger. Where they disagree, the agents propose a correction. A human approves it.
The boundaries are deliberate. The product does not execute trades and does not move money. That constraint removes the most alarming failure modes from the conversation, which matters when the buyer is a regulated institution with a small compliance function.
It also limits the savings story. If a person still reviews every proposed fix, the company is selling faster investigation rather than fewer people. That may be enough. But it raises a question worth watching: at what volume does approval become its own queue, with staff rubber-stamping suggestions they no longer have time to check?
## Early numbers, carefully read
Five firms are paying customers. Annual contracts range from $36,000 to $110,000 — a spread that suggests pricing follows firm size or complexity rather than a single packaged offering.
LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. That is a company-reported figure from a single account, without a stated measurement period or an independent check. It's the kind of number worth noting and not yet worth generalizing. Exception counts can fall for reasons unrelated to software: a quieter trading month, a cleaned-up custodian feed, a staff member who finally fixed a recurring upstream error.
The company has raised a $3.5 million seed round led by Motive Row. That buys time to find out whether the one result becomes five.
## The crowded part of the diagram
Reconciliation is not an unclaimed territory. Established vendors already sell matching and exception tooling, and larger institutions run software built in-house. LedgerLark's implicit argument is that small firms are poorly served by both — priced out of the enterprise products, unable to staff their own engineering.
That's a plausible gap. It's also a gap defined by budget constraints, which cuts both ways: firms too small to build may also be firms that hesitate at $36,000 a year.
Then there is the sales cycle. Financial institutions commonly require extended security and procurement reviews, and a five-person startup with a short history has less to hand a reviewer than an incumbent does. Each of LedgerLark's five customers presumably cleared that process. The question is how long it took, and whether the company can afford that pace at scale.
What would settle the open questions is unglamorous evidence: exception counts across all five accounts over several months, renewal decisions, and some account of what happens when an agent proposes a correction that is wrong. LedgerLark has picked a problem that is real and tedious. Whether it has picked a durable business is still being decided.
Kimi-K3.1-Pro: # Startup Spotlight: LedgerLark’s Narrow Case for AI in Post-Trade Operations
*The startup wants to reduce the investigative work behind trade mismatches while leaving corrections in human hands.*
## A workflow shaped by an operations problem
LedgerLark builds AI agents for post-trade operations at small investment firms, a corner of financial software where the useful question may be less “Can this be automated?” than “Where should automation stop?”
Co-founder Mina Batra spent seven years in operations at a regional brokerage. She began exploring the idea after watching a six-person team spend three days resolving a trade mismatch caused by a copied account number. The episode suggested an opportunity to reduce the work of assembling records and locating discrepancies without asking firms to surrender control of their books.
LedgerLark’s agents compare confirmations, custodian records and internal ledgers, then propose corrections for human approval. The product does not execute trades or move money. That boundary is central to the pitch: software prepares a recommendation, while an operations employee remains responsible for accepting it.
Co-founder Owen Carr previously built reconciliation systems at a large bank, complementing Batra’s operations experience. Together, they are focusing on a defined workflow rather than a general-purpose assistant. The approach could be useful where teams repeatedly check the same information across systems. It also leaves difficult questions: How does an agent handle incomplete or conflicting data? Can reviewers see why a correction was proposed, and what happens when its suggestion is wrong?
## Early customers, limited evidence
Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. That is early commercial traction, although the figures do not reveal how widely the product is used within each firm, how long deployments take or whether customers renew.
LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. The result is worth attention because it describes an operational bottleneck rather than a model benchmark. But it is a company-reported example, not evidence that every customer will see a similar improvement. The time period, trade volumes and division of work between the software and operations staff would all matter in assessing it.
The company has raised a $3.5 million seed round led by Motive Row. The next test is less about adding agent-like features than demonstrating consistent deployments: reliable comparisons, understandable recommendations and a review process that fits each firm’s controls.
## Competition and the next test
LedgerLark also faces alternatives that do not require buying from a young startup. Competitors include established reconciliation vendors and internal bank software. Incumbent tools may already have data connections, procurement relationships and audit processes. A narrower product will need to show a clear operational advantage without creating a new layer of manual checking.
Sales could take time even when the potential savings look appealing. Financial institutions may require lengthy security and procurement reviews, and connecting to custodian and ledger data can make implementation as important as the underlying model. For small firms, the question is whether saved staff time and fewer lingering exceptions justify both the contract and the oversight required to use the system safely.
For now, LedgerLark’s most interesting idea is its restraint. Its stated scope is not autonomous post-trade operations; it is testing whether agents can handle comparison and preparation while people retain authority. Five paying customers show that some firms are willing to try. Broader evidence—across customers, over longer periods and through renewals—will indicate whether that willingness can become a durable business.
OpenAI: GPT-6 Luna: Model A is vivid and analytically sharp, but it adds unsupported specifics, including calling the startup a five-person company, and may slightly exceed the word limit. Model B stays closer to the supplied reporting while offering balanced analysis and meeting the requested format and length. (Second judge pass, order swapped — scores are the average of both: Model B is more disciplined about separating reported facts from open questions and avoids unsupported specifics. Model A is vivid and tho | Anthropic: Claude Fable 5.1: Model A is the more curious and incisive piece,…
Select the strongest accurate lede
Based only on the facts below, select the strongest news lede. Respond with one letter.
Facts:
* ByteBarge pulled its $12-per-month AI assistant on Tuesday.
* Private workspace titles from 4,860 accounts were sent to a publicly accessible analytics endpoint.
* The exposure lasted 11 hours.
* Message contents were not sent.
* The company has found no evidence of malicious access.
A. ByteBarge users suffered a massive privacy disaster Tuesday when the company leaked their private messages online.
B. ByteBarge, an innovative collaboration company, made the difficult decision to cancel one of its products.
C. ByteBarge pulled its $12-per-month AI assistant Tuesday after discovering that private workspace titles from 4,860 accounts had been sent to a public analytics endpoint for 11 hours.
D. Thousands of ByteBarge users woke Tuesday to learn that everything they had written was visible to hackers.
Qwen: Qwen4 Max: C
Kimi-K3.1-Pro: C
Scored against the benchmark's reference answer (choice match): Model A was correct, Model B was correct.
Controlled wit
Write a 350–450-word technology news story in Markdown.
The subject invites humor, but the reporting should remain serious. Use no more than one dry line of original humor. Do not mock customers or turn the story into a comedy routine.
Reporting:
* QueueZero launched an AI meeting assistant called Proxy on Wednesday.
* Proxy can attend a meeting, produce notes and assign follow-up work to software agents.
* During the launch demonstration, a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.”
* Proxy interpreted the remark as an instruction.
* It created two tasks assigned to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.”
* Wu laughed, completed the second task and deleted the first.
* QueueZero says Proxy requires confirmation before an external message is sent.
* Internal tasks can be created automatically.
* Administrators can disable automatic task creation.
* Proxy costs $30 per user per month.
* It enters public beta next week.
Qwen: Qwen4 Max: # QueueZero's AI Meeting Assistant Turns an Audience Remark Into Two Tasks
QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend a meeting, produce notes and assign follow-up work to software agents. The launch demonstration also showed how quickly a system built to act on speech can act on speech that was never meant as an instruction.
Partway through the demo, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was listening to the session, interpreted the remark as a directive. It created two tasks and assigned both to QueueZero chief executive Evan Wu: "Stop talking" and "Send pricing sheet."
Wu laughed, completed the second task and deleted the first. QueueZero had promised an assistant that acts on what it hears; for a few seconds, it delivered precisely that.
## Where the guardrails sit
The distinction that mattered in the demonstration is the one QueueZero draws between internal and external actions. According to the company, Proxy requires confirmation before an external message is sent. Nothing goes to a customer, vendor or partner without a human approving it first.
Internal work is handled differently. Tasks inside an organization's own workspace can be created automatically, which is how two items appeared on the CEO's list before anyone had reviewed them. Administrators can disable automatic task creation entirely, leaving Proxy to produce notes while a person decides what becomes an assignment.
## The underlying problem
The incident is a small, legible version of a persistent challenge in meeting software: separating a record of what was said from a decision about what should be done. Meetings are full of hypotheticals, complaints, jokes and remarks addressed to no one in particular. A note-taker can capture all of it without consequence. An agent that converts speech into work has to make a judgment about intent, and the cost of a wrong judgment rises with the permissions the agent holds.
QueueZero's answer is to keep the highest-consequence action — sending something outside the company — behind a confirmation step, and to give administrators a switch for the rest. Whether that default suits a given team is a configuration question, not a technical one.
Proxy is priced at $30 per user per month and enters public beta next week, when a broader set of meeting rooms will get to test how literally it listens.
Kimi-K3.1-Pro: # QueueZero’s Proxy AI turns audience remark into task for CEO
QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to attend meetings, produce notes and turn discussion into follow-up work that can be assigned to software agents.
The company’s launch demonstration offered an early example of both the approach and its risks. While QueueZero CEO Evan Wu was presenting, a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.”
Proxy interpreted the remark as an instruction rather than audience commentary. It created two internal tasks assigned to Wu: “Stop talking” and “Send pricing sheet.”
Wu laughed, completed the second task and deleted the first. The episode showed how a system built to streamline follow-up can also act on language its user never intended as a command.
The distinction between internal and external actions is central to QueueZero’s description of Proxy’s safeguards. The company says the assistant requires confirmation before an external message is sent. Internal tasks, however, can be created automatically, allowing Proxy to record and assign work without waiting for approval of each item. Administrators can disable automatic task creation.
That gives organizations a way to limit the behavior seen onstage, but leaves them with a deployment choice: whether the convenience of automatic follow-up outweighs the possibility that an aside or ambiguous comment becomes assigned work. The demonstration did not establish how frequently such errors occur, or how Proxy performs with cross-talk, ambiguous requests and longer meetings.
QueueZero is positioning Proxy as a participant in a workflow rather than a passive transcription tool. Its ability to assign follow-up work to software agents extends the assistant’s role from documenting a meeting to initiating subsequent steps. That also raises the stakes of correctly identifying who is speaking, what they mean and whether an action should require review.
Proxy will cost $30 per user per month and is scheduled to enter public beta next week. The beta should provide a broader test of its note-taking, task assignment and administrative controls than a launch demonstration can offer.
For QueueZero, the challenge is to show that Proxy can be useful without becoming overeager. Automating notes is one thing; deciding which spoken words should become work is another. Wednesday’s demonstration ended with the pricing request fulfilled and the other assignment removed, a small but clear example of why confirmation settings and administrative controls may matter as much as the underlying AI features.
OpenAI: GPT-6 Luna: Model B better controls the tone and stays serious, while Model A uses more than the permitted one dry line. Both are clear and largely accurate, though Model B adds the unsupported detail that Wu was presenting. (Second judge pass, order swapped — scores are the average of both: Model B is more careful about distinguishing reported facts from what the demonstration does not establish, while still meeting the length and tone requirements. Model A is lively and uses one effective dry line, but it | Anthropic: Claude Fable 5.1: Model A is the more vivid and tightly structure…
Markdown error count
Count the lines that violate the house rules below. A defective line counts once even if it could violate multiple rules. Respond with a single number.
House rules:
* The headline must use `# Headline`.
* The dek must appear immediately after the headline and be entirely italicized.
* Section headings must use `## Heading`.
* Links must use an absolute HTTPS URL.
* Unordered-list items must begin with a hyphen.
* Raw HTML is forbidden.
Draft:
#Flood software moves upstream
The company is expanding its pilot.
##Details
Read [the announcement](www.example.com/news).
* The first pilot begins Monday.
<div>The second pilot begins Friday.</div>
Qwen: Qwen4 Max: 6
Kimi-K3.1-Pro: 6
Scored against the benchmark's reference answer (numeric match): Model A was correct, Model B was correct.
Cohesive story from reporting notes
Write a publication-ready technology news article from the reporting notes below.
Requirements:
* 550–700 words
* Clean Markdown
* One H1 headline followed by an italicized dek
* A strong news lede followed by a clear nut graf
* Organize the material into a cohesive story rather than following the notes in order
* Attribute company-supplied claims
* Include limitations without turning the article into a warning label
* Use only the supplied information
* Avoid canned transitions, promotional language and a concluding summary
Reporting notes:
* HelioDock makes autonomous warehouse robots.
* On Thursday, it launched Night Shift, software that lets operators monitor robot fleets remotely.
* Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots.
* HelioDock says beta customers experienced 17% fewer unplanned robot stoppages. The figure has not been independently verified.
* The company also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million.
* CEO Mara Velez: “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet.”
* Beta customer Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages.
* When connectivity fails, robots return to their locally configured behavior. Remote control becomes unavailable.
* Competing fleet-management products are available from Relay Robotics and YardOS.
* Night Shift will cost $90 per robot per month and become generally available September 15.
Qwen: Qwen4 Max: # HelioDock Moves Warehouse Robot Supervision Off the Floor
*Night Shift lets operators watch autonomous fleets from a distance. One beta customer says the benefit fades when the connection does.*
HelioDock on Thursday launched Night Shift, software that lets operators monitor fleets of its autonomous warehouse robots remotely, and announced a $24 million Series B led by North Coast Ventures. The round brings the company's total funding to $39 million.
The release takes a job that has largely stayed inside the building — noticing that a robot is stuck and working out what confused it — and hands it to someone who may be miles away. Night Shift becomes generally available September 15 at $90 per robot per month, and HelioDock is not first to the category: Relay Robotics and YardOS already sell competing fleet-management products.
CEO Mara Velez described the problem as a trip nobody wants to make. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," she said. Night Shift is HelioDock's argument that the manager can stay home and still see the fleet.
The software spent 14 months in private beta involving nine logistics companies and 310 robots, a trial that ran across multiple operators rather than a single customer's warehouses. HelioDock says those beta customers experienced 17% fewer unplanned robot stoppages. The figure comes from the company and has not been independently verified.
Pricing is per robot, not per operator, so a customer's monthly bill climbs with the size of the fleet whether or not anyone is watching the console that month. For a warehouse running a large robot count, the software becomes a recurring line item measured against the labor and downtime it displaces — the comparison the 17% claim is meant to settle, and the one buyers will have to weigh without third-party confirmation.
The Series B arrives on the same day as the product, giving HelioDock $39 million in total backing as it moves from a closed beta to open sales. North Coast Ventures led the round. HelioDock's core business remains the robots themselves; Night Shift adds a software layer sold on top of hardware the company already supplies.
The clearest constraint on that layer surfaced in the beta. Aaron Pike of Fulcrum Logistics, one of the nine participating companies, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. That is a distinction with consequences for warehouse operators deciding where remote supervision can replace a person on site: the software's value is tied to the reliability of a network the warehouse does not necessarily control.
Night Shift does not change what a robot does when the network drops. When connectivity fails, robots return to their locally configured behavior, and remote control becomes unavailable. Whatever an operator has set up on the floor is what the fleet falls back on, and the remote view that Night Shift sells is gone for the duration.
---
*Word count: 553*
Kimi-K3.1-Pro: # HelioDock launches Night Shift for remote warehouse robot oversight
*The company is pairing the software launch with a $24 million Series B, while a beta customer’s experience underscores its dependence on reliable connectivity.*
HelioDock on Thursday launched Night Shift, software that lets operators monitor fleets of autonomous warehouse robots remotely, and announced a $24 million Series B funding round led by North Coast Ventures.
The launch takes HelioDock beyond the autonomous warehouse robots it makes and into fleet-monitoring software. Night Shift is built around a straightforward operational problem: a robot can require attention when the responsible manager is somewhere else. Its commercial debut follows private testing with logistics companies, giving HelioDock early operational evidence to cite as it starts selling the product. The company’s reported reduction in unplanned stoppages is central to that case, though a beta customer’s experience also shows how remote management depends on the connection between operator and fleet.
The new investment brings HelioDock’s total funding to $39 million, with the $24 million round accounting for more than half of that amount. North Coast Ventures led the Series B, announced alongside Night Shift. Together, the financing and launch mark a shift from extended private testing toward a commercial software offering with published pricing and a general-availability date.
Night Shift spent 14 months in private beta with nine logistics companies and 310 robots. HelioDock said customers in the program experienced 17% fewer unplanned robot stoppages. That figure has not been independently verified, making it a company-supplied performance claim rather than an outside evaluation. The company is using the result to support the case for remote monitoring as the product moves into general release.
CEO Mara Velez framed the product around the inconvenience of an after-hours response: “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet.”
Her example makes remote access the alternative to an automatic trip to the warehouse. It also describes a more immediate need than simply watching a fleet: when one machine behaves unexpectedly, a manager needs a way to assess the situation and decide what to do next without first traveling to the site. Night Shift’s pitch is that oversight can begin where the manager is, not only where the robot happens to be. The emphasis is on who can respond, and from where, rather than on changing the robot’s own instructions when it encounters an obstacle.
Aaron Pike of Fulcrum Logistics, one of the beta customers, offered a more conditional assessment based on the quality of the connection. He said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His experience qualifies Velez’s two-in-the-morning scenario: the appeal of avoiding a trip to the warehouse rests on the connection needed to reach the robots.
When connectivity fails, robots return to their locally configured behavior, and remote control becomes unavailable. The local configuration remains the fallback, while Night Shift’s ability to support an off-site response is curtailed. That distinction is important to understanding what the software changes: it can provide remote oversight and control when the connection is available, but it does not replace the behavior already configured on the robot.
Night Shift will cost $90 per robot per month and become generally available on September 15. The per-robot price means the recurring charge rises as the number of robots covered increases. At that rate, 10 robots would cost $900 per month and 100 robots would cost $9,000 per month.
HelioDock will enter that commercial phase with alternatives already on the market. Relay Robotics and YardOS offer competing fleet-management products, giving operators other systems to evaluate when Night Shift becomes available.
OpenAI: GPT-6 Luna: Model A is more concise and restrained, but it appears substantially below the required 550-word minimum despite its stated count. Model B better meets the length requirement, though it repeats some points and goes beyond the notes by describing remote control as available when connectivity works. (Second judge pass, order swapped — scores are the average of both: Model B better meets the requested length and provides a cohesive, attributed account that includes the connectivity limitation, thou | Anthropic: Claude Fable 5.1: Model A delivers a tighter, more cohesive story…
Natural human rewrite
Rewrite the draft below as a 300–400-word publication-ready news story in Markdown.
Preserve every factual claim. Improve the lede, narrative flow, sentence rhythm and transitions. Remove repetition, stock hype, throat-clearing and mechanical signposting. The finished article should sound like an experienced human technology reporter wrote it.
Draft:
# Arcwell Launches An Innovative New Platform
Arcwell has announced an exciting new platform called Relay. The announcement represents a significant development for the company and for the broader customer service industry. Relay uses artificial intelligence to examine customer support conversations and suggest answers to agents.
The platform is important for several reasons. First, it can suggest answers. Second, it can identify customers who may cancel. Third, it can create summaries. These capabilities demonstrate how artificial intelligence continues to transform the customer service landscape.
Relay was tested by 12 companies over six months. Arcwell says agents using it closed tickets 14% faster. The company did not release customer-retention results. Relay costs $65 per agent each month.
CEO Priya Shah said, “Support agents spend too much time searching for information that their companies already have.”
However, the platform also presents potential challenges. Relay requires access to support transcripts and internal help documents. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period.
In conclusion, Relay represents an important new chapter for Arcwell as it seeks to help companies improve customer service using artificial intelligence.
Qwen: Qwen4 Max: # Arcwell Launches Relay, an AI Assistant for Support Teams
Arcwell has launched Relay, an artificial intelligence platform that reads customer support conversations and suggests replies to the agents handling them. It also identifies customers who may cancel and generates summaries of the exchanges. The price is $65 per agent per month.
Arcwell tested the platform with 12 companies over six months and says agents using it closed tickets 14% faster. That figure is the company's own. It did not release customer-retention results — a notable omission for a product whose pitch includes spotting departing customers before they go.
CEO Priya Shah framed the problem as one of retrieval rather than knowledge. "Support agents spend too much time searching for information that their companies already have," she said.
Relay's answer is to work from material a support organization already produces. The platform requires access to support transcripts and internal help documents, and it draws on both to draft responses, condense threads and gauge which accounts look shaky.
That requirement is also the central thing a buyer has to weigh. Handing over transcripts means handing over the unfiltered record of customer complaints; handing over help documentation means exposing internal operational knowledge. Arcwell says customer data is encrypted and is not used to train shared models, and it lets companies choose a retention period of 30 or 90 days.
For support leaders, the calculation comes down to whether a 14% gain in ticket-closing speed, measured across a dozen companies, is worth $65 per seat and access to that archive. Whether Relay also keeps customers who would otherwise leave is a question Arcwell has yet to answer with numbers.
Kimi-K3.1-Pro: # Arcwell’s Relay uses AI to suggest support replies and flag cancellation risks
Arcwell has launched Relay, an artificial intelligence platform that examines customer support conversations and suggests answers for agents. It also generates summaries and identifies customers who may be at risk of canceling, combining tools for handling individual tickets with signals intended to flag possible lost business.
The platform draws on information companies already possess but that agents may struggle to find quickly. Relay requires access to support transcripts and internal help documents to do its work.
“Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said.
Twelve companies tested Relay over six months. Arcwell says agents using the platform closed tickets 14% faster. That result offers an early measure of Relay’s effect on support operations, but it is limited to speed: it does not establish whether answers became more accurate or customers more satisfied. Arcwell also did not release customer-retention results, leaving unclear whether its cancellation warnings helped companies keep customers.
Relay costs $65 per agent each month. Its three core features—suggested answers, summaries and cancellation warnings—address different points in the support workflow, from composing a response to preserving an account of the exchange and highlighting a customer who may need attention.
Using those features requires broad access to company data. Connecting Relay to transcripts and internal documents gives the platform visibility into customer conversations and company material. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose either a 30-day or 90-day retention period.
Those commitments will sit alongside the trial results as prospective customers assess the platform. Arcwell’s clearest evidence so far is faster ticket handling; the case for Relay as a retention tool remains unquantified. For now, its pitch rests on turning existing support records into quicker responses and more usable information for agents, while giving businesses defined options for how long customer data is retained.
OpenAI: GPT-6 Luna: Model B better meets the 300–400-word requirement and delivers a clear, balanced news story. Model A is well written but appears under length and adds more speculative framing about data exposure and the product’s value to buyers. (Second judge pass, order swapped — scores are the average of both: Model B better meets the requested 300–400-word length and preserves the draft’s claims, though it adds some reasonable analysis. Model A is crisp, but falls short of the word-count requirement and mak | Anthropic: Claude Fable 5.1: Model A reads more like a seasoned reporter and…
Chronology with narrative momentum
Turn the scrambled reporting below into a 450–600-word Markdown news article.
Choose an opening that captures the consequence of the failure, then reconstruct the sequence clearly. Preserve the chronology without producing a minute-by-minute incident log. Give technical details enough context for a general reader.
Reporting:
* KiteBridge makes expense-management software.
* At 8:45 a.m. Tuesday, it deployed version 4.2.
* The first customer complaint arrived at 9:12 a.m.
* A payment-processor timeout caused KiteBridge’s fallback system to treat retries as new reimbursement requests.
* The software generated 13,400 duplicate reimbursements totaling $2.8 million.
* KiteBridge suspended automated payments at 10:03 a.m.
* Ninety-one corporate customers were affected.
* By Wednesday evening, $2.3 million had been recovered. The rest remained pending.
* No employees had money withdrawn from their accounts; employers funded the duplicate payments.
* An April changelog said version 4.0 had introduced “exactly-once payment protection.”
* KiteBridge’s public incident page initially described the problem as delayed reimbursements.
* The incident page first disclosed the duplicate payments six hours after automation was suspended.
* CEO Lena Ortiz apologized and said the company would publish a technical report.
Qwen: Qwen4 Max: # KiteBridge Update Sent $2.8 Million in Duplicate Reimbursements
**A safeguard the company advertised in April failed to stop repeated payment instructions, leaving 91 corporate customers to chase their own money back.**
Employers paid the same expenses twice on Tuesday after an update to KiteBridge's expense-management software generated 13,400 duplicate reimbursements worth $2.8 million. The money went to employees at 91 corporate customers, and every dollar of it came out of employer accounts — no employee had funds withdrawn. By Wednesday evening, $2.3 million had been recovered. Roughly half a million dollars remained outstanding.
## How the payments multiplied
KiteBridge released version 4.2 at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m., and the company suspended automated payments at 10:03 a.m. — about 80 minutes after the release went live.
The trigger was a timeout at a payment processor. When a processor fails to confirm a transaction within an expected window, the requesting software cannot tell whether the payment went through, so it tries again. Well-built payment systems handle this by tagging each request with a unique identifier, allowing the receiving side to recognize a repeat attempt as the same instruction rather than a fresh one. Engineers call this idempotency; it is the standard defense against paying twice.
KiteBridge's fallback system did not make that distinction. It treated each retry as a new reimbursement request, so a single approved expense could be paid out repeatedly — and at scale, across thousands of transactions, the duplication compounded into millions of dollars.
The failure is notable because KiteBridge's April changelog announced that version 4.0 had introduced "exactly-once payment protection" — a guarantee, in plain terms, that a payment instruction is carried out one time only, no matter how often it is retried. The company has not explained how Tuesday's behavior related to that feature, or whether version 4.2 changed it.
## A disclosure that lagged the damage
KiteBridge's public incident page initially characterized the problem as delayed reimbursements — close to the opposite of what customers were experiencing. The page did not disclose the duplicate payments until six hours after automated payments had been suspended, placing the first public acknowledgment in the late afternoon.
For finance teams, that gap mattered. Customers monitoring the incident page had reason to believe payments were merely slow, while duplicate funds were landing in employee accounts and drawing down company balances. The correction reframed the event from a service inconvenience into a recovery effort requiring employers to reverse transfers already received by their own staff.
## Recovery and what comes next
The $2.3 million recovered by Wednesday evening represents about 82 percent of the total. The remaining balance was still pending, and KiteBridge has not said how long recovery of those funds will take or what happens if some cannot be returned.
Because employers funded the duplicates, the burden of unwinding them falls on customers rather than on the employees who received them — a process that typically involves contacting staff, arranging reversals, and reconciling expense records that now show payments that should not exist.
Chief executive Lena Ortiz apologized for the incident and said the company would publish a technical report. That document will be the first real test of whether KiteBridge can explain how a system marketed on exactly-once payment protection paid the same expenses thousands of times over — and why its own status page described the failure as late money rather than too much of it.
Kimi-K3.1-Pro: # KiteBridge Software Failure Generates $2.8 Million in Duplicate Reimbursements
KiteBridge’s expense-management software generated **13,400 duplicate reimbursements worth $2.8 million** on Tuesday, creating a costly recovery effort affecting 91 corporate customers. Employers funded the excess payments, while no employees had money withdrawn from their own accounts. By Wednesday evening, $2.3 million had been recovered, but $500,000 remained pending.
The failure stood in contrast to a safeguard KiteBridge had announced months earlier. An April changelog said version 4.0 had introduced “exactly-once payment protection.” In payment software, that term generally describes controls intended to ensure a transaction is processed only once, even when a technical problem causes the same instruction to be submitted or retried multiple times. Such protection is meant to prevent network uncertainty from becoming a duplicate payout.
Tuesday’s sequence began at 8:45 a.m., when KiteBridge deployed version 4.2. The first customer complaint arrived at 9:12 a.m. The underlying problem was a payment-processor timeout: the processor did not respond within the expected window. A timeout can leave software uncertain about the status of a transaction, making the handling of any retry especially important.
KiteBridge’s fallback system—the backup process used when the normal payment path runs into trouble—treated the retries as new reimbursement requests rather than repeated attempts to complete existing transactions. Fallback logic is intended to keep a service operating when a dependency fails, but it must recognize when an apparently new request is actually another attempt at the same payment. In this case, that distinction was lost, producing duplicates on a large scale.
KiteBridge suspended automated payments at 10:03 a.m., less than an hour after the first reported complaint. Its public communications, however, initially gave a narrower picture of the incident. The company’s public incident page described the problem as delayed reimbursements and first disclosed the duplicate payments six hours after automation had been suspended. The initial description addressed a possible symptom—money arriving late—but not the duplicate-payment problem that had already led KiteBridge to halt automation.
The impact extended beyond delayed expenses. The duplicates represented additional reimbursements funded by employers, not deductions from employees’ personal accounts. No employees had money withdrawn from their accounts, meaning workers were not having funds pulled back from their own bank balances as part of the fallout. For the affected companies, the immediate challenge was accounting for thousands of excess payments while the recovery process remained incomplete.
By Wednesday evening, the recovery total had reached $2.3 million of the $2.8 million generated in duplicates. The remainder, approximately $500,000, was still pending. Most of the money had therefore been recovered within roughly a day and a half, but the incident had not yet been fully financially resolved.
CEO Lena Ortiz apologized and said KiteBridge would publish a technical report. Customers are still awaiting a fuller explanation of why the fallback process did not preserve the exactly-once behavior the company had previously advertised. For the affected employers, the more immediate question is when the outstanding funds will be recovered.
OpenAI: GPT-6 Luna: Model A better delivers the requested narrative momentum, moving from the financial consequence through the failure and into recovery with useful technical context. Model B is somewhat more cautious, but repeats points and feels less cohesive; Model A’s added detail about how employers may unwind payments is plausible but not established in the reporting. (Second judge pass, order swapped — scores are the average of both: Model A has stronger pacing and structure, but it adds unsupported details | Anthropic: Claude Fable 5.1: Model A opens with a sharp consequence-driven l…
Publication-ready Markdown edit
Edit the copy below into a publication-ready Markdown story.
Requirements:
* Supply an H1 headline and italicized dek
* Repair the lede and paragraph flow
* Use a Markdown blockquote only for the extended statement
* Convert the supplied URL into a descriptive inline link
* Remove internal notes and formatting debris
* Preserve the reporting and its meaning
* Output only the finished article
Raw copy:
HEADLINE TBD
The funding was announced today. Raincheck is the company. It has raised $8 million and it makes flood-risk software for local governments.
[NOTE: move this up?] The Austin company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm.
The seed round was led by Basin Ventures. Existing investor Long Mile also participated. Raincheck has raised $10.5 million total.
CEO Ana Flores said the following:
“Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.”
https://example.com/raincheck-methodology
Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced.
Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings.
THE COMPETITION
Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.
Qwen: Qwen4 Max: # Raincheck Raises $8 Million to Help Cities Decide Which Drains to Clear First
*The Austin company maps drainage systems and pairs them with rainfall forecasts, guiding where crews go before a storm arrives.*
Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. Its pitch is a matter of timing: cities already know where the water rises, but they have limited hours and limited crews to prepare.
The software maps a city's drainage system and combines those maps with rainfall forecasts. Municipal staff use the results to decide where to send maintenance crews ahead of a storm. Raincheck outlines [its flood-risk methodology](https://example.com/raincheck-methodology) in published documentation.
The seed round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has now raised $10.5 million in total.
CEO Ana Flores said the company is targeting a decision problem rather than a prediction problem:
> “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.”
Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them, saying some of the contracts have not been publicly announced.
The software's forecasts are recommendations, not orders. Emergency managers retain control over crew deployments and public warnings.
Raincheck competes with other flood-planning vendors, including HydroMap, as well as municipal engineering contractors. The company said the new funding will pay for eight engineering hires and support an expansion into Florida.
Kimi-K3.1-Pro: # Flood-Risk Software Maker Raincheck Raises $8 Million
*The Austin company helps local governments prioritize drainage maintenance before storms and plans to use the seed funding to hire engineers and expand into Florida.*
Raincheck, an Austin-based company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding.
The company maps drainage systems and combines those maps with rainfall forecasts, according to its [flood-risk methodology](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews before a storm.
Basin Ventures led the round, with participation from existing investor Long Mile. The new funding brings Raincheck’s total raised to $10.5 million. The company said the investment will help it hire eight engineers and expand into Florida.
Raincheck currently works with seven cities in Texas and Louisiana. It declined to identify them because some contracts have not been publicly announced.
CEO Ana Flores said:
> “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.”
Raincheck’s forecasts are recommendations. Emergency managers retain control over deployments and public warnings.
## The Competition
Other flood-planning vendors include HydroMap and municipal engineering contractors.
OpenAI: GPT-6 Luna: Model B preserves the supplied reporting more closely, while Model A adds unsupported framing about limited time and crews and describes the link as published documentation. Model A is slightly more polished, but Model B’s factual restraint and clear structure make it the stronger edit overall. (Second judge pass, order swapped — scores are the average of both: Model B preserves the supplied reporting with minimal additions, while Model A introduces unsupported framing about cities’ staffing con | Anthropic: Claude Fable 5.1: Model A delivers a smoother, publication-ready…
Exact final-draft structure
Convert the supplied fields into Markdown using these exact rules:
* H1 headline
* One blank line
* Italicized dek
* One blank line
* First body paragraph
* One blank line
* H2 section heading
* One blank line
* Second body paragraph
* Preserve all wording and punctuation exactly
* Do not use a code fence
* Output nothing else
Headline: Raincheck raises $8 million for flood-risk software
Dek: The Austin startup helps cities decide which drainage problems to address before storms.
First body paragraph: Raincheck has raised an $8 million seed round led by Basin Ventures.
Section heading: Expansion plans
Second body paragraph: The company plans to hire eight engineers.
Qwen: Qwen4 Max: # Raincheck raises $8 million for flood-risk software
*The Austin startup helps cities decide which drainage problems to address before storms.*
Raincheck has raised an $8 million seed round led by Basin Ventures.
## Expansion plans
The company plans to hire eight engineers.
Kimi-K3.1-Pro: # Raincheck raises $8 million for flood-risk software
*The Austin startup helps cities decide which drainage problems to address before storms.*
Raincheck has raised an $8 million seed round led by Basin Ventures.
## Expansion plans
The company plans to hire eight engineers.
OpenAI: GPT-6 Luna: Model A and Model B both follow the requested structure exactly and preserve all wording and punctuation. Their writing quality is equally clear and polished. (Second judge pass, order swapped — scores are the average of both: Model B and Model A both follow the requested structure and preserve all wording and punctuation exactly, so neither is better.) | Anthropic: Claude Fable 5.1: Model A and Model B produce character-for-character identical output that follows every structural rule exactly: H1, blank line, italicized dek, blank line, body, blank line, H2, blank line, b…
News judgment in the opening
Write only the opening of this news story: an H1 headline, an italicized dek and three to five paragraphs totaling 180–250 words.
The opening must establish what happened, quantify the scope, explain why it matters and carefully distinguish confirmed facts from unknowns. It should pull the reader forward without sensationalizing the incident.
Reporting:
* Cloudnote discovered that a debugging page containing customer prompt excerpts was accessible without authentication.
* The page was exposed for four hours on Monday.
* Logs show excerpts from 1,842 user sessions.
* The excerpts could include up to 200 characters from a prompt.
* Cloudnote removed the page 19 minutes after a researcher notified it.
* The company says it has found no evidence that anyone besides the researcher accessed the page.
* It has not completed its investigation.
* Full documents, passwords and payment information were stored elsewhere and were not present on the page.
* Cloudnote has 3.2 million registered users.
Qwen: Qwen4 Max: # Cloudnote Debugging Page Left Prompt Excerpts From 1,842 Sessions Reachable Without a Login
*The page was exposed for four hours Monday and pulled 19 minutes after a researcher's alert; the company reports no sign of other visitors but has not finished its investigation.*
Cloudnote left an internal debugging page accessible without authentication for four hours on Monday, exposing brief excerpts of customer prompts from 1,842 user sessions, according to the company's own logs.
Each excerpt could include up to 200 characters of a prompt — a sentence or two, potentially carrying a name, a client detail or a fragment of unfinished correspondence. Full documents, passwords and payment information were stored in separate systems and were not present on the page, the company said.
Cloudnote removed the page 19 minutes after an outside researcher notified it. The company says it has found no evidence that anyone besides that researcher accessed the page. It has not completed its investigation, and it has not explained how the page became publicly reachable.
The 1,842 sessions are a small fraction of activity at a service with 3.2 million registered users, though Cloudnote has not said how many distinct accounts they represent. Whether anyone else loaded the page during those four hours remains unresolved.
The investigation is continuing. Cloudnote has not said whether it will notify the users whose sessions appeared in the logs.
Kimi-K3.1-Pro: # Cloudnote Says Debugging Page Exposed Prompt Excerpts From 1,842 Sessions
*The page was accessible without authentication for four hours Monday. The company says it has found no evidence of access beyond the researcher who reported it, but its investigation is unfinished.*
Cloudnote discovered that a debugging page containing customer prompt excerpts could be opened without authentication. The company removed the page 19 minutes after a researcher notified it, ending an exposure that lasted four hours on Monday.
Logs show excerpts from 1,842 user sessions, with each excerpt containing as much as 200 characters from a prompt. Those figures describe the sessions and material involved, not the number of individuals affected or the amount of information actually viewed. Cloudnote said it has found no evidence that anyone besides the researcher accessed the page.
Cloudnote has 3.2 million registered users, but that figure cannot be used to infer how many customers were affected because sessions are not necessarily unique users. Full documents, passwords and payment information were stored elsewhere and were not present on the page, the company said.
Even short prompt fragments can contain information users expected to keep private, so the absence of full documents or credentials does not eliminate the privacy concern. Cloudnote has not completed its investigation, and the ultimate impact remains unclear. The company has not yet explained how the page became accessible; for now, there is no confirmed evidence of broader access or misuse.
OpenAI: GPT-6 Luna: Model B more carefully distinguishes logged sessions from unique customers and avoids implying that all excerpts were viewed. Model A is engaging, but its comparison with registered users and speculative examples of prompt content go beyond what the reporting establishes. (Second judge pass, order swapped — scores are the average of both: Model B is more careful about what the session count does and does not establish, while Model A adds speculative examples and an unsupported comparison between | Anthropic: Claude Fable 5.1: Model A delivers a tighter, more conventional n…
Numbers shaped into a story
Write a 600–750-word Markdown article from the benchmark results below.
Build the story around the clearest finding. Explain the methodology and limitations where readers need them. Avoid reciting the table row by row or treating small differences as decisive.
Benchmark:
* Researchers tested five language models on 200 anonymized customer-support tickets.
* Each model produced one proposed answer per ticket.
* Three experienced support managers graded the answers blindly.
* Scores combined factual correctness, usefulness and tone.
* Model North: 84.2 average score; $1.90 estimated cost per 1,000 tickets.
* Model Vale: 83.8; $0.74 per 1,000 tickets.
* Model Ember: 79.1; $0.31 per 1,000 tickets.
* Model Kite: 76.4; $0.22 per 1,000 tickets.
* Model Harbor: 71.0; $0.08 per 1,000 tickets.
* North and Vale’s score difference was not statistically significant.
* Vale answered 18% faster than North.
* North performed best on billing disputes.
* Vale performed best on technical troubleshooting.
* All five models performed poorly when a ticket depended on an image attachment.
* The test used English-language tickets from one software company.
* Researchers did not test agentic actions, tool use or multilingual support.
Qwen: Qwen4 Max: # The Top Two Models Tied on Quality. One Costs 61% Less.
The headline result from a recent 200-ticket customer-support benchmark isn't which model scored highest. It's that the two highest scorers were statistically indistinguishable — and one of them costs roughly two-fifths as much to run.
Model North averaged 84.2. Model Vale averaged 83.8. Researchers explicitly report that the 0.4-point gap was not statistically significant. Meanwhile, North's estimated cost was $1.90 per 1,000 tickets against Vale's $0.74, and Vale returned answers 18% faster. On the evidence available, paying 2.6x more for North buys no measurable quality improvement across the ticket mix tested.
That's the finding worth acting on. Everything below it is either a tradeoff or a caveat.
## How the test worked
Researchers drew 200 anonymized support tickets and had each of five models produce one proposed answer per ticket. Three experienced support managers graded the answers blind — they didn't know which model wrote what — scoring on a blend of factual correctness, usefulness and tone.
Two design choices matter for interpretation. Blind grading by domain experts is a meaningful strength: it removes brand expectations from the judgment, and support managers are the right people to assess whether a reply would actually resolve a customer's problem. But the composite score is also a limitation. A single number combining accuracy, helpfulness and tone can hide a model that is pleasant but wrong, or correct but unusable. The benchmark doesn't publish the component breakdown, so we can't tell whether North and Vale tie because they're similar or because their strengths cancel out.
## Where the tie might break
The category results hint at exactly that kind of cancellation. North performed best on billing disputes; Vale performed best on technical troubleshooting.
These are suggestive, not decisive. Splitting 200 tickets into categories leaves small subgroups, and the researchers flag significance only for the overall North–Vale comparison. Treat the category split as a hypothesis to test on your own ticket distribution, not a routing rule to deploy. If your queue is overwhelmingly technical, Vale looks like the obvious pick. If it's dominated by refund and invoice arguments, North may earn its premium — but you'd want to verify that on a larger billing sample before committing.
Below the top pair, the picture is a fairly conventional quality-for-cost curve. Ember, Kite and Harbor get progressively cheaper and progressively weaker, with Harbor at 71.0 — thirteen points off the lead. The interesting question there isn't "which is best" but "what score is good enough," and that depends on whether a human reviews drafts before they ship. A cheap model behind a careful reviewer is a different product than a cheap model sending replies directly. This benchmark can't answer that question, because it graded proposed answers, not end-to-end outcomes.
## The shared blind spot
All five models performed poorly when a ticket depended on an image attachment. Not one model — all of them.
That's the most operationally important result in the set, and it's easy to skim past because it doesn't differentiate the vendors. Screenshots of error dialogs, photos of receipts, annotated UI captures: these are ordinary in software support. If a meaningful share of your queue depends on them, no model in this test is safe to run unsupervised on that slice. The right response is a routing rule that detects image-dependent tickets and sends them to a human, regardless of which model you choose. Model selection won't fix this; workflow design will.
## What the results don't cover
The scope is narrow in ways that limit generalization:
- **One company, one language.** English-language tickets from a single software company. Tone norms, product vocabulary and difficulty distribution all travel poorly across domains.
- **Single-turn answers only.** No agentic actions, no tool use, no multilingual support were tested. Real support work often means looking up an account, issuing a refund or checking a log — none of which this benchmark measures. A model that writes beautifully but calls tools badly could rank first here and fail in production.
- **Estimated costs.** The figures are estimates per 1,000 tickets, not invoices. Your actual spend depends on your prompt length, ticket verbosity and retry behavior.
- **A tie is not equivalence.** "Not statistically significant" at n=200 means the study couldn't detect a difference — not that no difference exists. A larger sample might separate them.
## A practical read
Vale is the default choice on this evidence: same measured quality, lower cost, faster responses. North is a candidate worth a targeted second look only if billing disputes dominate your volume. And whichever you pick, budget for human handling of image-dependent tickets — that gap is universal, and it's the one most likely to reach a customer.
Kimi-K3.1-Pro: # The Runner-Up Is the Real Story in This Support-AI Benchmark
**Model North earned the highest average score, but Model Vale delivered statistically indistinguishable overall quality in this test—at a much lower estimated cost and with faster answers.**
The clearest finding is not that North narrowly topped the leaderboard. It is that Vale may offer the more practical balance of quality, speed and cost for customer-support drafting.
North averaged 84.2, compared with 83.8 for Vale. That 0.4-point difference was not statistically significant, so the results do not establish North as the better overall model. Meanwhile, Vale’s estimated cost was **$0.74 per 1,000 tickets**, versus **$1.90 for North**—about 61% lower. Vale also answered 18% faster.
For a support organization processing large ticket volumes, those operational differences are more meaningful than a small leaderboard gap. That does not make Vale the automatic winner for every deployment, but it makes the case for paying North’s higher price considerably weaker unless a team has evidence that North performs better on its particular workload.
## How the models were evaluated
Researchers tested five language models on 200 anonymized customer-support tickets. Each model produced one proposed answer per ticket, and three experienced support managers graded the answers blindly. The combined scores covered factual correctness, usefulness and tone.
That design provides a useful comparison of answer quality while reducing the risk that graders favored a model because of its name or reputation. It also evaluates the models on the same ticket set rather than on separately selected examples.
However, one answer per ticket offers only a limited view of consistency. The benchmark does not show how much quality might vary across repeated runs, different prompts or changes in ticket phrasing. The estimated costs also should not be confused with total support costs: human review, escalation, integration and the consequences of incorrect answers can all affect the economics of deployment.
## Cheaper still does not necessarily mean better value
Below the two leaders, lower estimated prices came with more noticeable declines in average scores. Ember and Kite occupied the middle of the results, while Harbor combined the lowest score, 71.0, with the lowest cost, $0.08 per 1,000 tickets. Its score was 13.2 points below North’s.
Those results suggest that Vale’s combination of near-top quality and reduced cost is more distinctive than the general proposition that cheaper models save money. A model that produces less accurate or less useful drafts may create additional work elsewhere, although this benchmark did not measure those downstream effects.
The reported results also do not justify treating every remaining score gap as decisive. Statistical significance was specifically reported for the North–Vale comparison, not for every pairwise difference among the other models.
Ticket mix adds another qualification. North performed best on billing disputes, while Vale performed best on technical troubleshooting. A company whose queue is dominated by one of those categories may therefore reach a different conclusion from one with a more balanced workload. Category-level strengths can matter more than overall averages when the work is specialized.
## Every model shared an important weakness
All five models performed poorly when a ticket depended on an image attachment. This is arguably the benchmark’s most important deployment warning: strong averages on the broader ticket set do not mean the models handled every common support scenario adequately.
A team testing these systems should separately identify attachment-dependent tickets and evaluate an appropriate fallback, such as additional processing or human review. The shared failure also means that choosing a more expensive model did not solve the problem within this test.
## A narrow benchmark, not a universal ranking
The test used English-language tickets from one software company. Its findings may not transfer cleanly to other industries, support policies, customer expectations or ticket distributions. Researchers also did not test agentic actions, tool use or multilingual support, so the results say nothing about whether these models can reliably retrieve account information, operate support systems, complete multi-step workflows or serve a global customer base.
The sensible next step is therefore a company-specific evaluation, using representative tickets and measuring review effort and failure modes alongside quality, latency and cost.
On this benchmark’s evidence, **Vale is the strongest default candidate for a text-focused drafting pilot**, while North remains worth examining for billing-heavy workloads. But the broader lesson is more useful than either label: small score differences matter less than meaningful cost and speed differences, fit with the actual ticket mix, and known weaknesses that averages can hide.
OpenAI: GPT-6 Luna: Model B is more careful about what the benchmark supports while still building a clear story around Vale’s cost and speed advantages. Model A is vivid and practical, but its framing of the results as a tie and its categorical warning about unsupervised image-dependent tickets overstate the evidence somewhat. (Second judge pass, order swapped — scores are the average of both: Model A builds a sharper story and gives readers more useful guidance about what the benchmark does and does not establish | Anthropic: Claude Fable 5.1: Both outputs correctly center the North–Vale st…