Head to head: Anthropic: Claude Fable 5.1 vs Qwen: Qwen3.8 2.4T A95B

Anthropic: Claude Fable 5.1: # Foundry Hall to Close After Losing Its Largest Sponsor The lights in Foundry Hall will go off for the last time on October 31. The nonprofit startup accelerator, which opened in Milwaukee in 2018, announced this week that it cannot afford to run another full program. Its largest corporate sponsor, Blank River Bank, declined to renew a sponsorship that covered 38 percent of the organization's annual budget. Foundry Hall has enough money to keep the doors open through the end of October, and no further. "We kept waiting for the replacement check that never came," said executive director Camille Ross. Six employees will lose their jobs. Twelve startups that currently work out of the building will have to find somewhere else to go. In seven years, Foundry Hall supported 212 companies. Together, those companies employ 1,140 people — more than the population of some Wisconsin towns, and a number that will keep growing after the building that produced it is empty. For many of those founders, Foundry Hall was the first place anyone took them seriously. Dev Malik, who went through the program and now runs his own company, remembers pitching from the back table of a coffee shop before he arrived. "My first investor meeting happened in that building," he said. Blank River Bank, in a statement, described the decision as a shift in its community investment priorities and said it had nothing to do with Foundry Hall's performance. The bank did not say what those priorities are now. The distinction offers little comfort to the people inside. An accelerator that depends on a single sponsor for more than a third of its budget is, by design or by necessity, fragile. When the sponsor walks away, there is no slow decline — only a date on the calendar. City officials have begun looking for a landing spot for the twelve companies still in residence. Deputy mayor Susan Cole said the city is in conversations with building owners and local universities about workspace, though no arrangement has been announced and the clock is already running. What the city cannot easily replace is harder to name. Foundry Hall was not the only place in Milwaukee where a founder could rent a desk. It was a place where a small group of people made it their job to believe in strangers' ideas, and then to introduce those strangers to people with money. That kind of work does not show up on a balance sheet until it stops. Through October, the building will operate as it always has. Companies will keep their desks, meetings will keep happening, and the staff will keep showing up to a job that has an end date. Then the furniture will be moved out, the twelve companies will scatter to wherever the city finds room for them, and 1,140 jobs will stand as the record of something that no longer exists.

Qwen: Qwen3.8 2.4T A95B: # The Last Lights in Foundry Hall Foundry Hall will close on October 31, ending a seven-year run in Milwaukee’s startup scene. The nonprofit accelerator, opened in 2018, became a first address for founders who had ideas but no lobby, no conference room, no one to ask where the printer paper lived. By the end, it had supported 212 startups whose companies employ 1,140 people, a tally that suggests how much small, unglamorous work can accumulate in a city when someone gives it a place to happen. The closure is not the result of a sudden scandal or a failed cohort. It is simpler and harder than that. Blank River Bank, Foundry Hall’s largest corporate sponsor, provided 38% of the organization’s annual budget. When the bank chose not to renew that support, the hall did not have enough runway to fund another full program. It can operate through October, but not beyond. Six employees will lose their jobs. Twelve startups currently using the workspace must leave. Executive Director Camille Ross described the months after the bank’s decision as a search for a bridge that never appeared, saying, “We kept waiting for the replacement check that never came.” Blank River Bank said its decision reflected a shift in community investment priorities and was not tied to Foundry Hall’s performance. The explanation is tidy, but for the people inside the building, it changes little. A budget line moved; a door closes. Founder Dev Malik remembers the hall as the place where his company first looked real. “My first investor meeting happened in that building,” he said. Before that, he had been pitching from the back table of a coffee shop. Those tables, rooms, and borrowed squares of whiteboard were part of the model. Accelerators often speak of mentorship, capital, and pipelines, but much of their labor is shelter: a desk, a sign-in code, a staff member who remembers your name. Foundry Hall gave early companies a setting in which they could be taken seriously before they had earned the usual credentials. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about space for the displaced companies. That effort may soften the scramble ahead, yet it cannot replace the organization itself. The startups may find desks elsewhere. The jobs lost are immediate. The six employees who kept programs running, introduced founders to resources, and probably knew which doors stuck in the afternoon will now have to look for their own next rooms. On Halloween, the hall will go dark. There will be boxes in the hallway, logos taken down from walls, and a quiet that comes after a place has been useful. Milwaukee still has founders, and some will carry Foundry Hall with them into offices, labs, kitchens, and leased floors. But for now, the story ends with a date, a shortfall, and 1,140 jobs that once had a beginning here.

Anthropic: Claude Fable 5.1: Let me check each line: 1. `#Flood software moves upstream` — missing space after `#` → violation. 2. `The company is expanding its pilot.` — dek not italicized → violation. 3. `##Details` — missing space after `##` → violation. 4. `Read [the announcement](www.example.com/news).` — not an absolute HTTPS URL → violation. 5. `* The first pilot begins Monday.` — uses asterisk instead of hyphen → violation. 6. `<div>The second pilot begins Friday.</div>` — raw HTML → violation. **6**

Qwen: Qwen3.8 2.4T A95B: 6

Anthropic: Claude Fable 5.1: # QueueZero Launches Proxy, an AI Meeting Assistant That Turns Conversation Into Tasks QueueZero introduced an AI meeting assistant called Proxy on Wednesday, pitching it as a tool that can sit in on a meeting, write up the notes and hand off follow-up work to software agents without a human relaying the instructions. The launch demonstration showed how literally the product can take that job. Midway through the presentation, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was listening to the session, treated the remark as an instruction. Within moments it had created two tasks and assigned both to QueueZero chief executive Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. The exchange was brief, but it illustrated the central design question for products like Proxy: how an assistant should decide which parts of a conversation are commitments and which are commentary. QueueZero's answer, at least for now, is a distinction between internal and external actions. According to the company, Proxy will not send an external message — an email to a client, for example — without a human confirming it first. Internal tasks, such as items assigned to colleagues inside the organization, can be created automatically from what the assistant hears. That is the path the audience remark took. Administrators who prefer a more cautious setup can disable automatic task creation entirely, leaving Proxy to propose tasks for review rather than file them on its own. The company did not say whether it plans to change the default behavior after Wednesday's demonstration. Proxy is one of a growing number of products that aim to move AI assistants beyond transcription and summarization and into delegation. Earlier generations of meeting software produced a record of what was said; the newer category tries to act on it, assigning work to people or to other automated agents that can draft documents, update records or schedule follow-ups. That ambition raises the stakes for errors. A mistaken summary can be corrected after the fact. A mistaken assignment lands in a real person's queue before anyone has reviewed it, and in a large organization the person receiving it may have no idea where it came from. Wednesday's incident was easy to catch because the task's recipient was standing on stage. Most will not be. Proxy is priced at $30 per user per month. QueueZero said the product will enter public beta next week.

Qwen: Qwen3.8 2.4T A95B: # QueueZero Launches Proxy AI Meeting Assistant After Live Demonstration Creates Tasks for CEO QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant designed to attend meetings, take notes and assign follow-up work to software agents. The product enters public beta next week at $30 per user per month. The company demonstrated Proxy during a launch event, but the demonstration took an unexpected turn when a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the comment as an instruction. According to QueueZero, the assistant 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. The episode proved that an AI assistant can be extremely attentive to the wrong sentence. It also highlighted a broader challenge facing enterprise AI vendors: ensuring that systems can distinguish between conversational remarks, requests and commands. QueueZero said Proxy is designed with safeguards around outbound communication. The assistant requires confirmation before an external message is sent, preventing it from automatically emailing customers or contacting outside parties based on overheard comments. Internal tasks, however, can be created automatically. That feature allows Proxy to convert meeting discussions into to-do items, assign follow-ups to software agents and place work into existing project management systems without waiting for manual input. QueueZero said administrators can disable automatic task creation for organizations that prefer stricter controls. The setting may appeal to enterprises deploying AI assistants in sensitive meetings where casual comments could be misread as directives. Meeting assistants have become one of the most visible categories in enterprise AI. Vendors are racing to reduce time spent on note-taking, summarizing calls and tracking action items. As companies integrate AI agents more deeply into workplace software, the challenge is shifting from transcription accuracy to operational judgment. Proxy’s launch demonstration underscored that gap. The system captured the audience remark, parsed it into actionable items and routed them to the appropriate person, even though the comment was not addressed directly to the software. QueueZero is positioning Proxy as an agent that can move beyond summarization. In addition to producing meeting notes, it can trigger follow-up work, potentially allowing teams to close the distance between discussion and execution. The public beta will give customers an early look at how those controls work in practice.

Anthropic: Claude Fable 5.1: # Raincheck Raises $8 Million to Help Cities Decide Where to Send Crews Before a Storm *The Austin startup's flood-risk software pairs drainage maps with rainfall forecasts. The seed round, led by Basin Ventures, brings its total funding to $10.5 million.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million. The 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. Raincheck has published an outline of [how it builds those forecasts](https://example.com/raincheck-methodology). The seed 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 software is meant to address the step that comes after a city already knows its trouble spots: > 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 only; emergency managers retain control over deployments and public warnings. Raincheck competes with other flood-planning vendors, including HydroMap, as well as municipal engineering contractors. The company says the new funding will help it hire eight engineers and expand into Florida.

Qwen: Qwen3.8 2.4T A95B: # Raincheck Raises $8 Million to Help Cities Prepare for Flooding *The Austin company combines drainage maps with rainfall forecasts to help local governments prioritize pre-storm work.* Raincheck, an Austin company that develops flood-risk software for local governments, said today it has raised $8 million in a seed round led by Basin Ventures. Existing investor Long Mile also participated, bringing the company’s total funding to $10.5 million. The company maps drainage systems and combines those maps with rainfall forecasts, giving cities a way to decide where to send maintenance crews before a storm arrives. Chief Executive Ana Flores said customers generally know where flooding occurs, but struggle to determine which fixes should come 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, whose approach is described in [Raincheck’s methodology](https://example.com/raincheck-methodology), 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 intended as recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck said the new funding will help it hire eight engineers and expand into Florida.

Anthropic: Claude Fable 5.1: # HelioDock Launches Remote Robot Monitoring Software, Raises $24 Million *Night Shift lets warehouse operators oversee robot fleets from off site, but its usefulness drops when the network does.* HelioDock, a maker of autonomous warehouse robots, on Thursday released Night Shift, software that lets operators monitor and manage robot fleets remotely, and disclosed a $24 million Series B round led by North Coast Ventures. The product is aimed at a problem familiar to anyone running automated warehouses around the clock: a single stalled robot can force a human to show up in person, often at inconvenient hours, to sort out what went wrong. Night Shift gives operators a way to see fleet status and intervene without being on the floor. HelioDock says customers in its private beta recorded 17 percent fewer unplanned robot stoppages, a figure the company supplied and that has not been independently verified. The software will cost $90 per robot per month and becomes 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," said Mara Velez, HelioDock's chief executive. Night Shift spent 14 months in a private beta before Thursday's launch. Nine logistics companies participated, running the software across 310 robots in total. HelioDock has not said how the 17 percent reduction in stoppages was measured or whether results varied across those nine deployments. Among the beta participants was Fulcrum Logistics, where Aaron Pike said the system performed well in facilities with strong connectivity. The picture changed in locations that rely on cellular networks. During rural LTE outages, Pike said, Night Shift became less useful. That trade-off is built into how the software works. Night Shift depends on a live connection between the warehouse and the remote operator. When that link fails, robots fall back to their locally configured behavior, and remote control becomes unavailable until connectivity is restored. The robots do not stop operating, but the off-site oversight the product is designed to provide does. For warehouses in areas with unreliable coverage, that means the two-in-the-morning drive Velez described may still happen on bad nights. HelioDock is not alone in selling fleet-management tools to warehouse operators. Relay Robotics and YardOS both offer competing products, and HelioDock will need to make the case that its software, priced on a per-robot basis, is worth the recurring cost on top of the hardware itself. At $90 per robot per month, a 100-robot facility would pay $108,000 a year for the service. The Series B brings HelioDock's total funding to $39 million. North Coast Ventures led the round; the company did not disclose other participants or how the proceeds will be allocated. The timing pairs the capital raise with the first commercial availability of a software product from a company previously known for its hardware, suggesting HelioDock sees recurring software revenue as part of its growth path alongside robot sales. Whether Night Shift's beta results hold up across a broader customer base remains to be seen. The 310-robot pilot is a meaningful test but a limited one, and the connectivity constraints Pike described point to a product whose value will vary by site. Operators with wired, redundant networks stand to get the most from it. Those running robots in remote distribution centers served by a single cellular carrier will likely find the software most helpful precisely when they need it least. General availability on September 15 opens the product to customers outside the original nine beta companies. HelioDock has not said whether existing beta participants will receive different pricing or whether the $90 monthly rate includes support.

Qwen: Qwen3.8 2.4T A95B: # HelioDock Launches Night Shift Remote Monitoring Software, Raises $24 Million *Night Shift follows a 14-month private beta with nine logistics companies and 310 robots, with general availability set for September 15.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely, and announced a $24 million Series B led by North Coast Ventures. The release pairs HelioDock’s robots with a new software product for operating them from a distance. Night Shift is built around remote monitoring, giving operators a way to watch fleets without standing beside the robots. The company is entering a field where fleet-management products are already available from Relay Robotics and YardOS. Night Shift is software, not a robot. HelioDock makes autonomous warehouse robots, and Night Shift provides the monitoring tools for those robots. The product will move to general availability on September 15, when the $90 per-robot monthly price will apply. Night Shift spent 14 months in private beta with nine logistics companies, which ran the software on 310 robots. HelioDock said beta customers saw 17% fewer unplanned robot stoppages. The company said that figure has not been independently verified. Unplanned stoppages are at the center of the operational problem HelioDock describes. When a warehouse robot stops, becomes uncertain, or cannot proceed, a person may need to intervene. Night Shift is intended to allow that response to begin remotely, reducing the need for an operator to be physically present when a robot requires attention. Chief Executive Mara Velez described the situation as a late-night dispatch problem. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” Velez said. Velez’s remark identifies a specific scenario: a manager driving across town late at night because one robot is confused by a pallet. Night Shift’s remote-monitoring function is the tool HelioDock offers for handling that kind of event without an in-person trip. The beta included both the claimed reduction in stoppages and a practical limit tied to connectivity. Aaron Pike of Fulcrum Logistics said Night Shift worked well in facilities with strong connectivity. He said the system became less useful during rural LTE outages. When connectivity fails, robots return to their locally configured behavior. Remote control becomes unavailable. Night Shift’s remote-monitoring function depends on a working connection. When the connection drops, the robot continues under local configuration, and the operator loses remote control until connectivity is restored. Network reliability sits among the factors that will shape how useful Night Shift is in a given warehouse. In a facility with strong connectivity, the software can support remote monitoring as HelioDock describes. In a rural facility experiencing LTE outages, the remote-monitoring role becomes less useful, according to the beta customer account. HelioDock said Night Shift will cost $90 per robot per month. The per-robot charge means the monthly cost rises with the size of the fleet covered by the software. General availability begins September 15, following the private beta period. Relay Robotics and YardOS offer competing fleet-management products. HelioDock’s entry into that market is tied to its autonomous warehouse robots and the remote-monitoring function announced Thursday. The beta provides the company’s current performance claim, the 17% reduction figure that has not been independently verified, and one beta customer’s account of how connectivity affected the system’s usefulness. The Series B brings HelioDock’s total funding to $39 million. North Coast Ventures led the round. HelioDock announced the financing on the same day it introduced Night Shift.

Anthropic: Claude Fable 5.1: # Startup Spotlight: LedgerLark ## Teaching software to chase down the trades that don't match Every day, investment firms process trades that mostly reconcile without incident. The ones that don't — a confirmation that disagrees with a custodian record, a ledger entry that doesn't line up — land on an operations desk as "exceptions." Resolving them is tedious, repetitive and, at small firms, often handled by a handful of people who also have other jobs. LedgerLark, a young company that recently raised a $3.5 million seed round led by Motive Row, is betting that this work is a good fit for AI agents. It is an early bet, and the company is candid that the hard parts of selling into financial institutions are still ahead. ## The problem, as the founders saw it Co-founder Mina Batra spent seven years in operations at a regional brokerage. She says the idea took shape after she watched a six-person team spend three days resolving a single trade mismatch. The cause turned out to be mundane: an account number had been copied incorrectly. Nothing about the error was exotic, but finding it meant comparing records across several systems by hand. Her co-founder, Owen Carr, approached the problem from the other end. He built reconciliation systems at a large bank, where the tooling was more sophisticated but also expensive and tightly coupled to internal infrastructure. The two concluded that smaller firms were working with a gap: too much volume to handle manually, not enough budget for enterprise-grade systems. ## What the product does — and deliberately doesn't LedgerLark's agents ingest trade confirmations, custodian records and a firm's internal ledgers, then look for discrepancies. When they find one, they trace likely causes and propose a correction. A human reviews and approves the change before anything is updated. That last step is central to the pitch. The product does not execute trades and does not move money. Batra frames this as a design choice rather than a limitation: operations teams need to trust the system before they will let it act, and compliance departments are unlikely to approve software that alters records autonomously. Whether that human-in-the-loop model remains a feature or becomes a bottleneck as volume grows is an open question. ## Early traction Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The company says one customer cut its unresolved daily exceptions from roughly 70 to 18. That is a notable figure, with caveats. It is a single customer, reported by the company, and it is unclear how much of the improvement came from the software versus process changes made alongside it. Five customers is enough to suggest the product works in real environments; it is not enough to draw conclusions about how broadly the results generalize. ## The competitive and structural headwinds LedgerLark is not entering an empty field. Established reconciliation vendors have served this market for years, and many of them are adding automation features of their own. Larger institutions often rely on internal software that would be difficult to displace regardless of how capable an outside tool is. The sales process is another constraint. Financial institutions commonly require lengthy security and procurement reviews before adopting new software, particularly software that touches trade data. For a company with a small team and a limited runway, long sales cycles can be as consequential as any technical challenge. The seed round buys time, but not an unlimited amount of it. ## What to watch The interesting question is not whether AI can spot mismatched records — that seems well within reach — but whether small firms will trust an outside agent with the task, and whether LedgerLark can get through the door before incumbents close the gap. The founders' operational background gives them a credible read on the problem. The next year will show whether that translates into a durable business.

Qwen: Qwen3.8 2.4T A95B: # Startup Spotlight: LedgerLark Targets the Messy Middle of Post-Trade Operations When a trade breaks, the problem rarely announces itself as a grand failure. More often, it appears as a copied account number, a confirmation that does not match a custodian record, or an exception report that grows while staff move between email, spreadsheets and legacy systems. LedgerLark, a startup for post-trade operations at small investment firms, is betting that a narrow slice of that work can be handed to software agents. The company’s AI agents compare confirmations, custodian records and internal ledgers, then propose corrections for human approval. It does not execute trades or move money. The idea grew out of co-founder Mina Batra’s seven years in operations at a regional brokerage. According to the company, Batra began exploring the problem after watching a six-person team spend three days resolving a trade mismatch caused by a copied account number. That episode points to a familiar operations pattern: the error itself may be small, but finding it, verifying it and fixing it can pull staff across systems and documents. Batra is joined by co-founder Owen Carr, who previously built reconciliation systems at a large bank. That combination — front-line operations experience and infrastructure work — suggests a practical understanding of where reconciliation fails. It is not always a lack of data. It is often too much data, scattered across formats and owned by different parties. LedgerLark’s product stays deliberately inside a controlled lane. The agents identify discrepancies and draft corrections, but people remain in the loop. For financial firms, that distinction matters. Automation that merely recommends actions can be easier to review, test and explain than software that makes changes on its own. It also leaves open the question of how much trust clients will place in the agents over time, especially when exceptions are unusual or high value. Early traction is modest but specific. The company says five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. It also says one customer reduced its unresolved daily exceptions from 70 to 18. Because that figure comes from the company and relates to one customer, it is not yet clear whether the result reflects a typical use case, a particularly messy starting point or a limited set of exception types. The funding news is equally straightforward: LedgerLark has raised a $3.5 million seed round led by Motive Row. For a startup selling into financial operations, that capital will likely need to support implementation, security reviews and client-specific configuration rather than rapid, frictionless expansion. That is because the market is not empty. Established reconciliation vendors already serve financial institutions, and many banks rely on internal software built over years. Those incumbents may be slower or more cumbersome, but they also carry existing relationships, compliance history and integration depth. LedgerLark will need to show not only that its agents reduce backlog, but that they fit cleanly into firms’ existing controls, audit trails and reporting requirements. There is also the pace of selling itself. Financial institutions often require lengthy security and procurement reviews, and small investment firms may have limited budgets for new software. Even if the pain is real, converting it into recurring revenue can be slow. Still, the startup’s focus is sensible. Post-trade operations is unglamorous, but errors are costly and staff time is scarce. If LedgerLark can reliably reduce routine exceptions while keeping humans responsible for approvals, it may find a durable niche. The harder task will be proving that the system works across enough edge cases, clients and workflows to justify displacing tools firms already know.

Anthropic: Claude Fable 5.1: # The Benchmark That Says "Pay Less": Two Support Models Tie on Quality, and One Costs Less Than Half as Much When researchers put five language models through 200 real customer-support tickets, the headline result wasn't which model scored highest. It was that the top two scored essentially the same — and one of them costs 61% less to run. ## The finding that matters Model North posted the best average score, 84.2 out of 100. Model Vale came in at 83.8. The researchers checked whether that 0.4-point gap means anything, and it doesn't: the difference was not statistically significant. Given 200 tickets and three human graders, that spread is well within the noise you'd expect from grader judgment and ticket-to-ticket variation. Where the two models do differ is everything around the score. Vale costs an estimated $0.74 per 1,000 tickets; North costs $1.90. Vale also returned answers 18% faster. For a support operation handling tens of thousands of tickets a month, that's a meaningful difference in both budget and queue time, purchased at no measurable cost in answer quality. If you take one thing from this study, take that: the most expensive option in the test did not buy a detectably better answer. ## How the test worked The setup was straightforward but worth understanding before drawing conclusions. - Researchers collected 200 anonymized tickets from a single software company, all in English. - Each of the five models produced one proposed response per ticket. - Three experienced support managers graded every response blind — they didn't know which model wrote what. - Grades combined three dimensions: factual correctness, usefulness to the customer, and tone. That design is a reasonable proxy for "would a support manager be happy sending this?" It's a sharper question than generic language benchmarks ask. But it also means the scores reflect a particular company's tickets and a particular team's taste, a point we'll return to. ## The rest of the field Below the top pair, scores dropped off more clearly: | Model | Avg. score | Cost per 1,000 tickets | |---|---|---| | North | 84.2 | $1.90 | | Vale | 83.8 | $0.74 | | Ember | 79.1 | $0.31 | | Kite | 76.4 | $0.22 | | Harbor | 71.0 | $0.08 | The cheaper three trade real quality for savings. Ember, Kite and Harbor all sit several points behind, and Harbor's 13-point gap from the leaders is large enough that it likely shows up as a noticeably different customer experience. Whether that tradeoff is worth it depends on the stakes of your tickets; a 71 might be perfectly adequate for password resets and a liability for refund disputes. The study didn't test whether the gaps among these three are significant, so be cautious about ranking them against each other too confidently. ## Strengths weren't uniform The aggregate scores hide one useful texture. North did best on billing disputes, while Vale did best on technical troubleshooting. The researchers didn't report how large those category gaps were, so it's hard to say whether routing tickets by type would pay off. But it suggests that "which model is best" may have a different answer depending on what your inbox actually contains. A company whose tickets skew heavily toward one category might reasonably prefer one model over the other despite the overall tie. ## One weakness they all shared Every model struggled when a ticket depended on an image attachment — a screenshot of an error, a photo of a damaged product, a marked-up invoice. This was the clearest across-the-board failure in the study, and it cuts against the cost story: it doesn't matter how cheap a model is if a chunk of your tickets require something none of them can handle well. Teams piloting any of these systems should expect to route image-dependent tickets to humans or add a separate vision step. ## What the study can't tell you The limitations are specific and should temper how far you generalize: - **One company, one language.** All tickets were English and came from a single software firm. Results may not transfer to hardware support, financial services, or multilingual queues. - **Static answers only.** Models produced text; they didn't look up order histories, issue refunds, or call tools. Real deployments increasingly depend on those agentic steps, and this test says nothing about them. - **Three graders.** Experienced, blind, but still a small panel with shared context. Another team might weight tone or correctness differently. - **Estimated costs.** Pricing was researcher-estimated and will shift with volume discounts and vendor changes. ## The bottom line Within the bounds of this test, the sensible reading is this: if you're choosing between the two top models, the quality case for the expensive one isn't there, and the cost and speed case for the cheaper one is. Below that tier, you're making a genuine quality-for-price trade. And whatever you pick, plan for the tickets with screenshots.

Qwen: Qwen3.8 2.4T A95B: # For Support AI, the Top Models Are Close, but Cost Stands Out The clearest finding from this benchmark is not that one model dominated. It is that North and Vale, the two highest-scoring systems, were statistically indistinguishable in overall quality while differing sharply in price and speed. North averaged 84.2 and Vale averaged 83.8 across 200 anonymized customer-support tickets, but the 0.4-point gap was not statistically significant. In practical terms, the study did not show a reliable quality advantage for North. It did show that Vale costs far less and responds faster. That makes the benchmark less a simple leaderboard and more a purchasing signal. For teams whose tickets resemble the test set, Vale looks like the stronger general value: nearly identical quality to North, a much lower estimated cost, and 18% faster answers. North remains attractive where its specialty, billing disputes, is central. The lesson is that small average-score differences can hide meaningful differences in workflow fit, cost, and latency. ## Methodology and limitations The test used 200 anonymized tickets from one software company. Each of the five models produced one proposed answer per ticket. Three experienced support managers graded the answers blindly, so they did not know which model wrote each response. Scores combined factual correctness, usefulness, and tone, which reflects how real support teams evaluate customer-facing writing. Blind grading by experienced managers is a strength, because it reduces brand bias and values practical quality over abstract metrics. Still, the results apply to a narrow setting. The tickets were English-language tickets from one company, so the findings may not transfer to other products, customer bases, or languages. The researchers also did not test agentic actions, tool use, or multilingual support. And every model performed poorly when a ticket depended on an image attachment, a serious limitation for queues where screenshots and error images matter. ## The top tier and its trade-offs North and Vale formed a clear top tier. Their scores were close enough that choosing between them on overall quality alone would overread the data. The bigger distinction is operational. Vale cost an estimated $0.74 per 1,000 tickets, compared with $1.90 for North, and it answered 18% faster. At scale, that price gap can change whether AI drafting is economical across large queues. Faster answers can also reduce backlog pressure and make agents more willing to use suggestions. The models also had different strengths. North performed best on billing disputes, where careful wording and policy sensitivity can matter more than raw speed. Vale performed best on technical troubleshooting, where clear diagnosis and next steps are often decisive. A team should therefore weigh its ticket mix before choosing. A billing-heavy support organization may favor North despite the higher cost, while a technically oriented queue may favor Vale. ## What about the lower-scoring models? Ember scored 79.1, Kite 76.4, and Harbor 71.0. Those differences are large enough to affect customer experience, so these models are less compelling for broad customer-facing use. However, lower cost can still make them useful in limited roles, such as internal drafts, triage suggestions, or low-risk categories with mandatory human review. Harbor’s $0.08 per 1,000 tickets is especially cheap, but the quality gap means buyers should not treat low price as an automatic substitute for capability. The image-attachment weakness deserves special attention. All five models performed poorly when understanding the ticket required an image. That means the benchmark’s headline rankings are most relevant to text-only support. For teams with many screenshot-based tickets, the first problem is not choosing between North and Vale; it is building a workflow that can process images or route those tickets to humans. ## Bottom line The benchmark does not crown a single winner. It points to a practical one: Vale offers top-tier quality at a much lower cost and higher speed, making it the most sensible general choice under these test conditions. North is the model to watch when billing disputes dominate. Before deploying any model widely, support leaders should test it on their own tickets, including attachments, edge cases, and the languages their customers actually use.

Anthropic: Claude Fable 5.1: # Software Glitch Sends $2.8 Million in Duplicate Reimbursements to Employees at 91 Companies A faulty software update at KiteBridge, a maker of expense-management software, caused the company's systems to issue 13,400 duplicate reimbursement payments totaling $2.8 million on Tuesday, leaving 91 corporate customers to recover money that should never have left their accounts. By Wednesday evening, KiteBridge said it had clawed back $2.3 million. The remaining roughly $500,000 was still pending. ## A routine update, then a cascade The trouble began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its platform. Less than half an hour later, at 9:12 a.m., the first customer complaint arrived. The underlying cause, according to the company, was a timeout at a payment processor — the third-party service that actually moves money from an employer's account to an employee's. When a payment request takes too long to receive a response, software like KiteBridge's typically retries the request, on the assumption that the first attempt may have failed. That is where the new version broke down. KiteBridge's fallback system, which is supposed to handle these retries safely, instead interpreted each retry as a brand-new reimbursement request. Payments that had, in many cases, already gone through were sent again. Because the system was operating automatically and at scale, the duplicates multiplied quickly: 13,400 extra payments were issued before anyone intervened. KiteBridge suspended automated payments at 10:03 a.m., roughly an hour and twenty minutes after the deployment and about fifty minutes after the first complaint. ## Who was affected The duplicates flowed to employees at 91 companies that use KiteBridge to reimburse staff for business expenses such as travel and meals. The money came from employers' accounts, not employees' — no employee had funds withdrawn. Employees who received a double payment were, in effect, briefly overpaid by their employers, and KiteBridge has been working with those companies to reverse the extra transfers. ## A protection that was supposed to prevent this The failure is notable because KiteBridge had advertised a safeguard against precisely this scenario. A changelog published in April stated that version 4.0 of the software had introduced "exactly-once payment protection" — a guarantee, in industry terms, that a single reimbursement request will result in exactly one payment, no matter how many times the system retries it. Exactly-once guarantees are a well-understood challenge in payment software. Networks fail, processors time out, and systems must decide whether a request that received no response actually went through. The standard solution is to tag each payment with a unique identifier so that retries are recognized as repeats rather than new requests. KiteBridge has not yet explained how version 4.2 bypassed or undermined the protection it introduced two versions earlier. ## Slow to disclose KiteBridge also drew criticism for how it communicated during the incident. Its public incident page initially described the problem as *delayed* reimbursements — language suggesting that employees were waiting for money, not receiving it twice. The page did not disclose the duplicate payments until six hours after the company had already halted automation, meaning affected employers spent much of Tuesday unaware of the true nature of the problem. ## Apology and a promised report CEO Lena Ortiz apologized for the incident and said the company would publish a technical report explaining what went wrong. She did not give a timeline for the report or for recovering the outstanding funds. For KiteBridge's customers, the episode is a reminder that automation in payment systems cuts both ways: the same speed that makes reimbursements convenient allowed a single bug to issue thousands of erroneous payments in under an hour and a half.

Qwen: Qwen3.8 2.4T A95B: ## A Routine Update Became a $2.8 Million Payment Error A software update at KiteBridge on Tuesday morning produced an immediate and expensive failure: 13,400 duplicate reimbursements worth $2.8 million were sent across 91 corporate customers before automated payments were shut down. The trouble began with version 4.2, deployed at 8:45 a.m. Tuesday. At 9:12 a.m., the first customer complaint arrived. By late morning, engineers had confirmed that the problem was not a simple delay. A payment-processor timeout had activated KiteBridge’s fallback system. In normal operation, that backup process is meant to keep reimbursements moving when a payment request fails or stalls. In this case, however, the fallback treated retry attempts as fresh reimbursement requests, causing the same employee expenses to be paid more than once. The error struck at a basic expectation of business software: a payment should not be issued twice for the same claim. KiteBridge had publicly promoted that safeguard. In an April changelog, the company said version 4.0 had introduced “exactly-once payment protection,” a technical control designed to prevent duplicate transactions even when systems need to retry a payment. Tuesday’s failure raised questions about how that protection worked with the new version and under the timeout conditions. KiteBridge suspended automated payments at 10:03 a.m., but its public explanation lagged behind the technical reality. The company’s incident page initially described the issue as delayed reimbursements. It did not disclose the duplicate payments until six hours after the automation had already been suspended. That gap left affected employers trying to understand why funds were moving incorrectly before they had a full description of the incident. By Wednesday evening, KiteBridge said it had recovered $2.3 million of the duplicated funds. The remaining balance was still pending, leaving some customers waiting for reconciliation. The company said no employees had money withdrawn from their accounts; instead, the duplicate payments were funded by employers, meaning the financial exposure fell on companies using KiteBridge’s platform rather than individual workers. The episode underscores how quickly automation can turn an administrative process into a financial cleanup. Expense-management systems often connect corporate accounts, employee claims and outside payment networks, so a single flaw in retry logic can spread across many employers before internal controls catch it. Chief Executive Lena Ortiz apologized for the incident and said the company would publish a technical report. The review is expected to examine how version 4.2 handled payment retries, why the fallback system failed to recognize repeated attempts, and whether the exactly-once safeguard was adequately tested before release. For KiteBridge, the immediate task is recovering the outstanding funds, restoring trust with the 91 affected corporate customers, and explaining the controls meant to prevent a repeat. The larger lesson may be more difficult: in automated payment systems, a fallback is only as safe as its ability to recognize when it has already acted.

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.

Qwen: Qwen3.8 2.4T A95B: # 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

Qwen: Qwen3.8 2.4T A95B: C

Anthropic: Claude Fable 5.1: # Cloudnote Left Customer Prompt Excerpts on Unsecured Debug Page for Four Hours *The note-taking service says snippets from 1,842 user sessions were exposed and that only the researcher who reported the flaw appears to have seen them, though its investigation is not complete.* Cloudnote, an AI-assisted note-taking service with 3.2 million registered users, said an internal debugging page containing excerpts of customer prompts was reachable without a password for about four hours on Monday. Server logs show the page displayed snippets from 1,842 user sessions, the company said. Each excerpt could include up to 200 characters of text a user had typed into the service's assistant—long enough to capture a name, a meeting agenda or a line of private correspondence. Full documents, account passwords and payment information are stored on separate systems and were not present on the page, according to Cloudnote. A security researcher noticed the exposure and alerted the company, which removed the page 19 minutes later. Cloudnote said it has found no evidence that anyone besides the researcher accessed the page. But it acknowledged its review is ongoing, and it has not said how the page became publicly accessible, how it is verifying who visited it, or whether affected users will be notified individually. The incident affects a small fraction of Cloudnote's users, but it underscores a growing risk for AI services: the prompts people type are often more revealing than the files they upload, and tools built for engineers to inspect them can become leaks if left unguarded.

Qwen: Qwen3.8 2.4T A95B: # Cloudnote debugging page exposed customer prompt excerpts for four hours *The incident affected a small fraction of the company’s 3.2 million registered users, and Cloudnote says there is no evidence anyone besides a researcher viewed the data.* Cloudnote said Monday that a debugging page containing excerpts from customer prompts was accessible without authentication for four hours, exposing a limited slice of user interactions with the service. Company logs indicate the page displayed excerpts from 1,842 user sessions, with each excerpt containing as many as 200 characters from a prompt. Cloudnote, which has 3.2 million registered users, said the page did not include full documents, passwords or payment information, which are stored elsewhere. The exposure ended 19 minutes after a researcher notified Cloudnote, though the page had already been reachable for four hours. The company said it has found no evidence that anyone other than the researcher accessed it, but cautioned that its investigation is not complete. It has not determined whether the exposed excerpts contained personal or sensitive information. The incident matters because even partial prompt data can reveal sensitive details, depending on what users asked the service to process. It also raises questions about why a page containing customer data was unauthenticated and how such exposures are detected before outside researchers find them. Cloudnote has not said whether affected users will be notified.

Anthropic: Claude Fable 5.1: # Arcwell Launches Relay, an AI Assistant for Customer Support Teams Arcwell has unveiled Relay, a platform that reads customer support conversations and feeds agents suggested replies, in a bid to cut the time support staff spend hunting for answers their companies already possess. The product does three things. It analyzes support exchanges and proposes responses for the agent handling them, it flags customers who appear likely to cancel, and it generates summaries of conversations. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Before launch, Arcwell ran Relay through a six-month trial with 12 companies. According to the company, agents using the tool closed tickets 14% faster. That figure covers speed alone. Arcwell has not released data on whether Relay's churn-detection feature helped the pilot companies keep customers — a notable gap, given that retention is one of the platform's three headline capabilities. Relay is priced at $65 per agent per month. How useful the platform proves will depend on how much it can see. To work, Relay needs access to a company's support transcripts and internal help documentation — precisely the material many businesses guard most closely. Arcwell says that data is encrypted and is not used to train models shared across customers. Buyers can also choose how long Relay holds their information, with retention windows of either 30 or 90 days. Those assurances answer the most obvious objection, but they also place the burden of trust squarely on Arcwell. Companies weighing Relay will be deciding not just whether a 14% speed gain justifies the monthly fee, but whether they are comfortable routing their support history through a third party. For Arcwell, Relay is a step into a crowded corner of the software market, where artificial intelligence is increasingly being sold as a way to reshape customer service. The company's bet is that agents will see the tool as something that surfaces what they need faster rather than something that replaces them. Whether customers notice the difference — and whether fewer of them walk away — is a question Arcwell has yet to answer with numbers.

Qwen: Qwen3.8 2.4T A95B: # Arcwell launches Relay to point support agents toward answers Arcwell has released Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to agents. The launch places Arcwell more squarely in the broader customer service industry’s AI rollout, with a product aimed at speeding up service, spotting signs of churn and reducing administrative work for support teams. Relay reviews support conversations and the internal help documents companies already maintain. From that material, it can recommend answers, identify customers who may cancel, and create summaries of interactions. The combination of answer suggestions, cancellation warnings and summaries reflects how AI is being applied across customer service by reading conversations, surfacing relevant information and condensing manual work. That approach points to a broader shift in support software: rather than serving only as a repository for tickets and policies, the system tries to interpret conversations and surface the next step for agents. The platform was tested by 12 companies over six months. Arcwell said agents using Relay closed tickets 14% faster, though the company did not release customer-retention results. Relay costs $65 per agent each month. CEO Priya Shah said the product is aimed at a familiar problem in support organizations. “Support agents spend too much time searching for information that their companies already have,” Shah said. Using Relay requires giving the system access to support transcripts and internal help documents, a requirement that could raise questions about data handling. Arcwell said customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period, giving customers some control over how long their support records remain in the system. For Arcwell, the central argument is that AI can help human agents work with information their company already owns. Relay offers suggested answers, churn signals and summaries. The clearest disclosed evidence is faster ticket closure, while the retention impact remains undisclosed.