A newsroom-focused matchup tests Qwen4 Max and Claude Sonnet 5.5 on reported-story writing, editing, accuracy, and structure. The comparison weighs both polished narrative work and tightly constrained editorial tasks.
Qwen4 Max takes the saved verdict, with an aggregate score of **105.0** to Claude Sonnet 5.5’s **97.5**. It won seven tasks; Claude won none, and five ended in ties. That makes this a decisive result in this head-to-head—not a statistical claim about broader model performance.
The wins span several kinds of newsroom work: controlled wit, cohesive stories from reporting notes, publication-ready edits, weaving quotes into narrative, shaping numbers into a story, chronology, and natural rewrites. The breadth matters: Qwen’s edge wasn’t confined to one format or one narrow test.
The ties are meaningful, too. Both models matched on exact final-draft structure, Markdown error count, strongest accurate lede, startup spotlight, and news judgment in the opening. Those results show Claude could meet the bar on constrained or accuracy-sensitive assignments, even without a recorded task win.
**Final call: Qwen4 Max wins this saved task sample decisively, on the strength of seven task wins and the higher aggregate score. This assessment applies only to the saved task sample; it does not establish general capability superiority.**
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 Takes an Audience Remark Literally, Assigns Its CEO Two Tasks
**The company launched Proxy on Wednesday. The demonstration surfaced a question the product category has not settled: what counts as an instruction.**
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 that capability can misfire when a system cannot tell a directive from an aside.
Partway through the demo, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy parsed the remark as an instruction. It generated two tasks and assigned both to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet."
Wu laughed, completed the second task and deleted the first. For a product marketed on reducing meeting overhead, Proxy spent part of its debut creating work for the executive running the presentation.
The episode was brief, but it mapped a real boundary in the product's design. QueueZero says Proxy requires confirmation before any external message is sent, meaning a request to email a pricing sheet to a customer would stop for human approval. Internal tasks carry no such gate and can be created automatically. Administrators can disable automatic task creation entirely.
That split — friction on outbound communication, none on internal work — reflects a common assumption in the meeting-assistant category: that the cost of a mistaken email is high and the cost of a mistaken to-do is low. Wednesday's demonstration suggested the second half of that assumption is not free. Task lists that accumulate items nobody asked for still have to be read, triaged and cleared by someone.
The underlying problem is linguistic rather than mechanical. Meeting speech is full of conditionals, complaints, jokes and remarks addressed to a neighbor rather than the room. "Someone should" is a hedge in conversation and an imperative to a parser looking for verbs and owners. Systems that listen continuously to ambient discussion inherit that ambiguity, and the more faithfully they transcribe, the more of it they carry forward into action.
QueueZero's administrative toggle is the current answer: organizations that want notes without autonomous task creation can turn the latter off. Whether that setting becomes the default posture for cautious buyers is a question the beta will begin to answer.
Proxy costs $30 per user per month and enters public beta next week.
Anthropic: Claude Sonnet 5.5: # QueueZero Launches Proxy, an AI Meeting Assistant That Assigns Work to Software Agents
QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend meetings, produce notes and assign follow-up work to software agents. The company said Proxy enters public beta next week and will cost $30 per user per month.
The launch demonstration showed both the product's appeal and the risk of automating action items. A customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy treated the comment as an instruction and 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.
## Why the moment matters
The exchange was an offhand remark, not a command, and Proxy turned it into a task list. Meeting assistants have mostly been summarization tools. Proxy goes further by converting conversation into assigned work, which means it must decide which statements are requests and who is responsible for them. A system that gets this wrong can produce clutter, or something worse.
QueueZero says it has built in limits. Proxy requires confirmation before an external message is sent, so a misread remark should not reach a customer or other outside party without a person approving it. The tasks in the demonstration were internal, and the company said internal tasks can be created automatically.
Automatic creation is not mandatory. Administrators can disable it, which lets organizations require review before Proxy adds items to anyone's list. Companies with strict approval processes or sensitive client relationships may prefer that setting, at least at first.
## What to watch in the beta
The public beta will test how Proxy performs outside a controlled demonstration. Meetings are full of jokes, hypotheticals, interruptions and half-formed suggestions, and the assistant's value will depend on telling those apart from real commitments. Customers will likely look at how often it creates tasks nobody wanted, how easily those tasks can be corrected and how clearly the confirmation step works when an external message is involved.
Wu, to his credit, did not appeal the first task; he deleted it.
The $30 monthly per-user price puts Proxy in the range of other workplace software subscriptions, and the cost will add up for larger teams. QueueZero has not suggested that the demonstration changed its plans, and the product is on track for next week's beta.
For now, the incident shows both what Proxy does and where its design needs the most care. The assistant acted on what it heard, and the safeguards QueueZero describes are meant to keep a person in control of what happens next.
OpenAI: GPT-6 Luna: Model A offers sharper analysis of the distinction between external-message safeguards and automatic internal tasks, while Model B adds a few unsupported comparisons and status claims. Both meet the length and tone constraints, with one restrained humorous line each. (Second judge pass, order swapped — scores are the average of both: Model A offers sharper analysis of the distinction between external messages and automatically created internal tasks, while keeping the story focused and readable. | Anthropic: Claude Fable 5.1: Both hit the word range, include every reportin…
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 Oversight Off the Warehouse Floor
*Night Shift exits a 14-month beta across 310 robots with a company-reported 17% drop in unplanned stoppages — and a dependency on the network at each site.*
HelioDock on Thursday launched Night Shift, software that lets operators monitor fleets of autonomous warehouse robots without being in the building. The release closes a 14-month private beta and lands alongside a $24 million Series B led by North Coast Ventures, bringing the robotics maker's total funding to $39 million.
The product takes a task that has conventionally kept a person near the machines — noticing that a robot has stalled and working out what to do about it — and turns it into something an operator can handle from somewhere else. HelioDock is selling into a market where Relay Robotics and YardOS already offer fleet-management products, and where the practical value of remote oversight rests on a variable the software company does not control: the quality of the connection at each customer site.
Night Shift's beta ran for 14 months with nine logistics companies and 310 robots, according to HelioDock. The company says those customers experienced 17% fewer unplanned robot stoppages during the trial. That figure is company-supplied and has not been independently verified, and it is the central number HelioDock is putting in front of buyers who already have alternatives.
The problem the company describes is a specific and unglamorous one. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," CEO Mara Velez said. A single stalled machine, in other words, is enough to pull a person out of bed and into a vehicle — the kind of interruption that remote monitoring is meant to absorb.
Aaron Pike of Fulcrum Logistics, one of the beta participants, said the system worked well in facilities with strong connectivity. It became less useful, he said, during rural LTE outages — a caveat that maps onto the geography of warehouse networks, where sites are not uniformly well served.
The software's behavior during those outages is defined: when connectivity fails, robots return to their locally configured behavior, and remote control becomes unavailable. The machines do not stop for want of an operator, but the operator also cannot reach them. Whatever response has been configured on site is the response that runs.
That makes Night Shift's benefit uneven across a customer's footprint. A facility with dependable coverage gets the full version of the pitch — a manager who can stay home. A rural site on a shaky LTE link gets monitoring that thins out precisely when something goes wrong, and falls back on local settings until the link returns. Buyers running both kinds of sites are effectively purchasing two different levels of service at one price.
The price is $90 per robot per month, which scales with fleet size rather than with facility count. Applied to a fleet the size of the beta group — 310 robots — that comes to roughly $27,900 a month, or about $335,000 a year. Whether the reported reduction in unplanned stoppages justifies that spend is a calculation each operator will have to make against its own downtime costs, since the 17% figure comes from HelioDock's own beta cohort rather than an outside assessment.
HelioDock is not the first vendor to argue that robot fleets should be managed from a dashboard. Fleet-management products from Relay Robotics and YardOS are already available, which means Night Shift arrives as a competitive entry rather than a new category, and will be evaluated against tools that warehouse operators may already be running.
The Series B, led by North Coast Ventures, gives HelioDock $39 million in total funding to work with as it moves from a controlled beta of nine customers to open sales. The beta group's robots were monitored under conditions the company selected; the general-availability cohort will choose its own sites, connectivity and local fallback configurations.
Night Shift becomes generally available on September 15.
Anthropic: Claude Sonnet 5.5: # HelioDock Launches Remote Monitoring Software for Warehouse Robots and Raises $24 Million
*Night Shift lets operators oversee robot fleets from afar. It will cost $90 per robot per month when it becomes generally available Sept. 15.*
HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. The company also announced a $24 million Series B round led by North Coast Ventures.
Night Shift is meant to reduce the number of times a person has to go to the warehouse when a robot stops working. HelioDock says customers in a private beta had 17% fewer unplanned robot stoppages. That figure has not been independently verified. A beta customer said the software's usefulness dropped when internet service did, and HelioDock will be selling into a market that already has competing fleet-management products.
## What the beta showed
Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock supplied the 17% figure and did not provide further detail on how it was measured.
CEO Mara Velez framed the product around the inconvenience of overnight troubleshooting. “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.
## Where it falls short
Aaron Pike of Fulcrum Logistics, a beta customer, said the system worked well in facilities with strong connectivity. It became less useful during rural LTE outages, he said.
HelioDock confirmed how the software behaves when a connection fails. Robots return to their locally configured behavior, and remote control becomes unavailable. In practice, an operator relying on Night Shift cannot intervene from off-site during an outage, and the robots continue under whatever settings were configured locally.
## Pricing and competition
Night Shift will cost $90 per robot per month. At that rate, a 100-robot fleet would cost $9,000 a month. The software becomes generally available Sept. 15.
Relay Robotics and YardOS both sell competing fleet-management products. HelioDock did not provide comparisons with either company's offerings.
## Funding
The Series B brings HelioDock's total funding to $39 million. The company did not detail in the notes how it plans to use the new capital.
OpenAI: GPT-6 Luna: Model A better meets the requested length and builds a more cohesive story, though it adds some unsupported implications about beta conditions and outage behavior. Model B is clearer and more restrained in places, but falls well short of the 550-word minimum and introduces unsupported claims about what HelioDock did not provide. (Second judge pass, order swapped — scores are the average of both: Model A is much closer to the required length and builds a more cohesive, publication-ready story, wh | Anthropic: Claude Fable 5.1: Model A delivers a cohesive, within-range artic…
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 Fix Drainage Before the Storm
*The Austin startup's seed round, led by Basin Ventures, brings its total funding to $10.5 million as it plans to hire eight engineers and expand into Florida.*
Raincheck, an Austin company that builds flood-risk software for local governments, announced $8 million in seed funding today. The software maps municipal drainage systems and combines those maps with rainfall forecasts, giving cities a way to decide where to send maintenance crews before a storm arrives.
The round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has raised $10.5 million in total.
CEO Ana Flores said the product is aimed at a prioritization problem rather than a mapping one:
> "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."
The company describes how it generates those priorities in [its published forecasting methodology](https://example.com/raincheck-methodology). The outputs are recommendations: emergency managers retain control over deployments and public warnings.
Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced.
It competes with other flood-planning vendors, including HydroMap, as well as municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.
Anthropic: Claude Sonnet 5.5: # Raincheck Raises $8 Million to Help Cities Prepare for Floods
*The Austin software company, which maps drainage systems and pairs them with rainfall forecasts, plans to hire engineers and expand into Florida.*
Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million, it announced today.
The software maps drainage systems and combines those maps with rainfall forecasts. Cities use it to decide where to send maintenance crews before a storm. More detail on the approach is available in [Raincheck's methodology](https://example.com/raincheck-methodology).
Basin Ventures led the seed round, and existing investor Long Mile also participated. Raincheck has now raised $10.5 million in total.
CEO Ana Flores said the challenge is not identifying where flooding occurs:
> "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 because some contracts have not been publicly announced.
Its forecasts are recommendations, and emergency managers retain control over deployments and public warnings.
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.
OpenAI: GPT-6 Luna: Model B stays closer to the supplied reporting and avoids Model A’s unsupported characterization of the link as a published forecasting methodology. Model A has slightly stronger paragraph flow, but its headline also suggests Raincheck directly fixes drainage problems. (Second judge pass, order swapped — scores are the average of both: Model B preserves the reporting closely and uses a clear, accurate headline and link. Model A is polished, but its headline implies Raincheck itself fixes drainag | Anthropic: Claude Fable 5.1: Both outputs satisfy every structural requireme…
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: # Foundry Hall to Close on October 31
Foundry Hall, the nonprofit startup accelerator that opened in Milwaukee in 2018, will close on October 31 after seven years. It has money to operate through the end of the month, but not to run another full program.
The cause is arithmetic. Blank River Bank, its largest corporate sponsor, supplied 38 percent of Foundry Hall's annual budget. This year the bank declined to renew. A spokesperson said the decision reflected a change in the bank's community investment priorities and was unrelated to Foundry Hall's performance.
A sponsor of that size is not quickly replaced. Losing it removed more than a third of the year's funding at once, and what remained covered the runway to October and nothing past it. Nonprofit accelerators run on money that is discretionary somewhere else: a shift in a bank's giving strategy is a line in its annual report and a closing date in someone else's calendar.
Executive director Camille Ross said the ending came slowly, across months in which the gap never closed.
"We kept waiting for the replacement check that never came," she said.
In seven years, Foundry Hall supported 212 startups, which collectively employ 1,140 people. Those companies will continue. The 1,140 jobs were the number the hall used to justify itself; none of them were its own. Foundry Hall's six employees will lose their jobs, and the twelve startups now using its workspace must be out by the end of October.
For those twelve, the closure means solving in a few weeks the problem the hall existed to solve for them: somewhere to work, and someone who will take the meeting.
Founder Dev Malik counts the loss in a smaller unit.
"My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop."
The distinction he draws is not between success and failure but between having a room and not having one.
Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies. Whether any of those conversations produces a lease before the deadline remains open.
What is closing is not only a workspace. It is a set of circumstances — a room where a first meeting could happen, and a staff whose job was to arrange it. That work leaves no monument and few records anyone will read later: introductions made, pitches rehearsed, a founder told to try the sentence again, differently.
Until the last day of the month, the lights stay on. The twelve companies keep working. The six employees keep coming in.
Anthropic: Claude Sonnet 5.5: # After Seven Years of Early Pitches, Foundry Hall Prepares to Close
Foundry Hall, the Milwaukee nonprofit startup accelerator that opened in 2018, will close October 31.
In that time, it has supported 212 startups. Together, those companies employ 1,140 people, a figure that carries more weight now that the doors are closing.
The end came down to arithmetic. Blank River Bank, Foundry Hall's largest corporate sponsor, supplied 38% of its annual budget. The bank declined to renew. A spokesperson said the decision reflected a change in the bank's community investment priorities and had nothing to do with how Foundry Hall performed.
Executive director Camille Ross said the staff spent months hoping the gap would be filled. "We kept waiting for the replacement check that never came."
Foundry Hall has enough money to operate through October, but not enough to run another full program. Six employees will lose their jobs. Twelve startups that currently work out of the building must leave.
For the founders who have passed through, the loss is practical, but it is also something else. Dev Malik remembers what the building gave him when he had little else.
"My first investor meeting happened in that building," Malik said. "Before Foundry Hall, I was pitching from the back table of a coffee shop."
He did not describe the meeting in detail, and he did not need to. A conference room, a door that closed, and a place to be taken seriously can be hard to find at the start, and easy to take for granted once found.
The immediate concern is where the twelve current companies will go. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for them. The conversations are meant to keep the companies together, or at least keep them working, while they sort out what comes next.
What happens to the rest of what Foundry Hall built is harder to say. Its alumni will still have their companies and their payrolls. The mentors who came through its rooms will still know one another's names. But the place where those connections began will be gone.
Until October 31, the building remains open. The staff will go on doing what they have done since 2018, helping founders prepare for their next meeting, even as they prepare to leave the building themselves.
OpenAI: GPT-6 Luna: Model A is within the requested 400–500 words and weaves the facts into a restrained narrative with selective quotations. Model B is substantially underlength and adds unsupported claims about the city’s plans and Foundry Hall’s mentors. (Second judge pass, order swapped — scores are the average of both: Model A is stronger: it meets the requested length and weaves the selected quotations into a more developed emotional arc. Model B is restrained and clear but falls short of 400 words; Model A’s | Anthropic: Claude Fable 5.1: Both entries hit every fact, quote the two huma…
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.
Anthropic: Claude Sonnet 5.5: # 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 produce identical output that follows the required structure and preserves all supplied wording and punctuation. (Second judge pass, order swapped — scores are the average of both: Model B and Model A are identical and follow the requested structure, preserve the wording and punctuation, and add nothing else.) | Anthropic: Claude Fable 5.1: Model A and Model B produced byte-identical outputs that follow every structural rule exactly: H1, italicized dek, body paragraph, H2, second paragraph, each separated by one blank line, with all wording preserved, 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 most expensive model didn't win
The headline result of this 200-ticket support benchmark isn't which model scored highest. It's that the highest score and the second-highest score are statistically indistinguishable — and one of them costs 2.6 times as much to run.
Model North posted an 84.2 average. Model Vale posted 83.8. The researchers explicitly report that the 0.4-point gap is not statistically significant. North costs an estimated $1.90 per 1,000 tickets; Vale costs $0.74. Vale also answered 18% faster.
On this evidence, the premium bought nothing measurable.
## How the test worked
Researchers pulled 200 anonymized customer-support tickets from a single software company, all in English. Each of five models produced one proposed answer per ticket. Three experienced support managers graded the answers blind, scoring a composite of factual correctness, usefulness and tone.
Two features deserve credit. Blind grading removes brand expectation from the ratings, and using working support managers rather than crowd workers means the judgments reflect what a real queue owner would accept.
Two features deserve caution. A composite score can hide a lot: a model that is warm and readable but occasionally wrong can land near a model that is terse and reliable. And no inter-rater agreement figures were reported, so we can't tell how much the three managers concurred.
## Reading the tie correctly
"Not statistically significant" is not the same as "proven equal." With 200 tickets, the test may simply lack the power to detect a small real difference. What it does establish is that this dataset gives you no defensible reason to pay extra for North.
It's also worth keeping the money in proportion. At 100,000 tickets a month, the gap between North and Vale is about $116 — less than a couple of hours of a support manager's time. Inference price is rarely the dominant cost in a support operation; human review, escalations and repeat contacts are. The practical importance of the tie isn't the savings. It's that price stops being a reason to choose, which frees you to decide on other grounds.
## Where the choice gets interesting
The category results are more useful than the leaderboard. North performed best on billing disputes; Vale performed best on technical troubleshooting. That's a genuine split, and it suggests routing by ticket type rather than picking a single winner.
But no per-category scores or sample sizes were published, so we can't size the effect or know whether it survives statistical scrutiny. With 200 tickets divided across categories, the subsets are small. Treat routing as a hypothesis to test on your own ticket mix, not a conclusion.
## The finding that should change your workflow
All five models performed poorly when a ticket depended on an image attachment.
That is the most actionable result in the study, and it isn't a model-selection problem. Screenshots of error dialogs, photos of receipts, annotated UI captures — none of the five handled them well. Switching vendors won't fix it. Detecting image-dependent tickets and routing them to a human, or adding a transcription step, will move your quality numbers far more than 0.4 points at the top of the table.
## The rest of the field
Ember, Kite and Harbor clustered between 71.0 and 79.1, roughly five to thirteen points behind the leaders, at costs from $0.08 to $0.31 per 1,000 tickets. That gap is much wider than the one at the top, and in a domain where the composite includes factual correctness, a wrong answer generates a repeat contact that erases the savings. Still, no significance tests were reported for these pairs either, so don't read the middle of the pack as a firm ranking.
## What this test can't tell you
The scope is narrow: English only, one company, one product domain, 200 tickets. Researchers did not test agentic actions, tool use or multilingual support — so nothing here predicts how these models behave when issuing a refund or querying an account.
Each model also produced a single answer with no retrieval, retries or human editing, which is not how most teams deploy them. And the costs are estimates, excluding integration, review and escalation.
**Bottom line:** Vale is the sensible default here — comparable quality, lower cost, faster. North is worth a targeted trial on billing disputes. And whatever you pick, build an escape hatch for image attachments first.
Anthropic: Claude Sonnet 5.5: # At the Top of the Support-Bot Rankings, the Pricier Model Didn't Win
A new benchmark of five language models answering customer-support tickets suggests that paying the most did not buy a measurably better answer. The top two models finished in a statistical tie, and one cost about 60% less than the other.
## What the researchers did
The team gave five models, North, Vale, Ember, Kite and Harbor, the same 200 anonymized customer-support tickets. Each model wrote one proposed reply per ticket. Three experienced support managers graded the replies blind, without knowing which model produced which answer. Each score combined factual correctness, usefulness and tone, so a reply had to be accurate and helpful as well as appropriately worded.
## The clearest finding: a tie at the top
North averaged 84.2 and Vale averaged 83.8. The researchers report that this gap is not statistically significant, so the data gives no good reason to call North the better writer of support replies.
The costs are very different. North's estimated price is $1.90 per 1,000 tickets, while Vale's is $0.74. Vale was also 18% faster to answer.
At small volumes the difference is trivial. At a million tickets, the estimated gap is about $1,160. For a large support operation, though, getting statistically indistinguishable quality for roughly 60% less money, with faster replies, is hard to ignore.
## Where the trade-off does appear
Below the top pair, the scores drop and the prices drop with them. Ember scored 79.1 at $0.31 per 1,000 tickets, Kite 76.4 at $0.22, and Harbor 71.0 at $0.08. Harbor is the cheapest by a wide margin, but it sits about 13 points behind the leaders.
Whether that gap matters depends on the job. A team drafting replies that humans review might tolerate a weaker, cheaper model. A team sending answers directly to customers probably would not. The benchmark can show the size of the quality gap, but not how much that gap is worth to a given business.
## Different strengths, with caution
North did best on billing disputes, and Vale did best on technical troubleshooting. That could matter for routing tickets to different models, but it should be treated as a hint rather than a conclusion. With only 200 tickets spread across categories, each category contains relatively few, and the results don't say whether these category-level differences were statistically meaningful. Since the models' overall scores were so close, small category edges could easily be noise.
## A shared weakness: images
All five models performed poorly when a ticket depended on an image attachment, such as a screenshot of an error message or a photo of a billing statement. This was the one problem the whole field shared, and price made no difference to it. Any team that gets many image-based tickets should expect to keep a human in the loop for those, or test separately whether a model handles them well.
## Limits of the study
Several caveats affect how far these results travel:
- **One company, one language.** The tickets were in English and came from a single software company. Products, customer behavior and ticket types vary widely, so the rankings might shift in other industries, such as healthcare, travel or hardware.
- **Narrow task.** Each model wrote one proposed answer per ticket. The researchers did not test agentic actions, tool use or multilingual support. Issuing a refund, looking up an account or replying in Spanish would be different tests.
- **Estimated costs.** The prices are estimates per 1,000 tickets. Real bills depend on ticket length, prompt design, volume discounts and how often answers are regenerated.
- **Human judgment.** Three managers is a reasonable panel and blinding is a strength, but graders can differ in what they consider helpful or well-toned. The combined score also folds three qualities into one number, which can hide, for example, a model that sounds friendly but is occasionally wrong.
- **Small differences.** Gaps of a point or less, like the one between North and Vale, shouldn't be read as meaningful, and the same caution applies to modest gaps elsewhere in the ranking.
## The takeaway
The main lesson isn't that Vale is the best model or that cheap models are good enough. It's that, in this test, the most expensive model did not separate itself from a much cheaper rival on answer quality. Buyers comparing models for support work should test on their own tickets, weigh cost and speed alongside quality scores, and plan for human review wherever images are involved.
OpenAI: GPT-6 Luna: Model B gives a clear account of the top-model tie while more consistently distinguishing what the benchmark shows from what remains uncertain. Model A is engaging and actionable, but it makes several stronger operational claims than the results support. (Second judge pass, order swapped — scores are the average of both: Model B is more consistently grounded in the supplied results and handles the study’s limits with appropriate caution. Model A is punchier, but it presents unsupported details a | Anthropic: Claude Fable 5.1: Model A builds a sharper, more insightful narra…
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
Anthropic: Claude Sonnet 5.5: **6**
Six lines break the rules:
1. `#Flood software moves upstream` has no space after the `#`, so it is not a valid `# Headline`.
2. `The company is expanding its pilot.` is the dek, and it is not italicized.
3. `##Details` has no space after the `##`.
4. `Read [the announcement](www.example.com/news).` uses a link that is not an absolute HTTPS URL.
5. `* The first pilot begins Monday.` begins its list item with an asterisk instead of a hyphen.
6. `<div>The second pilot begins Friday.</div>` contains raw HTML.
Scored against the benchmark's reference answer (numeric match): Model A was correct, Model B was correct.
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: # Software Update Sent $2.8 Million in Duplicate Reimbursements, Leaving Employers to Cover the Cost
Ninety-one companies paid out $2.8 million they never owed after an update to KiteBridge's expense-management software turned routine payment retries into 13,400 duplicate reimbursements. Employees who received the extra money did not have it pulled back from their accounts. Their employers funded the windfall — and are now waiting to get it back.
## A retry that became a new payment
KiteBridge deployed version 4.2 at 8:45 a.m. Tuesday. The first customer complaint arrived 27 minutes later, at 9:12 a.m.
The trigger was a timeout at a payment processor: the outside service that moves money failed to confirm, within the expected window, whether a payment had gone through. That left KiteBridge's software in an ambiguous state, and its fallback system responded by trying again. But instead of treating each retry as another attempt at the same reimbursement, the fallback treated it as a brand-new request. Every attempt produced another payment.
The result was 13,400 duplicate reimbursements totaling $2.8 million. KiteBridge suspended automated payments at 10:03 a.m., roughly 78 minutes after the update went live.
## A safeguard the company had advertised
Duplicate payments are a well-known hazard in software that moves money, because a system can never be fully certain whether an instruction it sent was carried out. The standard defense is a unique identifier attached to each payment, so that a repeated instruction is recognized as the same transaction rather than an additional one — a property engineers call "exactly-once" delivery.
KiteBridge had marketed precisely that. An April changelog announcing version 4.0 said the release introduced "exactly-once payment protection." Tuesday's incident occurred on a later version of the same product, in the fallback path meant to handle payments whose status was unclear.
## The public account lagged the internal one
For hours, customers reading KiteBridge's public incident page were told something different from what was happening. The page initially characterized the problem as delayed reimbursements — money arriving late, rather than money arriving repeatedly.
The duplicate payments were first disclosed on the incident page six hours after automated payments had been suspended, placing the disclosure in the middle of Tuesday afternoon. Employers checking the page earlier in the day would have understood the incident as a slowdown, not as unauthorized outgoing payments from their own accounts.
## Most of the money is back
By Wednesday evening, $2.3 million of the $2.8 million had been recovered. The remaining roughly $500,000 was still pending.
Because no funds were withdrawn from employees' bank accounts, the financial exposure sat entirely with the 91 corporate customers whose reimbursement budgets absorbed the duplicate transfers.
Chief executive Lena Ortiz apologized for the incident and said the company would publish a technical report. That report will be the first opportunity for KiteBridge to explain how a product carrying an advertised exactly-once guarantee issued the same reimbursement thousands of times — and why its public description of the failure initially pointed in the opposite direction.
Anthropic: Claude Sonnet 5.5: # KiteBridge Software Error Sent $2.8 Million in Duplicate Reimbursements to 91 Companies
**A payment glitch at the expense-management firm left employers funding 13,400 duplicate payments. About $500,000 had not been recovered by Wednesday evening.**
A software failure at KiteBridge, which makes expense-management software, produced 13,400 duplicate employee reimbursements worth $2.8 million. The error hit 91 corporate customers. By Wednesday evening, KiteBridge had recovered $2.3 million, and the remainder was still pending.
Employees did not lose money. No one had funds withdrawn from their accounts, because employers paid for the duplicate reimbursements. The cost fell on the companies that use KiteBridge's platform.
## What happened
KiteBridge deployed version 4.2 of its software at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m.
The problem began when a payment processor, the outside company that moves money between accounts, timed out. A timeout means the processor did not respond in the expected window, leaving KiteBridge's system unsure whether a payment had gone through.
KiteBridge's fallback system is meant to handle such failures, typically by trying again. In this case it treated the retries as new reimbursement requests rather than repeat attempts at the same payment. Each retry became a fresh payment, and the software generated thousands of duplicates.
KiteBridge suspended automated payments at 10:03 a.m., about 78 minutes after the deployment.
## A safeguard in question
The failure stands out because of what KiteBridge had said about its own product. An April changelog, the public list of software updates, stated that version 4.0 had introduced "exactly-once payment protection." The term describes a safeguard designed to ensure a payment request is carried out one time, even if the system has to retry it.
Duplicate payments of this scale are the outcome that kind of protection is meant to prevent. KiteBridge has not yet explained how the failure got past it.
## Communication under scrutiny
KiteBridge's public incident page initially described the problem as delayed reimbursements. It first disclosed the duplicate payments six hours after automation was suspended.
The gap matters because customers needed to know that money had been sent in error, not merely that it was late. Finance teams at affected companies had to identify the extra payments and work toward recovering them.
## Recovery and next steps
By Wednesday evening, KiteBridge had recovered $2.3 million of the $2.8 million, or about 82 percent. The remaining amount was pending.
CEO Lena Ortiz apologized and said the company would publish a technical report. That report is likely to be the first detailed account of why the fallback system mishandled the retries and why the earlier protection did not stop the duplicates.
Until then, several questions remain open. KiteBridge has not said when automated payments will resume, how the remaining money will be recovered, or whether affected customers will be compensated for any costs beyond the duplicate payments themselves.
OpenAI: GPT-6 Luna: Model A better balances the consequences, technical explanation, and chronology, while meeting the requested article length. Model B is clear, but adds unsupported claims about unresolved customer questions and recovery costs, and its ending is more speculative. (Second judge pass, order swapped — scores are the average of both: Model A has stronger narrative momentum and gives the failure a clearer technical explanation while preserving the main sequence and consequences. Model B meets the task | Anthropic: Claude Fable 5.1: Model A opens with a sharper consequence-first…
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: # Startup Spotlight: LedgerLark
## The insight came from a three-day cleanup
LedgerLark, which builds AI agents for post-trade operations at small investment firms, traces its origin to an incident rather than a market map. Co-founder Mina Batra, who spent seven years in operations at a regional brokerage, watched a six-person team spend three days resolving a trade mismatch caused by a copied account number. The error was trivial. The search for it was not.
That gap — between how small a mistake can be and how much labor it can absorb — is the company's premise. Batra's co-founder, Owen Carr, built reconciliation systems at a large bank, which gives the team an unusual pairing: one person who has lived the queue, another who has built the machinery meant to shrink it.
The implicit bet is that reconciliation at small firms is less a software problem than an attention problem. Large institutions buy expensive platforms and staff around them. Smaller firms often improvise with spreadsheets, email threads, and institutional memory. LedgerLark is wagering that agents can occupy that middle ground.
## Narrow by design
The product compares trade confirmations, custodian records, and internal ledgers, then proposes corrections for a human to approve. It does not execute trades or move money.
That boundary matters more than it might sound. It keeps the system out of the transaction path, which changes the risk conversation a compliance officer has to have. It also caps the ambition: LedgerLark is selling faster resolution, not automated operations. Every proposed fix still needs a person to look at it, so the value depends on how often the proposals are right, how quickly a reviewer can judge them, and whether reviewers start rubber-stamping suggestions once they trust the system. Those are open questions, not settled ones.
## Early traction, carefully read
Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The company has raised a $3.5 million seed round led by Motive Row.
The pricing spread suggests the product is being scoped to something like exception volume or operational complexity rather than sold as a flat tool — a reasonable approach when a five-person operations desk and a fifty-person one have different problems. It also implies the sales motion is consultative, which is slow.
LedgerLark says one customer cut its unresolved daily exceptions from 70 to 18. That is a striking number, and it deserves the qualifiers attached to it: it is company-reported, it describes one customer, and no baseline period, timeframe, or independent verification has been disclosed. Exception counts can also fall for reasons unrelated to software — a quieter trading month, a cleaned-up reference data set, a new hire. Five customers is enough to show that someone will pay. It is not yet enough to show what happens across a full market cycle, a custodian migration, or an audit.
## The competition is already installed
LedgerLark's rivals are not other startups. They are established reconciliation vendors with years of custodian integrations and audit history, and internal software that banks and larger brokers built themselves and consider adequate. Displacing either requires proving that the incumbent is not just worse but worth the switching cost — a higher bar than being newer.
Then there is procurement. Financial institutions may require lengthy security and procurement reviews, and a seed-stage vendor with five references has limited leverage to shorten them. Long sales cycles are survivable; long sales cycles that consume most of a $3.5 million round are a different problem.
## What to watch
The interesting question is not whether agents can spot a mismatched account number. It is whether the trust required to accept their corrections accumulates fast enough to justify the contract — and whether small firms, once served well, turn out to be a durable market or a stepping stone to bigger, slower buyers. LedgerLark has a credible founding story and a deliberately narrow product. The next evidence worth waiting for is a second and third customer reporting the same kind of number.
Anthropic: Claude Sonnet 5.5: # Startup Spotlight: LedgerLark Points AI Agents at the Back Office
*A young company is betting that small investment firms need help with post-trade errors that are tedious, costly and easy to miss.*
## A mismatch that took three days
Mina Batra spent seven years in operations at a regional brokerage. She began exploring the idea for LedgerLark after watching a six-person team spend three days resolving a trade mismatch. The cause was a copied account number.
The insight is less about any single error than about the labor around it. Post-trade work involves confirming that a firm's records agree with those of custodians and counterparties. When they don't, people trace the discrepancy by hand across documents and systems. Smaller firms may have fewer people to absorb that work, and a few bad days can crowd out everything else.
Batra's co-founder, Owen Carr, built reconciliation systems at a large bank. The pairing gives the company experience on both the operational and the technical sides of the problem.
## What the product does
LedgerLark's agents compare three sources: trade confirmations, custodian records and internal ledgers. When they find a discrepancy, they propose a correction, and a person approves it before anything changes.
The product does not execute trades or move money. That boundary limits what the software can do, but it also limits the damage a bad suggestion can cause. The design shifts the risk to the review step. Human-approval systems work only if reviewers keep scrutinizing proposals as volume grows, and that is something customers will need to watch.
## Early traction
Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. The company has raised a $3.5 million seed round led by Motive Row.
LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. That is an encouraging figure, but it comes from the company and describes one customer. It doesn't say over what period the drop occurred, how exceptions were counted, or whether the customer's trading volume changed in the meantime. Until more customers report comparable results, it is best read as a promising data point rather than a benchmark.
## Competition and open questions
LedgerLark is entering a field with established players. Reconciliation vendors already sell to financial institutions, and some firms rely on software built in-house, including internal bank systems. Those incumbents have existing integrations, procurement relationships and track records, which are hard advantages for a newcomer to match.
The startup's case rests on a narrower claim: that AI agents can handle the messy exceptions, the ones that conventional matching leaves to people, well enough to change a small firm's workload. That is testable, but five customers is too few to show how the product performs across different custodians, data formats and kinds of errors.
Sales may also move slowly. Financial institutions can require lengthy security and procurement reviews, and software that touches confidential records will draw close scrutiny. Long review cycles can delay revenue and stretch a seed round further than planned. They also give an edge to vendors that have already cleared the same hurdles.
## What to watch
Three things will say more than the headline numbers:
- **Renewals.** The first annual contracts will show whether customers keep paying once the novelty fades.
- **Consistency.** If other customers see exception counts fall the way the first did, the result starts to look like a pattern.
- **Rejection rates.** How often humans reject or edit the agents' proposed corrections will show how reliable the suggestions are.
LedgerLark has picked a real and unglamorous problem, and its decision not to touch trades or money is sensible. Whether its agents earn a place alongside entrenched competitors is still an open question.
OpenAI: GPT-6 Luna: Model A offers the stronger narrative and more probing analysis while still treating the traction claim and sales risks with appropriate caution. Model B is somewhat more disciplined about unsupported inferences and adds useful metrics to watch, but Model A’s richer explanation better fulfills the spotlight brief. (Second judge pass, order swapped — scores are the average of both: Model B is more disciplined about the supplied facts while still addressing traction, competition, and uncertainty. | Anthropic: Claude Fable 5.1: Both are accurate, well-hedged and appropriately…
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
Anthropic: Claude Sonnet 5.5: C
Scored against the benchmark's reference answer (choice match): Model A was correct, Model B was correct.
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 Customer Prompt Excerpts Open to the Web for Four Hours
*Logs show excerpts from 1,842 user sessions were reachable without a login on Monday. Cloudnote says it has found no evidence anyone but the researcher who reported the flaw viewed them — and its investigation is unfinished.*
Cloudnote, a note-taking service with 3.2 million registered users, said a debugging page displaying excerpts of customer prompts was accessible without authentication for four hours on Monday. Company logs show excerpts from 1,842 user sessions were exposed. Cloudnote has not said how many distinct accounts those sessions represent.
Each excerpt could contain up to 200 characters of a prompt — enough to include a name, a client or the opening of a private question, depending on what a user typed. Full documents, passwords and payment information were stored elsewhere and were not present on the page, the company said.
The exposure ended 19 minutes after an outside researcher notified Cloudnote, which removed the page. The company said it has found no evidence that anyone other than that researcher accessed it. That finding is preliminary: the investigation is not complete, and Cloudnote has not said whether its logs would record every visit.
Still unresolved is how an internal debugging tool became reachable without a login, and whether the affected sessions belong to paying customers, free accounts or company testers. Cloudnote has not said when it will notify affected users or release its findings.
Anthropic: Claude Sonnet 5.5: # Cloudnote exposed excerpts of user prompts on an unprotected debugging page for four hours
*The company says it has found no sign that anyone but the researcher who reported the problem viewed the page, but its investigation is not finished.*
Cloudnote said a debugging page containing excerpts of customer prompts was accessible without authentication for about four hours on Monday. The company, which has 3.2 million registered users, has removed the page.
Logs show the page held excerpts from 1,842 user sessions. Each excerpt could include up to 200 characters of a prompt, so some entries may have been fragments of longer messages. Prompts can contain personal or work-related details, which is why even partial excerpts matter.
Cloudnote took the page down 19 minutes after a researcher notified it. The company says it has found no evidence that anyone else accessed the page. Because the investigation is incomplete, that conclusion could change.
According to Cloudnote, full documents, passwords and payment information were stored elsewhere and were not on the page.
Several questions remain unanswered. It is not yet clear how many individual users are affected, since the figure counts sessions rather than people. Nor is it clear how the page became publicly reachable, or whether affected customers will be notified.
OpenAI: GPT-6 Luna: Model A better meets the requested length in the paragraphs themselves and more fully explains why brief prompt excerpts matter while distinguishing confirmed facts from unknowns. Model B is clear, but its paragraphs fall short of the 180-word minimum and it softens the reported four-hour duration to “about four hours.” (Second judge pass, order swapped — scores are the average of both: Model B more carefully separates reported facts from reasonable unknowns, while Model A adds unsupported possi | Anthropic: Claude Fable 5.1: Model A delivers a tighter, more vivid opening…
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 agents
Arcwell has launched Relay, an artificial intelligence platform that reads customer support conversations and suggests replies to the agents handling them. In a six-month trial with 12 companies, the company says, agents using the system closed tickets 14% faster.
Relay does more than draft answers. It also flags customers who appear likely to cancel and generates summaries of support interactions, positioning it as both a writing aid and an early-warning system for accounts at risk. Arcwell did not release customer-retention results from the trial, so the evidence it has put forward so far concerns speed rather than the churn predictions it is also selling.
The pitch is that the answers already exist inside a company and take too long to find. "Support agents spend too much time searching for information that their companies already have," said CEO Priya Shah.
Relay costs $65 per agent each month.
That help comes with a data requirement. To work, Relay needs access to a company's support transcripts and its internal help documents — the conversations customers have already had and the material agents consult to resolve them. Arcwell says customer data is encrypted and is not used to train shared models. Companies can also choose how long that material stays in the system, selecting a retention period of either 30 or 90 days.
Those terms are likely to matter as much as the performance claim. Support transcripts are among the more sensitive records a company holds, and handing them to an outside platform is a decision that typically involves security and legal teams, not just the support manager who wants faster ticket handling. Encryption, a promise not to pool customer data into shared models, and a choice of retention windows are the concessions Arcwell is offering to make that easier to approve.
What the trial establishes is narrower: 12 companies, six months, and a 14% improvement in how quickly agents closed tickets, as measured and reported by Arcwell. Whether Relay's churn flags actually keep customers from leaving remains an open question, and one the company has yet to answer with numbers.
Anthropic: Claude Sonnet 5.5: # Arcwell Launches Relay, an AI Tool That Suggests Answers to Support Agents
Arcwell has released Relay, a platform that uses artificial intelligence to analyze customer support conversations and suggest answers to agents.
Relay also identifies customers who may cancel and creates summaries. Arcwell is pitching it as a fix for a simple problem. "Support agents spend too much time searching for information that their companies already have," said CEO Priya Shah.
## Speed gains, but no retention data
Twelve companies tested Relay over six months. Arcwell says agents using it closed tickets 14% faster. The figure is the company's own, and it measures speed alone.
It is also the only performance result Arcwell has made public. It speaks to Relay's first job, suggesting answers, and says nothing about the other two. Arcwell did not release customer-retention results, so it is unclear whether Relay's ability to flag customers who may cancel actually keeps any of them.
Relay costs $65 per agent each month. At that rate, a team of 10 agents would pay $650 a month.
## Access and data handling
Shah's comment frames the problem as one of retrieval rather than missing knowledge: the information exists, and agents lose time finding it. Relay's design follows from that premise. It requires access to a company's support transcripts and internal help documents, the same material that lets it suggest answers in the first place.
That access is the trade-off for buyers. 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, but those are the only two options.
For now, then, Relay arrives with one measured benefit, a published price and a set of data assurances that come from the company itself. Whether its cancellation alerts and summaries deliver comparable value is not yet known.
OpenAI: GPT-6 Luna: Model B stays closer to the supplied facts and meets the requested length with a focused account of the trial, price and data practices. Model A is polished, but adds broader claims about sensitivity and approval processes and repeats the trial findings in its conclusion. (Second judge pass, order swapped — scores are the average of both: Model A has stronger narrative flow, but it adds unsupported claims about transcript sensitivity, typical security and legal reviews, and Arcwell’s data practi | Anthropic: Claude Fable 5.1: Model A delivers a stronger lede, smoother tran…