Head to head: Anthropic: Claude Fable 5.1 vs MoonshotAI: Kimi K3

Anthropic: Claude Fable 5.1: # Startup Spotlight: LedgerLark Wants to Make Trade Breaks Boring *AI agents for the unglamorous work of post-trade operations at small investment firms* ## The problem that started it Mina Batra spent seven years in operations at a regional brokerage. The moment she points to as the origin of LedgerLark was not a dramatic failure but a tedious one: a six-person team spent three days untangling a trade mismatch that traced back to an account number someone had copied incorrectly. Nothing about that episode was unusual, which is the point. Post-trade operations — matching confirmations against custodian records and internal ledgers, chasing down the discrepancies known as "breaks" — is repetitive, high-stakes work that small firms often handle with spreadsheets, email threads and institutional memory. Batra's co-founder, Owen Carr, saw the other end of the market, building reconciliation systems at a large bank with resources a 20-person asset manager will never have. ## What the product does LedgerLark's agents ingest confirmations, custodian records and a firm's internal ledger, compare them, and surface mismatches along with proposed corrections. A human operations staffer then approves, edits or rejects each suggestion. The constraints are as notable as the capabilities. The agents do not execute trades and do not move money. That design choice narrows the product's ambitions, but it also narrows the blast radius if the software is wrong — a reasonable trade-off in a domain where an incorrect fix can be worse than an unresolved one. It is also, presumably, easier to explain to a compliance officer. Whether "propose, then approve" remains the model as customers grow comfortable, or whether pressure builds to let agents close routine breaks on their own, is one of the more interesting open questions about the company. ## Early traction Five firms are currently paying for the product. Annual contracts run from $36,000 to $110,000, which suggests LedgerLark is pricing against the cost of staff time rather than against software budgets. The company says one customer cut its unresolved daily exceptions from 70 to 18. That is a meaningful figure if it holds up, though it is a single customer's result reported by the vendor, and it does not indicate how the remaining 18 compare in difficulty to the original 70. Anyone evaluating the product would want to know whether the agents are clearing the easy breaks and leaving the hard ones, or genuinely reducing the workload across the board. LedgerLark has raised a $3.5 million seed round led by Motive Row. ## The harder parts Reconciliation is not an empty field. Established vendors have sold matching and exception-management software to financial firms for decades, and larger institutions run internal systems built by teams like the one Carr worked on. LedgerLark's bet is that the small-firm segment is underserved by both — too small for enterprise vendors to court aggressively, too resource-constrained to build in-house — and that an agent-based approach can handle the messy, unstructured parts of the job that older rules-based systems hand back to humans. That bet may be right. It still has to survive the sales process. Financial institutions, even small ones, tend to run lengthy security and procurement reviews before letting outside software near their books. For a seed-stage company with five customers, each of those reviews consumes time and attention that is hard to spare. Five paying firms is a real signal; it is not yet evidence that the sales cycle is repeatable. There is also the question of what happens when an agent proposes a wrong correction and a tired analyst approves it at 6 p.m. The human-in-the-loop model shifts accountability but does not eliminate error, and how LedgerLark handles that first serious incident will say more about the company than any reduction in exception counts. ## What to watch The founders' backgrounds give them a plausible claim to understanding the work. The product's restraint gives them a plausible path through compliance. The next year should show whether a small firm's operations team, having tried the product, keeps paying for it — and whether a second and third cohort of customers arrives faster than the first.

MoonshotAI: Kimi K3: # Startup Spotlight: LedgerLark Takes AI Agents Into the Back Office *The startup's software hunts down trade mismatches and proposes fixes — but deliberately keeps humans in charge of the money.* **LedgerLark at a glance** - **Product:** AI agents for post-trade reconciliation at small investment firms - **Founders:** Mina Batra and Owen Carr - **Customers:** Five paying firms - **Pricing:** $36,000–$110,000 in annual contracts - **Funding:** $3.5 million seed round led by Motive Row Post-trade operations rarely make headlines, which is part of the problem Mina Batra wants to solve. After seven years in operations at a regional brokerage, she knew the work well: the quiet, meticulous process of making sure that what a firm traded, what its custodian recorded and what its own books say all match. Often they don't, and finding out why can swallow days. The episode that pushed her toward founding LedgerLark was almost comically mundane. A six-person team spent three days resolving a single trade mismatch that was eventually traced to a copied account number. No fraud, no system failure — just a transcription error with an outsized cleanup cost. Her co-founder, Owen Carr, had watched the same pattern from a different vantage point, building reconciliation systems at a large bank. Their shared read: the industry has plenty of tools for catching discrepancies, but small investment firms often lack the staff and budget to act on them quickly. ## What the agents actually do LedgerLark's software compares three sources of record — trade confirmations, custodian reports and internal ledgers — and flags where they diverge. Rather than simply surfacing exceptions for an analyst to chase, its agents draft proposed corrections, which a human then reviews and approves. Two design choices stand out. The product does not execute trades or move money, full stop. And nothing changes in a client's books without a person signing off. Those constraints likely make the company an easier sell to compliance-minded buyers, though they also cap how much labor the tool can realistically remove. The value proposition rests on the agents' proposals being right often enough that review is fast — a claim each customer will test for itself. ## Early traction, with caveats LedgerLark says five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. That is a modest base, but the spread suggests the product is being bought by firms of different sizes rather than a single friendly pilot customer. The company also reports that one customer cut its unresolved daily exceptions from 70 to 18. That is a striking figure — and an unaudited one, from a single deployment. Whether it replicates across the other four customers will say more about the product than the headline number does. Investors have shown early interest: LedgerLark recently closed a $3.5 million seed round led by Motive Row. ## The harder questions Competition comes from two directions. Established reconciliation vendors have decades of integrations and the trust of risk-averse buyers; large banks have internal software teams that can build their own tooling. LedgerLark's answer appears to be focus — small firms that incumbents underserve — but focus is a strategy, not a moat, and an incumbent adding agent-style features is a plausible response. There is also the calendar to contend with. Financial institutions tend to require lengthy security and procurement reviews before new software touches client data, which can stretch sales cycles well past what a seed-stage runway prefers. Selling to smaller firms may shorten that process, but it rarely eliminates it. None of this diminishes the underlying observation, which rings true: a large share of back-office cost comes from small errors that are tedious to find. The open question is whether LedgerLark's agents can earn enough trust — one approved correction at a time — before slower competitors notice the same opening.

Anthropic: Claude Fable 5.1: # Cloudnote Says Debugging Page Exposed Excerpts From 1,842 User Prompts *The page was reachable without a login for four hours Monday; the company says it has found no sign anyone but the researcher who flagged it looked.* Cloudnote said an internal debugging page containing excerpts of customer prompts was accessible without authentication for about four hours on Monday, exposing fragments of text from 1,842 user sessions. Each excerpt could include up to 200 characters from a prompt, according to the company. Full documents, passwords and payment information are stored on separate systems and were not present on the page, Cloudnote said. The company has 3.2 million registered users, so the affected sessions represent a small fraction of its base. A security researcher notified Cloudnote of the exposure, and the company said it removed the page 19 minutes later. Cloudnote said a review of its logs has turned up no evidence that anyone other than the researcher accessed the page while it was open. That finding is preliminary. Cloudnote said its investigation is not complete, and several questions remain open, including how the page became reachable, which users were affected and whether those customers have been or will be notified. Even short fragments can be sensitive. Two hundred characters is enough to capture a name, a question about a medical or legal issue, or the first lines of a private message. What the excerpts actually contained — and whether anyone saw them — will determine how serious the incident turns out to be.

MoonshotAI: Kimi K3: # Cloudnote Debugging Page Left Customer Prompt Excerpts Publicly Accessible for Four Hours *Prompt fragments from 1,842 user sessions were visible without a password on Monday. The note-taking company says it has found no evidence anyone beyond the researcher who flagged the page viewed it — but its investigation is not finished.* Cloudnote, a note-taking company with 3.2 million registered users, said a debugging page containing excerpts of customer prompts was accessible on the public internet without authentication for four hours on Monday. The company removed the page 19 minutes after a security researcher alerted it to the exposure. Logs show the page held prompt excerpts from 1,842 user sessions, each capped at 200 characters, according to the company. Full documents, passwords and payment information were stored elsewhere and were not present on the page. Because one person can generate multiple sessions, the number of individuals affected is not yet known. Even so, the exposure matters. Prompts can contain personal details, drafts or work material users never intended to share, and even a 200-character fragment may be revealing. The episode also raises a basic question: why was a debugging tool containing live customer data reachable by anyone at all? Cloudnote said it has found no evidence that anyone besides the researcher accessed the page — a conclusion it acknowledged is preliminary, since its investigation remains underway. The company has not said how the page became public, whether it will notify the users whose prompts appeared, or when the inquiry will conclude.

Anthropic: Claude Fable 5.1: # QueueZero's New AI Meeting Assistant Turns Offhand Remark Into Tasks for Its Own CEO QueueZero introduced Proxy, an AI meeting assistant, on Wednesday, pitching it as a tool that can sit in on meetings, write up notes and hand off follow-up work to software agents without human intervention. The launch demonstration also offered an early look at how literally the system can take what it hears. During the presentation, a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." The comment referred to QueueZero chief executive Evan Wu, who was leading the demonstration. Proxy, which was running live during the event, interpreted the remark as an instruction. Within moments, the assistant had generated two tasks and assigned both to Wu. The first read "Stop talking." The second read "Send pricing sheet." Wu laughed, completed the second task and deleted the first. The demonstration then continued. QueueZero later characterized the episode as an example of the product functioning as designed rather than a malfunction. According to the company, Proxy requires explicit confirmation from a user before it sends any message to an external recipient, a safeguard intended to prevent the assistant from acting on misheard or ambiguous statements in ways that reach customers or partners. Internal tasks, however, can be created automatically once Proxy identifies what it believes to be an action item. The company said administrators can disable automatic task creation entirely, requiring every proposed task to be reviewed before it is added to a team's workload. QueueZero did not say whether it plans to adjust how Proxy distinguishes between directives and commentary, or whether the default settings would change before the public beta. The distinction between an instruction and an observation has long been a difficult problem for voice- and language-driven software. Meeting transcripts are full of statements that sound like directives but are not meant as such, and assistants that parse them too eagerly risk cluttering task lists or, in this case, assigning work to the wrong person. Assistants that are too conservative, by contrast, miss genuine commitments and require manual cleanup, reducing the value of automation in the first place. Proxy's broader function is to act as a participant rather than a transcriber. Beyond producing notes, the assistant can route follow-up work to other software agents, such as drafting documents, scheduling meetings or updating project-tracking systems. QueueZero has positioned the product for teams already using agent-based tools and looking to connect them to meetings. Proxy will cost $30 per user per month. It enters public beta next week. Wu, for his part, left the launch with one task fewer than the software had intended.

MoonshotAI: Kimi K3: # QueueZero's AI Meeting Assistant Takes a Quip Literally, Assigns Tasks to Its Own CEO QueueZero on Wednesday launched Proxy, an artificial intelligence assistant that can attend meetings, produce notes and assign follow-up work to software agents. The product's first public demonstration also became an unscripted test of a pressing industry question: when, exactly, should an AI assistant act on what it hears? During the launch event, as QueueZero CEO Evan Wu presented the product onstage, a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was monitoring the meeting, interpreted the remark as an instruction. It created two tasks assigned to Wu: "Stop talking" and "Send pricing sheet." Wu laughed when the tasks appeared, completed the second and deleted the first; the pricing sheet reached attendees during the event. Proxy, for its part, ended the demonstration with a 50 percent completion rate on executive directives. The moment drew laughter in the room, but it underscored a central challenge for the fast-growing category of "agentic" AI — software designed not just to summarize information but to take action on a user's behalf. An assistant that turns a joke into a task in a controlled demo could, in a less forgiving setting, turn a stray comment into a consequential one. QueueZero said Proxy requires user confirmation before any external message is sent, so the assistant could not have distributed the pricing sheet — or anything else — outside the organization on its own. Internal tasks, by contrast, can be created automatically, which is what happened onstage. Administrators can disable automatic task creation entirely if they prefer tighter control. Proxy costs $30 per user per month and enters public beta next week, the company said. The launch comes as workplace software vendors race to move AI meeting tools beyond transcription and note-taking toward products that schedule work, draft messages and execute multi-step processes. That shift has raised the stakes on guardrails, since mistakes compound quickly when software acts rather than suggests. For QueueZero, the episode functioned as an accidental live demonstration of both halves of its pitch: an assistant attentive enough to catch an offhand request, and guardrails that determine how far it can go with one. Wu resumed his presentation after deleting the first task; the second was completed before the event ended.

Anthropic: Claude Fable 5.1: **6**

MoonshotAI: Kimi K3: 6

Anthropic: Claude Fable 5.1: # Foundry Hall to Close After Seven Years as Milwaukee's Startup Launchpad Foundry Hall will shut its doors on October 31, ending a seven-year run as the nonprofit accelerator that gave Milwaukee's early-stage founders a place to work, pitch, and fail without paying for the privilege. The decision came down to arithmetic. Blank River Bank, the organization's largest corporate sponsor, supplied 38 percent of Foundry Hall's annual budget. When the bank declined to renew, the remaining funds were enough to keep the lights on through October — but not enough to run another full program cohort. "We kept waiting for the replacement check that never came," said executive director Camille Ross. The bank, in a statement, attributed the withdrawal to shifting community investment priorities and emphasized that it had nothing to do with Foundry Hall's performance. By most measures, that performance had been strong. Since opening in 2018, the accelerator has supported 212 startups, which together now employ 1,140 people — a figure Ross cites without elaboration, as though the number speaks for itself. ## What's lost Six staff members will lose their jobs. Twelve startups currently working out of the building must find somewhere else to go. Dev Malik, who founded his company inside Foundry Hall, described what the space meant in practical terms. "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 trajectory — from borrowed table to real conference room — was the point. Foundry Hall never promised capital. What it offered was legitimacy, an address, a room where a founder with a laptop and an idea could look like a business rather than a hobby. The twelve companies still working there are now in the position Malik was in before he arrived. Several have active fundraising rounds underway. Others are mid-build on products with launch dates that don't move just because the landlord does. ## The city's response Deputy Mayor Susan Cole said the city is in conversations with local universities and commercial building owners about finding space for the displaced companies. No agreements have been announced, and Cole did not offer a timeline. Ross has been careful not to frame the closing as a betrayal. Corporate sponsorships end. Budgets get reallocated. Still, the speed of it — one decision, one line item, and an organization that outlasted a pandemic disappears in a single fiscal year — has left her and her staff with little time to grieve what they built. The building itself will likely be leased to a new tenant. Whether it will be another place where someone's first investor meeting happens is, for now, nobody's decision but the market's. On October 31, the last cohort will pack up. Foundry Hall will become one more entry in the long ledger of things Milwaukee tried, briefly succeeded at, and then let go.

MoonshotAI: Kimi K3: # Foundry Hall, Milwaukee Startup Accelerator, to Close After Losing Its Largest Sponsor **MILWAUKEE —** Foundry Hall, the nonprofit startup accelerator that has occupied a converted industrial building here since 2018, will close on October 31, ending a run its leaders once hoped would be permanent. When it opened, the accelerator was part of a broader push to keep young companies from leaving Milwaukee for larger cities. It offered startups workspace, mentorship and a path to investors they could not easily reach on their own. For six years, it paid for all of it with a budget pieced together from grants, program fees and a small circle of corporate sponsors. The numbers it leaves behind are considerable: 212 startups supported, companies that collectively employ roughly 1,140 people. But the figure that decided its fate was 38 — the percentage of its annual budget supplied by Blank River Bank, its largest corporate sponsor. This year, the bank declined to renew. A bank spokesperson said the decision reflects a change in community investment priorities and is unrelated to Foundry Hall's performance. Whatever the reason, the loss left a gap the accelerator could not close. "We kept waiting for the replacement check that never came," said Camille Ross, Foundry Hall's executive director, who spent months courting new funders. The organization has enough money to operate through October but not enough to commit to another full program cycle. Six employees will lose their jobs. Twelve startups now working in its space must find somewhere else to go — a scramble none of them planned for. Ross said the organization chose to announce the closure now to give both groups time to plan. ## "Pitching from the back table of a coffee shop" For the founders who passed through its doors, the loss is harder to quantify. Dev Malik credits Foundry Hall with turning his idea into a company. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." Malik's company now has offices of its own. He has not forgotten the coffee shop. ## What Comes Next City officials are trying to ease the landing. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the twelve displaced companies. Keeping those companies in Milwaukee, she said, is a priority. Ross will spend the next two months winding down operations — helping her staff look for work and her tenants look for offices. When the doors lock on October 31, Milwaukee will still have the 1,140 jobs that trace back to Foundry Hall. What it will lose is harder to count. The back tables of the city's coffee shops may soon be full again.

Anthropic: Claude Fable 5.1: # HelioDock Takes Warehouse Robot Oversight Off-Site With Night Shift Launch *The autonomous-robot maker also closed a $24 million Series B led by North Coast Ventures, bringing its total raised to $39 million.* HelioDock, which builds autonomous warehouse robots, on Thursday released Night Shift, software that lets operators monitor and intervene in robot fleets without being inside the building. The company announced a $24 million funding round alongside the launch. The product is aimed at a recurring headache for warehouses that run robots: when a machine stalls, a person has to sort it out, frequently in person and frequently at inconvenient hours. Night Shift moves that work to a remote console. It will cost $90 per robot per month and become generally available September 15. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," HelioDock chief executive Mara Velez said. Before the public launch, Night Shift spent 14 months in a private beta that covered nine logistics companies and 310 robots. HelioDock says customers in that program experienced 17% fewer unplanned robot stoppages. The figure is the company's own and has not been independently verified. The beta also exposed the software's most significant constraint. Night Shift depends on a working network connection to the robots it oversees, and that connection is not something HelioDock controls. Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. When the link drops, Night Shift's remote-control functions become unavailable and the robots revert to whatever behavior has been configured locally on each machine. A robot that needs a human decision cannot get one remotely until the network comes back, which leaves an operator at a poorly connected site in roughly the position the software was designed to eliminate. How much that matters will vary by facility. A warehouse with reliable wired or cellular service may rarely see the fallback kick in. An operator running sites in areas with patchy LTE coverage, as Pike's experience indicates, gets the benefit of remote monitoring only as long as the signal holds. Night Shift enters a field that already has established options. Relay Robotics and YardOS both sell fleet-management software for warehouse robots, and HelioDock will be competing with them for operators who want a single view of their machines. The timing of the launch alongside a funding announcement gives HelioDock a chance to put both pieces of news in front of prospective customers at once. The Series B, led by North Coast Ventures, raises the company's total funding to $39 million. HelioDock did not detail how the new capital will be allocated. For now, the clearest signal of what the company thinks it has built comes from the pricing model. Charging per robot per month ties HelioDock's software revenue directly to the size of the fleets its customers deploy, and it sets up Night Shift as an ongoing service layered on top of the hardware rather than a one-time add-on. Whether operators judge $90 a month per machine as worth it will depend in part on how often their robots get confused by pallets, and in part on whether the network is up when it happens.

MoonshotAI: Kimi K3: # HelioDock Launches Remote Monitoring for Warehouse Robot Fleets, Lands $24 Million Series B *Night Shift lets operators keep watch over autonomous robots from anywhere — as long as the network connection holds.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor and control their robot fleets remotely instead of dispatching staff to the warehouse floor whenever a machine stalls. The launch caps a 14-month private beta spanning nine logistics companies and 310 robots, and HelioDock paired it with news of a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. The company is betting that operators running large fleets will pay a per-robot subscription to keep watch remotely — and to stay home when something goes wrong overnight. The software targets an awkward gap in warehouse automation: fleets are built to run unattended, yet a single stuck robot has still required a person on site to assess it. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," said Mara Velez, HelioDock's chief executive. Night Shift is the company's attempt to make that drive unnecessary, pairing remote monitoring with remote control when the connection holds. According to HelioDock, customers in the beta program experienced 17 percent fewer unplanned robot stoppages while using the software. The company shared the figure as part of Thursday's announcement; it is company-supplied and has not been independently verified. Still, the beta was substantial in scope: nine operators running 310 machines through live logistics environments for well over a year, a longer test runway than most robotics software gets before release. Unplanned stoppages are the events Night Shift is built to catch: a robot halts mid-task — confused, in Velez's telling, by something as ordinary as a pallet — and sits idle until someone steps in. Across a fleet of hundreds, those interruptions accumulate. Reducing them, even by a fraction, translates directly into more hours of productive machine time. At least one beta participant offered a qualified review. Aaron Pike of Fulcrum Logistics, one of the nine companies in the program, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His account points to Night Shift's central dependency. Oversight travels over the network, so when connectivity fails, the link between operator and fleet is severed: robots return to their locally configured behavior and remote control becomes unavailable. The machines do not stop — they keep running under whatever settings were configured on site — but the remote oversight that is the product's core function goes dark until service resumes. In other words, the robots' autonomy is local, but the operator's window into it is not. That makes Night Shift's usefulness a function of network quality: a manageable constraint in a well-connected urban warehouse, a more pressing one at the rural sites Pike described. For operators with facilities in coverage-poor regions, the fallback behavior will shape how much of the product's promise materializes day to day. Night Shift becomes generally available on September 15, opening the software to operators beyond the nine beta companies. It will cost $90 per robot per month — pricing that scales with fleet size. A customer running 100 machines would pay $9,000 a month, or $108,000 a year, for the oversight. HelioDock is not entering an empty category. Relay Robotics and YardOS both offer competing fleet-management products, so the company must convince operators that its version — and its beta numbers — are worth switching to. September's debut will provide the first broad test.

Anthropic: Claude Fable 5.1: # The Cheapest Way to Match the Best Support Model Costs 61% Less The most expensive model in a new customer-support benchmark did not win by a meaningful margin. It barely edged out a competitor that costs less than half as much and answers faster. For teams deciding which model to put in front of real customers, that is the headline, and it is worth understanding both why the result is credible and where it stops applying. ## What was tested Researchers ran five language models against 200 anonymized customer-support tickets drawn from a single software company. Every model produced one proposed answer per ticket. Three experienced support managers then graded each answer blindly, without knowing which model had written it. Scores were a composite of three qualities a real support team cares about: whether the answer was factually correct, whether it actually helped the customer, and whether the tone was appropriate. That design matters. A blind panel of practitioners is a better proxy for production quality than automated metrics or self-reported model evaluations, because the graders are judging the same things a customer would notice. The sample size, 200 tickets, is large enough to separate models that are clearly far apart but small enough that close results should be read cautiously. ## The finding that matters Model North posted the highest average score at 84.2. Model Vale came in at 83.8. The researchers report that this gap was **not statistically significant**. On 200 tickets, a 0.4-point difference is well within the noise of how three human graders might score the same answers on a different day. What is not noise is the price. North is estimated to cost $1.90 per 1,000 tickets; Vale costs $0.74. That is a 61% reduction for what the benchmark could not distinguish from equal quality. Vale also returned answers 18% faster than North, which in a support context translates to shorter queues and quicker first responses. Put differently: if these two models were the only options, the data gives no quality-based reason to choose the pricier one for general ticket handling. The decision would rest on cost and latency, and both favor Vale. ## Where the models diverged The composite scores hide some texture. North did best on billing disputes; Vale did best on technical troubleshooting. These category-level results come from an even smaller slice of the 200 tickets, so they should be treated as hints rather than conclusions. But they suggest a plausible operational strategy: a team heavy on account and payment questions might lean toward North, while a team fielding mostly technical issues has even less reason to pay the premium. The three cheaper models, Ember, Kite and Harbor, formed a clear second tier. Ember scored 79.1 at $0.31 per 1,000 tickets, Kite 76.4 at $0.22, and Harbor 71.0 at $0.08. The drop from Vale to Ember is nearly five points, roughly twelve times the gap between North and Vale, and the drop to Harbor is thirteen points. These are differences a support manager would feel. Whether they are acceptable depends on the business: a low-stakes product with generous human review might happily run Harbor at a tenth of Vale's price, while a company where a wrong answer triggers a refund or a churned customer probably cannot. ## A shared weakness One result applied to every model regardless of price: all five performed poorly when a ticket depended on an image attachment. Screenshots of error messages, photos of damaged products, and annotated interface captures are routine in real support queues. If a meaningful share of your tickets rely on visual context, none of these models as tested will handle them well, and that gap needs a human or a dedicated workflow, not a bigger model budget. ## Limits of the benchmark The results should be read inside their boundaries. Every ticket was in English and came from one software company, so the findings may not transfer to other industries, product types or languages. The models were tested only on drafting replies; the study did not examine agentic actions, tool use, or the ability to look up account data or trigger refunds. Modern support automation increasingly relies on those capabilities, and a model that writes a good answer may not be the one that executes a good action. Cost figures are estimates, and the scoring rubric, while sensible, reflects three managers' judgment rather than measured customer outcomes. ## The practical takeaway Within those limits, the lesson is straightforward. For English-language support drafting, the second-most-expensive model matched the most expensive one in quality while costing far less and responding faster. Teams should treat the top-price option as something to justify, not the default, and should run their own tickets through at least two candidates before committing. The best model on someone else's benchmark is only a starting point.

MoonshotAI: Kimi K3: # The Best Model in the Test Wasn't the Best Buy **A benchmark of five language models on real customer-support work delivered a result that should interest anyone paying per ticket: the top-scoring model didn't meaningfully outperform its closest rival — but it cost more than twice as much and answered more slowly.** ## A tie at the top, a gap in the bill Model North earned the highest average score, 84.2, narrowly ahead of Model Vale at 83.8. The researchers found the difference — less than half a point — was not statistically significant. In practical terms, the graders could not reliably tell the two models' answers apart. The economics were anything but a tie. North's estimated cost was $1.90 per 1,000 tickets against Vale's $0.74, and Vale answered 18% faster. At a volume of one million tickets a month, that works out to roughly $1,900 versus $740 — a difference approaching $14,000 a year — for quality the test could not distinguish. Each leader also had distinct strengths: North performed best on billing disputes, Vale on technical troubleshooting. For teams whose queues skew technical, Vale's cost and speed advantage appears to come with no quality trade-off at all. ## How the test was run Researchers assembled 200 anonymized customer-support tickets and had each model produce one proposed answer per ticket. Three experienced support managers graded every response blindly — without knowing which model wrote it — using a combined measure of factual correctness, usefulness, and tone. Blind grading by domain veterans matters here. Support managers know what a deployable answer looks like, and hiding model identities removes the brand halo that can distort evaluations of AI output. Two caveats are worth noting. Two hundred tickets is a respectable but modest sample, and with only one answer per model per ticket, the test measures average quality rather than consistency — a model that is brilliant 95% of the time and erratic the rest would look much like a steady performer. And the non-significant result at the top cuts both ways: the data can't tell us Vale is *better* than North, either — only that they're effectively indistinguishable on this test. ## The budget tiers behave like budget tiers Below the leaders, quality stepped down alongside price. Model Ember averaged 79.1 at $0.31 per 1,000 tickets; Model Kite, 76.4 at $0.22; Model Harbor trailed at 71.0 but cost just $0.08 — roughly one twenty-fourth of North's price. These gaps — roughly three to five points per tier — are many times larger than the North–Vale split, though the benchmark only reports significance testing for the top pair, so precise confidence levels are unknown. Still, a 13-point spread between North and Harbor is the kind of difference end customers would likely notice. The curve suggests the market is behaving the way buyers would hope: each step down the ladder buys substantial savings for a visible but not catastrophic drop in quality. For high-volume, low-stakes queues — password resets, order status — the cheaper tiers may be rational. For nuanced or sensitive tickets, the savings look harder to justify. ## Every model failed on the same thing The most uniform finding had nothing to do with price. All five models performed poorly when a ticket depended on an image attachment — a screenshot of an error, a photo of a damaged item. Given how many real support tickets include exactly that kind of visual evidence, this may be the benchmark's most actionable result: no model here, at any price, earned trust with image-dependent cases. ## What this test can't tell you The findings come with firm boundaries. All 200 tickets were in English and came from a single software company, so results may not transfer across industries, languages, or brand tones. The researchers did not test agentic behavior — issuing refunds, pulling account data, calling external tools — which is where much of the real value, and risk, of support automation sits. And while the scoring was blind and expert, it remains human judgment compressed into one number; the costs, too, are estimates rather than metered billing. ## The bottom line For a buyer deciding today, the benchmark points to a clear default: Vale matched the top model on graded quality, beat it on speed, and cost roughly 60% less. North retains an edge on billing disputes that may matter for some ticket mixes. And regardless of which tier they choose, support leaders should keep humans on image-heavy queues — every model in this test showed why.

Anthropic: Claude Fable 5.1: # 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.

MoonshotAI: Kimi K3: # 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 Fable 5.1: C

MoonshotAI: Kimi K3: **C** It's the only lede that's accurate and complete. Options A and D falsely claim message contents were exposed (the facts say only workspace titles were sent, and contents were not), and D also invents "hackers" despite no evidence of malicious access. Option B buries the actual news—the security exposure—behind PR-style framing. Option C correctly captures the key facts: the product pull, what data was exposed, how many accounts, and the duration.

Anthropic: Claude Fable 5.1: # Arcwell's Relay Puts an AI Assistant Beside Every Support Agent Arcwell has launched Relay, an artificial intelligence platform that reads customer support conversations and suggests answers to the agents handling them. The suggestions are the headline feature, but not the only one. Relay also flags customers who appear likely to cancel and generates summaries of conversations, sparing agents the chore of writing up each case by hand once a ticket is closed. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. The pitch is that Relay surfaces that information at the moment an agent needs it, rather than leaving them to hunt for it across knowledge bases and old tickets while a customer waits. Arcwell tested the platform with 12 companies over six months. According to the company, agents using Relay closed tickets 14% faster. What Arcwell has not shared is whether the churn-detection feature actually kept any customers from leaving. No retention results were released, a notable gap given that identifying at-risk customers is one of the three capabilities the company is promoting. Relay costs $65 per agent per month. Data access is the obvious sticking point. To do its job, Relay needs to read a company's support transcripts and its internal help documents, which together amount to a detailed record of what customers complain about and how the business responds. That is precisely the kind of material many companies are most reluctant to hand to an outside vendor. Arcwell's answer is a set of controls rather than a workaround. Customer data is encrypted, the company says, and is not used to train models shared across its client base, so one customer's transcripts will not shape the suggestions another customer's agents see. Companies can also choose how long Relay holds their data, with retention periods of either 30 or 90 days. Those terms place Relay in a familiar position for AI tools aimed at the enterprise: the value depends on deep access to sensitive internal information, and the sale depends on convincing buyers that the access is safe. The 14% figure gives Arcwell a concrete efficiency claim to work with. The retention numbers it has chosen not to publish will be the ones prospective customers ask about next.

MoonshotAI: Kimi K3: # Arcwell Launches Relay, an AI Assistant for Customer Support Teams Arcwell has launched Relay, an artificial intelligence platform that reads customer support conversations and suggests answers to the agents handling them. It's the company's entry into a customer service industry that AI is steadily reshaping. Relay goes beyond suggesting replies. The platform also flags customers whose conversations suggest they may be preparing to cancel, and it produces summaries of support interactions. All three features target the same problem Arcwell says it set out to solve: agents hunting for information their companies already have while customers wait. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. The company has early numbers to back the pitch, at least on speed. Twelve companies tested the platform over six months, and Arcwell says agents using it closed tickets 14 percent faster. That figure comes from Arcwell itself, and the company did not release results on the metric that may matter most: whether Relay's cancellation warnings actually improved customer retention. Relay is priced at $65 per agent per month. Adopting it means granting unusual access. Relay requires a company's support transcripts and internal help documents to function — sensitive raw material for any business. Arcwell says customer data is encrypted and is not used to train shared models, and customers can choose a 30-day or 90-day retention period for their data. Whether those assurances satisfy buyers' security teams will shape the platform's early adoption. For Arcwell, Relay opens a new chapter: the company is betting that businesses are ready to rebuild customer service around artificial intelligence. For buyers, the trade-off is clear — a measurable speed gain, priced per seat, in exchange for deep data access and an unanswered question about retention. Arcwell has put a number on the first half of that equation. The second half will have to wait for data the company isn't sharing yet.

Anthropic: Claude Fable 5.1: # A Software Update Meant to Save Time Paid Out $2.8 Million Twice Ninety-one companies woke up Tuesday to find their employees had been reimbursed for the same expenses twice—13,400 duplicate payments in all, totaling $2.8 million—after a routine software update at KiteBridge, a maker of expense-management software, went wrong within half an hour of going live. By Wednesday evening, KiteBridge said it had recovered $2.3 million of the misdirected money. The remaining half-million dollars was still pending. No employee had funds pulled from a personal account; the duplicate payments were drawn from the employers who use KiteBridge to process staff expense claims. ## How it unfolded KiteBridge deployed version 4.2 of its software at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m., less than half an hour later. By 10:03 a.m., the company had suspended all automated payments—but by then, thousands of duplicates had already gone out. The failure began with a mundane problem: a timeout with KiteBridge's payment processor, the outside service that actually moves money to employees' bank accounts. Timeouts are common in financial software, and systems are built to handle them by retrying the request. The trouble was what happened next. KiteBridge's fallback system—the backup logic that kicks in when the primary payment path stalls—did not recognize those retries as repeat attempts at the same payment. Instead, it treated each one as a brand-new reimbursement request. Every time the software tried again, it created another payment. The result was a cascade of duplicates that continued until the company shut off automation. ## A promise made in April The breakdown is particularly awkward for KiteBridge because of what it told customers earlier this year. In an April changelog accompanying version 4.0, the company announced it had introduced "exactly-once payment protection"—a safeguard designed to guarantee that a single reimbursement request results in exactly one payment, no matter how many times the system has to retry. That is precisely the guarantee that failed Tuesday. KiteBridge has not yet explained whether the protection was bypassed by the fallback system, disabled in the new release, or never covered this scenario. ## A slow public accounting KiteBridge's public incident page did not initially describe what had happened. Its first update characterized the problem as *delayed* reimbursements, suggesting employees might receive money late. In fact, many were receiving it twice. The company did not disclose the duplicate payments on that page until six hours after it had suspended automation—roughly 4 p.m. Tuesday. During that window, affected employers were seeing unexpected outflows from their accounts without an official explanation of the cause. CEO Lena Ortiz has apologized for the incident and said the company will publish a technical report detailing what went wrong. ## What remains unresolved Several questions hang over the incident. The first is whether the remaining roughly $500,000 can be fully recovered; reversing payments that have already landed in employee bank accounts typically depends on the cooperation of the recipients and their banks. The second is how the exactly-once safeguard, advertised only months earlier, failed to prevent the most basic duplicate-payment scenario it was built for. The third is the gap between what KiteBridge knew internally—enough to halt payments at 10:03 a.m.—and what it told customers publicly for the following six hours. For companies that rely on the software to move money on their behalf, the delay in candor may prove as damaging as the bug itself. KiteBridge has not said when automated payments will resume.

MoonshotAI: Kimi K3: # KiteBridge Bug Floods 91 Corporate Customers With $2.8 Million in Duplicate Reimbursements *A payment-processor timeout cascaded into more than 13,000 repeated payouts — and undercut a safeguard the company had advertised just months earlier.* An expense-management software update gone wrong left KiteBridge's corporate customers overpaid by $2.8 million this week, after a technical failure caused the company's systems to issue roughly 13,400 duplicate reimbursements in a little over an hour on Tuesday morning. KiteBridge, whose platform companies use to reimburse employees for business expenses, deployed version 4.2 at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m., 27 minutes later. Yet the payouts continued for nearly another hour, until 10:03 a.m., when KiteBridge suspended automated payments altogether — a window of roughly 78 minutes in which the platform generated more than 13,000 extra payments, on average nearly three every second. By then, 91 corporate customers had been affected. ## What went wrong According to the company, the failure began with a timeout at its payment processor, the outside service responsible for actually moving money. When a payment request times out, there is often no immediate way to know whether it succeeded, so systems are designed to try again; retries are a routine and necessary part of payment software. KiteBridge's fallback system — the backup logic meant to keep payments flowing when something goes wrong — treated each retry as an entirely new reimbursement request rather than a repeat of the same one. Every retry therefore triggered an additional payment, turning single reimbursements into many. Modern payment systems normally guard against this by assigning each transaction a unique identifier, so duplicates can be recognized and discarded before money moves. The incident is especially awkward for KiteBridge because it undermines a protection the company had explicitly advertised. An April changelog for version 4.0 promoted "exactly-once payment protection," an industry term for safeguards that guarantee a transaction is processed one time only, even when technical hiccups force systems to retry. Tuesday's failure was a breakdown of precisely that guarantee. ## Disclosure and recovery The company's communication during the incident has also drawn scrutiny. KiteBridge's public incident page initially described the problem only as delayed reimbursements. It did not disclose the duplicate payments until roughly six hours after automated payments were suspended — around 4 p.m. — leaving customers in the dark for much of Tuesday about the fact that money had gone out in error, not simply arrived late. By Wednesday evening, KiteBridge said it had recovered $2.3 million of the $2.8 million, with roughly $500,000 still pending. The company emphasized that no employees had money withdrawn from their personal accounts as a result of the error; the duplicate payments were funded by employers, whose corporate accounts are managed through the platform. CEO Lena Ortiz apologized for the incident and said KiteBridge will publish a technical report explaining what went wrong. For the 91 affected customers, that report will need to answer two uncomfortable questions: how a protection introduced with fanfare in April failed so completely within months, and why the company took six hours to tell customers what its systems had actually done.

Anthropic: Claude Fable 5.1: # Raincheck Raises $8 Million to Help Cities Decide Where to Send Crews Before a Storm *The Austin startup combines drainage maps with rainfall forecasts so local governments can prioritize repairs before the rain arrives.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The company maps municipal drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. The 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 challenge for cities is not identifying where flooding happens, but deciding what to fix first: > 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 has published [an explanation of how it builds its flood-risk forecasts](https://example.com/raincheck-methodology). The company stresses that those forecasts are recommendations; emergency managers retain control over deployments and public warnings. Raincheck currently works with seven cities in Texas and Louisiana. It declined to identify them because some contracts have not been publicly announced. The company 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.

MoonshotAI: Kimi K3: # Raincheck Raises $8 Million to Help Cities Fix Flood Risks Before the Storm *The Austin startup, whose software maps drainage systems against rainfall forecasts, has now raised $10.5 million and plans to expand into Florida.* Raincheck, a maker of flood-risk software for local governments, announced an $8 million seed round today. The Austin-based 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 round was led by Basin Ventures, with participation from existing investor Long Mile. It brings Raincheck's total funding to $10.5 million. The company says the new capital will help it hire eight engineers and expand into Florida. > “Our customers usually know which streets flood,” CEO Ana Flores said. “The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” Raincheck details its forecasting approach in a [published methodology](https://example.com/raincheck-methodology). The company currently works with seven cities in Texas and Louisiana, though it 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.