Head to head: Anthropic: Claude Opus 5.5 vs Qwen: Qwen3.8 2.4T A95B

Anthropic: Claude Opus 5.5: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 Sessions *The page was publicly reachable for four hours on Monday. The company says it has found no sign that anyone other than the researcher who reported it gained access, but its investigation is not finished.* A Cloudnote debugging page containing excerpts of customer prompts was accessible on the open internet without authentication for four hours on Monday, exposing fragments of text from 1,842 user sessions, according to the company's logs. Each excerpt could include up to 200 characters of a prompt, which is long enough to contain names, project details or other information users may not have expected anyone else to see. Cloudnote said full documents, passwords and payment information are stored on separate systems and did not appear on the page. The company removed the page 19 minutes after a security researcher alerted it to the problem. Cloudnote said it has found no evidence that anyone besides that researcher viewed the page. That finding is preliminary. Cloudnote has not completed its investigation, and several questions remain open: how the page became publicly accessible, whether the company's records can rule out other visitors, and how many individual people the 1,842 sessions represent, since one user can account for more than one session. The affected sessions are a small fraction of activity on a service with 3.2 million registered users. But the incident shows how a tool built for internal troubleshooting can surface what customers type, and it raises questions about what Cloudnote logs and who can reach it.

Qwen: Qwen3.8 2.4T A95B: # Cloudnote Debugging Page Exposed Customer Prompt Excerpts for Four Hours *The company says 1,842 user sessions may have been visible before the page was taken down, and the investigation is ongoing.* Cloudnote, which has 3.2 million registered users, said Monday that a debugging page containing excerpts from customer prompts was publicly accessible without authentication for four hours. The exposure was limited: logs indicate excerpts from 1,842 user sessions, each potentially including up to 200 characters of a prompt. Cloudnote did not say how the debugging page came to be exposed or whether internal alerts detected it before the outside report. The company said it removed the page 19 minutes after a security researcher notified it. Cloudnote said it has found no evidence to date that anyone other than the researcher accessed the page, but added that its investigation is not complete, leaving open the possibility that additional access occurred before detection. The incident matters because even brief fragments of prompts can reveal sensitive context about how users employ the service, including work drafts, questions or personal details. Cloudnote said full documents, passwords and payment information were stored elsewhere and did not appear on the exposed page. The company said it has not yet determined whether the excerpts were copied or retained.

Anthropic: Claude Opus 5.5: # Raincheck Raises $8 Million for Flood-Risk Software *The Austin startup combines drainage maps with rainfall forecasts to help local governments decide where to send crews before a storm.* 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's software maps drainage systems and combines those maps with rainfall forecasts. Cities use it to decide where to send maintenance crews before a storm. Basin Ventures led the round, with participation from existing investor Long Mile. The funding brings Raincheck's total raised to $10.5 million. CEO Ana Flores described the problem the company is trying to solve: > "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." Details of [Raincheck's methodology](https://example.com/raincheck-methodology) are available online. Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. The software's forecasts are recommendations only. Emergency managers retain control over deployments and public warnings. Raincheck's competitors include HydroMap and municipal engineering contractors. The company says it will use the new funding to hire eight engineers and expand into Florida.

Qwen: Qwen3.8 2.4T A95B: # Raincheck Raises $8 Million to Help Local Governments Prepare for Floods *The Austin company maps drainage systems and pairs them with rainfall forecasts to help cities prioritize maintenance crews before storms.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in a seed round. 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. The seed round was led by Basin Ventures. Existing investor Long Mile also participated. Raincheck has raised $10.5 million total. CEO Ana Flores said the harder task for cities is not knowing where flooding occurs, but deciding how to act before a storm arrives. > “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.” More details about the company’s approach are available in [Raincheck’s methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Anthropic: Claude Opus 5.5: # Foundry Hall to Close After Losing Its Largest Sponsor Foundry Hall, the Milwaukee nonprofit that has helped hundreds of young companies find their footing since 2018, will close its doors October 31. The startup accelerator is shutting down after Blank River Bank, its largest corporate sponsor, declined to renew its support. The bank's contribution made up 38% of Foundry Hall's annual budget, a gap the organization could not fill. It has enough money to operate through the end of October but not enough to run another full program. For executive director Camille Ross, the closure arrived less as a sudden blow than as the end of a long wait. "We kept waiting for the replacement check that never came," she said. The consequences reach well beyond the building. Six Foundry Hall employees will lose their jobs. Twelve startups now working out of its space have weeks to find somewhere else to go. A spokesperson for Blank River Bank said the decision reflected a shift in the bank's community investment priorities and was not a judgment on Foundry Hall's performance. By the numbers, that performance was substantial. Over seven years, Foundry Hall supported 212 startups, which together employ 1,140 people. For a city that has worked to build a homegrown technology and entrepreneurship scene, those figures represent payrolls, leases and a steady stream of people who chose to start something in Milwaukee rather than somewhere else. For the founders who passed through, the accelerator's value was often harder to count. Dev Malik remembers what his company looked like before he arrived. "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 distance, from a borrowed table to a room where investors took a meeting, is what accelerators are built to close. It is also what the 12 companies still in residence now risk losing. City officials have begun looking for a landing place. Deputy Mayor Susan Cole said the city is in talks with building owners and local universities about space for the current companies. She did not say when those conversations might produce an answer, and no replacement for Foundry Hall's programming has been announced. Until October 31, the work inside the building will go on as it has, with founders at their desks, staff running the final weeks of operations, and companies preparing to pack up. After that, the 212 startups that came through Foundry Hall will carry on without the place that helped launch them. Most will keep growing. The building where many of them began will not be there to see it.

Qwen: Qwen3.8 2.4T A95B: # The Last Shift at Foundry Hall Foundry Hall will close on October 31, ending a run that began in 2018 with a simple promise: give Milwaukee founders a place to build. The nonprofit startup accelerator has supported 212 companies, which together employ 1,140 people. Those numbers suggest a sturdy institution. The final weeks suggest something more fragile: a room where the lights are still on, but the calendar has run out. The closing is not the result of a sudden collapse. Foundry Hall has enough money to operate through October. What it does not have is the funding needed to begin another full program. Its largest corporate sponsor, Blank River Bank, provided 38% of its annual budget. When the bank declined to renew its sponsorship, the organization began searching for a new anchor funder. None arrived in time. Executive director Camille Ross described the waiting as the hardest part. “We kept waiting for the replacement check that never came,” she said. That absence reshaped the building’s daily life. Six employees will lose their jobs. Twelve startups currently using Foundry Hall’s workspace must leave, scattering teams that had settled into desks, meeting rooms, and the casual hallway exchanges that often turn into partnerships. Some companies may find other space; others will lose more than square footage. They will lose the proof that someone had decided their work deserved a home. For Dev Malik, a founder who passed through the program, the building represented a threshold. He recalled that his first investor meeting took place there, after a period of pitching from the back table of a coffee shop. His memory captures what Foundry Hall offered beyond programming: legitimacy, an address, and a place where ambition could be spoken aloud without apology. Blank River Bank said its decision reflected a shift in community investment priorities and was not tied to Foundry Hall’s performance. That explanation may be accurate, but it does little to soften the result: a central support removed from an organization that had built measurable momentum. City officials are trying to manage the landing. Deputy mayor Susan Cole said the city is talking with building owners and local universities about possible space for the displaced companies. Those conversations matter, but they are happening after the closure is already set. In the meantime, Foundry Hall is doing what accelerators do: preparing founders for what comes next, even as its own future ends. The last cohorts will leave with introductions, lessons, and perhaps the unsettling knowledge that support systems can vanish faster than the companies they are meant to help. When the doors close at the end of October, Milwaukee will not lose only a building. It will lose a signal—a place that told entrepreneurs they belonged in the city’s economic story. The challenge now is whether that signal survives without it.

Anthropic: Claude Opus 5.5: # Startup Spotlight: LedgerLark Wants AI to Handle the Cleanup After the Trade *The company builds agents that flag reconciliation errors at small investment firms and propose fixes. A person still approves every change.* Most conversations about trading technology focus on the moment a trade is placed. LedgerLark is interested in what happens afterward: the confirmations, custodian records and internal ledgers that are supposed to agree with one another and sometimes don't. ## A copied account number Co-founder Mina Batra spent seven years in operations at a regional brokerage. She began exploring the idea after watching a six-person team spend three days untangling a single trade mismatch. The cause was a copied account number. The episode illustrates a familiar problem in post-trade work. Errors are often small and mundane, but finding them means comparing records across systems that weren't designed to work together. Large institutions typically have dedicated software and staff for this. At smaller firms, the work can fall to a handful of people working through spreadsheets and exception reports. Co-founder Owen Carr brings a different vantage point. He built reconciliation systems at a large bank. Together, the founders have seen the problem from the operations desk and from the engineering side. ## What the product does and doesn't do LedgerLark's AI agents compare trade confirmations, custodian records and a firm's internal ledgers, looking for discrepancies. When they find one, they propose a correction, which a person then reviews and approves. The limits are as important as the features. LedgerLark does not execute trades or move money. That boundary may reassure prospective customers, since it constrains what can go wrong if an agent misreads a record. It also positions the product as support for operations staff rather than a substitute for their judgment. Whether the human-approval step works well in practice is worth watching. If the agents generate many proposals, reviewers could face a new kind of backlog or start approving suggestions without much scrutiny. How clearly the system explains its reasoning may shape how useful it turns out to be. ## Early signs LedgerLark has five paying customers, with annual contracts ranging from $36,000 to $110,000. The company says one customer reduced its unresolved daily exceptions from 70 to 18. That figure comes from LedgerLark and has not been independently verified. Results at one firm may not carry over to others, since exception volumes depend on trading activity, counterparties and existing processes. The company has raised a $3.5 million seed round led by Motive Row. ## The harder questions LedgerLark is entering a field with established players. Reconciliation vendors have sold software to financial firms for years, and large banks often rely on internal systems. If incumbents add similar AI capabilities, a newcomer will have a harder time standing out on that basis alone. The focus on small investment firms could give LedgerLark some room if those firms are poorly served by tools built and priced for larger institutions. But smaller firms also have tighter budgets, and with five customers so far, each sale carries weight. Sales cycles are another constraint. Financial institutions may require lengthy security and procurement reviews before adopting new software, particularly tools that touch account data. Even interested firms could take months to sign, which could slow growth regardless of how well the product performs. ## What to watch The next stage will test whether results like the one LedgerLark reports hold up across more customers, and whether operations teams come to trust the agents' proposals. For now, the company is betting that a narrowly scoped tool with a human checkpoint can earn a place in back offices where a single copied digit can still cost days of work.

Qwen: Qwen3.8 2.4T A95B: # Startup Spotlight: LedgerLark Tames Post-Trade Paperwork for Small Investment Firms LedgerLark is building AI agents for a part of finance that rarely attracts glamour but can consume entire teams: post-trade operations. The startup’s software compares trade confirmations, custodian records and internal ledgers, then proposes corrections for human approval. It does not execute trades or move money, a boundary that may matter to compliance officers evaluating how much autonomy to give automated systems. The insight comes from co-founder Mina Batra, who spent seven years in operations at a regional brokerage. She began exploring the idea after watching a six-person team spend three days resolving a trade mismatch caused by a copied account number. That kind of error is not dramatic, but it can be expensive in time, risk and client service. Co-founder Owen Carr previously built reconciliation systems at a large bank, giving the team familiarity with both the manual workflow and the legacy technology landscape. LedgerLark’s pitch is aimed at small investment firms that may not have the budget or engineering staff of a global bank. Instead of replacing existing systems, the agents sit across records and flag discrepancies. The company says one customer reduced its unresolved daily exceptions from 70 to 18. That figure is early evidence rather than proof of broad impact, and it depends on the client’s starting conditions, data quality and how exceptions are categorized. Early traction is modest but concrete. LedgerLark has five paying customers, with annual contracts ranging from $36,000 to $110,000. For a seed-stage company, those numbers suggest firms are willing to test the product against real operational pain. The price range also positions the software between spreadsheets handled internally and larger enterprise platforms, though it remains to be seen whether the product can expand into more complex operations. The company has raised a $3.5 million seed round led by Motive Row. The capital will likely be used to deepen integrations, strengthen security controls and support customer onboarding. In financial services, however, selling can be slow. Institutions may require lengthy security reviews, procurement processes and proof that the software can handle sensitive data without introducing new risks. LedgerLark’s human-approval design is a sensible response, but buyers will still ask how the model handles errors, audit trails and changing data formats. The agents’ recommendations also need trust. An AI system that misreads a confirmation or mislabels a legitimate exception could create more work, not less. LedgerLark will have to show that its software can explain why it proposes a change and how it reached a conclusion. That matters in an industry where operations staff are accountable for every entry and regulators may ask for documentation. Competition is another open question. Established reconciliation vendors already serve parts of this market, and many banks rely on internal software built over years. Those systems may be clunky, but they are also embedded in regulated workflows. LedgerLark will need to show that its agents can learn quickly enough to be useful across different custodians, asset classes and reporting standards. The opportunity is not that AI will magically fix back offices. It is that small firms may be underserved by tools built for much larger institutions. If LedgerLark can reduce manual checking without overpromising autonomy, it may find a practical niche. The next test will be whether its early results hold across more customers, more markets and the slower rhythms of financial compliance.

Anthropic: Claude Opus 5.5: C

Qwen: Qwen3.8 2.4T A95B: C

Anthropic: Claude Opus 5.5: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

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 Opus 5.5: # HelioDock Launches Night Shift to Let Operators Manage Warehouse Robots Remotely *The autonomous robot maker also closed a $24 million Series B, though one beta customer says the software is only as useful as the network beneath it.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor and manage their robot fleets from off site. The company also announced a $24 million Series B round led by North Coast Ventures. Night Shift targets one of the persistent headaches of warehouse automation: robots that stall when no one is around to get them moving again. HelioDock says customers who tested the software saw 17% fewer unplanned robot stoppages, a figure that has not been independently verified. Testers also found that the remote tools depend heavily on reliable connectivity, which can be scarce at rural facilities. ## A Fix for the 2 a.m. Call HelioDock is pitching Night Shift as a way to resolve routine robot problems from a distance instead of sending a person to the floor. "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. Velez's example captures the scenario the product is built around. An overnight disruption may be minor, but if the robot cannot sort it out on its own, someone still has to step in. Night Shift is meant to let operators see what their fleets are doing and intervene remotely when needed. The software spent 14 months in a private beta with nine logistics companies running a combined 310 robots. The 17% reduction in unplanned stoppages comes from that group, according to HelioDock. The company has not released independent data to support the number. ## Connectivity Sets the Limits Aaron Pike of Fulcrum Logistics, one of the beta customers, said Night Shift performed well in facilities with strong connectivity. During LTE outages at rural sites, he said, the system became less useful. That reflects how the software handles a lost connection. When the network fails, robots revert to their locally configured behavior and remote control becomes unavailable. The fallback means the robots keep operating under their existing settings instead of stopping when the link drops. It also means the remote oversight Night Shift is designed to provide disappears until the connection returns. For operators, the practical value of the software may therefore vary from site to site. A warehouse with dependable network coverage stands to get the full benefit. A facility that relies on rural LTE may find that remote access is unavailable during the same outages when an off-site manager would most want it. ## Pricing and Competition Night Shift will cost $90 per robot per month, or $1,080 per robot per year, and becomes generally available September 15. At that list price, a fleet the size of the entire beta group, 310 robots, would cost about $27,900 a month. The product enters a category with established alternatives. Relay Robotics and YardOS both offer fleet-management products, so HelioDock is competing for operators who already have other options for overseeing their robots. ## New Funding The Series B brings HelioDock's total funding to $39 million, meaning the company had raised $15 million before this round. North Coast Ventures led the new financing. HelioDock's case to customers will rest largely on its stoppage claim. The company says the 17% reduction shows that remote oversight can keep fleets running with fewer interruptions. Until that figure is independently verified, prospective buyers will be relying on HelioDock's own account of the beta results and on reports from testers such as Pike, whose experience suggests that results depend on the quality of a site's network connection.

Qwen: Qwen3.8 2.4T A95B: # HelioDock Launches Night Shift for Remote Warehouse Robot Monitoring *The software arrives after a 14-month beta with nine logistics companies and a $24 million Series B led by North Coast Ventures.* HelioDock, a maker of autonomous warehouse robots, launched Night Shift on Thursday, software that lets operators monitor robot fleets remotely. Alongside the product release, the company announced a $24 million Series B led by North Coast Ventures. HelioDock said the financing brings its total funding to $39 million. The announcement links a new software product, beta performance figures and subscription pricing as HelioDock prepares to make Night Shift generally available. The product is aimed at reducing situations where a person must go to a warehouse to deal with a robot that has stopped or become confused by a pallet. The Series B gives HelioDock additional disclosed capital as it moves Night Shift from private testing to a public release date. North Coast Ventures led the round, according to the company. The $24 million raised in this round is part of the $39 million total funding HelioDock now reports. The timing ties the funding announcement to the product’s release schedule. HelioDock is making Night Shift generally available on September 15, after the beta period. HelioDock supplied a quote from CEO Mara Velez describing the operational issue behind the software. “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 quote captures HelioDock’s description of the problem Night Shift is meant to solve: a single robot issue prompting an off-site visit. Night Shift has been tested with customers for more than a year. HelioDock said the private beta lasted 14 months and included nine logistics companies. Those companies operated the software with 310 robots, giving HelioDock a test base before the product’s public release. The beta was not limited to one company or one site. Nine logistics companies participated, and the software was deployed across 310 robots over the

Anthropic: Claude Opus 5.5: # QueueZero's AI Meeting Assistant Turns Audience Remark Into Tasks for Its CEO QueueZero launched Proxy, an AI meeting assistant, on Wednesday, and the launch demonstration offered an unscripted look at how the software decides what counts as an instruction. Proxy is designed to attend meetings, produce notes and assign follow-up work to software agents. It is aimed at the routine administrative work that follows many meetings: capturing what was discussed and turning it into action items. ## An unplanned test During the demonstration, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the remark as an instruction. It created two tasks, both assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. Of the two assignments, only one survived review by its assignee. The customer's underlying request was fulfilled. Proxy correctly identified a concrete follow-up item buried in an informal comment. But it also converted a figure of speech into a literal assignment, which shows the difficulty of building software that listens to an entire room and decides which words should become work. ## Where the guardrails sit The episode highlights a central design question for AI tools that act on what they hear: how an assistant separates a directive from an offhand comment, and what it is allowed to do once it concludes that it has received one. QueueZero says Proxy treats internal and external actions differently. According to the company, Proxy requires confirmation before an external message is sent. Internal tasks, however, can be created automatically. That automatic behavior is what produced the two tasks on stage. Administrators can disable automatic task creation, the company says. That distinction limited the consequences of Wednesday's misreading. The result was an internal task that a person reviewed and deleted within moments, not a message sent outside the organization. In a busier setting, however, automatically generated tasks may not get the same immediate attention. That could make the administrator setting an important consideration for organizations deciding how much independence to give the tool. ## Pricing and availability Proxy costs $30 per user per month. QueueZero says the product will enter public beta next week. The beta will give organizations outside the launch event their first chance to test how Proxy interprets their own meetings and whether its default settings fit their workflows.

Qwen: Qwen3.8 2.4T A95B: # QueueZero’s AI Meeting Assistant Proxy Turns Audience Remark Into Task List QueueZero on Wednesday introduced Proxy, an artificial-intelligence meeting assistant designed to attend meetings, take notes and turn discussion into follow-up work assigned to software agents. The product, priced at $30 per user per month, is scheduled to enter public beta next week. The company demonstrated the assistant by placing it in a live meeting scenario. Like other AI assistants, Proxy is intended to reduce administrative work by listening to conversations, summarizing outcomes and generating tasks. But the demonstration briefly showed how literal such systems can be when a remark is made near the assistant’s listening range. During the event, a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy treated the comment as an instruction. It created two tasks assigned to QueueZero chief executive Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed at the first task, completed the second by sending the pricing sheet, and deleted the first. The first task had a short lifespan, though it briefly solved a meeting problem older than artificial intelligence. QueueZero described the moment as an example of how Proxy can convert meeting chatter into actionable items, though it also highlighted the need for controls when AI systems parse spoken language in real time. The company said Proxy requires confirmation before any external message is sent. Internal tasks, however, can be created automatically, a design intended to let teams capture follow-up work without waiting for a human to review every item. Administrators can disable automatic task creation if they prefer a stricter approval process. The episode is likely to resonate with companies testing AI agents in workplace software. Meeting assistants are increasingly being connected to calendars, project-management systems and communication tools, where a single misplaced sentence can become a ticket, a reminder or a message draft. Vendors are trying to balance speed with safeguards, particularly as these systems move from drafting responses to taking action. For QueueZero, the launch was an opportunity to show that Proxy can move beyond transcription. Its pitch is that meetings should produce software-executed follow-through, not just another document that no one opens. The accidental task list provided a public demonstration of that mechanism, along with a reminder that AI assistants need clear boundaries between comment and command.

Anthropic: Claude Opus 5.5: **6** 1. `#Flood software moves upstream` has no space after `#`, so it doesn't follow the `# Headline` format. 2. `The company is expanding its pilot.` is the dek, but it isn't italicized. 3. `##Details` has no space after `##`, so it doesn't follow the `## Heading` format. 4. `Read [the announcement](www.example.com/news).` uses a link that isn't an absolute HTTPS URL. 5. `* The first pilot begins Monday.` starts the list item with an asterisk instead of a hyphen. 6. `<div>The second pilot begins Friday.</div>` uses raw HTML.

Qwen: Qwen3.8 2.4T A95B: 6

Anthropic: Claude Opus 5.5: # The Top Two Support Models Tied on Quality. One Costs Far Less. When researchers set five language models loose on 200 real customer-support tickets, the headline result wasn't which model won. It was that the top two effectively didn't separate, and one of them costs less than half as much to run. Model North posted the highest average score, 84.2. Model Vale came in at 83.8. The researchers found that gap was not statistically significant. In other words, the test can't reliably distinguish them on answer quality. On other measures they differ clearly: Vale's estimated cost is $0.74 per 1,000 tickets versus North's $1.90, and Vale answered 18% faster. For a team choosing between the two, that reframes the decision. The question is less "which model is best?" and more "is there a specific reason to pay roughly 2.5 times as much for North?" ## How the test worked The researchers used 200 anonymized support tickets. Each model produced one proposed answer per ticket. Three experienced support managers then graded the answers without knowing which model wrote which response. The score combined three qualities that matter in support work: whether the answer was factually correct, whether it actually helped the customer and whether the tone was appropriate. Blind grading by practitioners is a meaningful strength here. It reduces the chance that reviewers favor a brand name, and it measures answers against the judgment of people who handle real customers. Combining three qualities into one score has a trade-off, though. A single average can hide differences. A model that is slightly warmer in tone but occasionally wrong on facts could land near one that is accurate but curt. The published results don't break the score into its components, so readers can't tell which kind of strength drives each model's number. ## Where North and Vale differ The two leaders did show different strengths by ticket type. North performed best on billing disputes, while Vale performed best on technical troubleshooting. Those findings are useful as signals, but they deserve caution. The results don't say how many tickets fell into each category or whether the category-level gaps were statistically meaningful. With 200 tickets split across several types, the subsets may be small. A support team that handles mostly billing disputes might reasonably test North further, and a team buried in technical issues might lean toward Vale. Neither should treat this benchmark as settling the matter. ## The cost picture, in context The cost gap between North and Vale is large in relative terms. In absolute terms it's modest. At the estimated rates, handling one million tickets would cost about $1,900 with North and $740 with Vale. For a high-volume operation, that difference adds up. For a smaller team, it may be negligible next to staffing costs, and factors like integration effort or the speed advantage could matter more. Further down the list, the trade-off becomes more pronounced. Ember, Kite and Harbor are all cheaper still. Harbor costs just $0.08 per 1,000 tickets. But quality falls as price does: Ember scored 79.1, Kite 76.4 and Harbor 71.0. Whether those models are "good enough" depends on how answers are used. A draft that a human agent reviews before sending can tolerate more errors than a reply that goes straight to a customer. ## A shared blind spot One weakness cut across every model. All five performed poorly when a ticket depended on an image attachment, such as a screenshot of an error message. In software support, customers send screenshots constantly. Any team deploying these models should plan for that gap, perhaps by routing image-dependent tickets to human agents. ## What the benchmark doesn't tell you The test has clear boundaries. All tickets were in English and came from a single software company. Results could look different for another product, another industry or another customer base with different kinds of questions. The researchers also didn't test agentic actions, tool use or multilingual support. That leaves out a lot of what modern support automation involves: looking up account details, issuing refunds, updating tickets or responding in a customer's own language. The benchmark measures one thing well: the quality of a single written answer to an English-language ticket. ## The takeaway Within those limits, the clearest lesson is that the most expensive model isn't automatically the best value. North and Vale delivered statistically indistinguishable quality, and Vale did it faster and cheaper. Teams evaluating either should run their own tickets through both, paying particular attention to the ticket types they see most often and to the image-heavy cases that tripped up every model in this test.

Qwen: Qwen3.8 2.4T A95B: # Vale Matches Model North on Support Quality at a Much Lower Cost The clearest finding from a new benchmark of five language models for customer support is not that one model dominated, but that the top two were effectively indistinguishable on overall quality—while one cost far less to run. Model Vale posted an average score of 83.8, statistically tied with Model North’s 84.2. Yet Vale’s estimated cost was $0.74 per 1,000 tickets, compared with $1.90 for North, and it answered 18% faster. For teams choosing a model to draft routine support replies, that suggests the most expensive option may not offer the best value. The benchmark evaluated how well the models turned anonymized customer-support tickets into proposed answers. Researchers used 200 tickets from one software company; each model produced one reply per ticket. Three experienced support managers graded the answers blindly, combining factual correctness, usefulness, and tone. That setup matters: blind grading reduces stylistic or brand bias, and the scoring reflects what support teams actually need, not just fluent prose. But the methodology also sets limits. Two hundred tickets can reveal broad patterns, but it cannot make small score gaps decisive. The researchers reported that North and Vale’s difference was not statistically significant, so the 0.4-point gap should be treated as noise. The results also come from English-language tickets in a single company’s support environment, and the test did not include agentic actions, tool use, or multilingual support. It also cannot show how models would handle unusual products, seasonal spikes, or highly emotional complaints. Within that overall tie, the models showed different strengths. North performed best on billing disputes, where refunds, invoicing details, and customer sensitivity make precision and tone especially important. Vale performed best on technical troubleshooting, where clear steps and plausible diagnostics may matter most. A support organization with a heavy billing workload might still prefer North despite its higher cost. One with more technical questions may find Vale’s speed and price more attractive. The other models trailed by wider margins: Ember scored 79.1, Kite 76.4, and Harbor 71.0. Their costs were lower—$0.31, $0.22, and $0.08 per 1,000 tickets, respectively—pointing to a familiar trade-off. Cheaper models may reduce inference expense, but they may also require more human editing or correction. The benchmark does not measure downstream handling time, so the true operational cost of lower-scoring answers is unknown. The estimated costs should be treated as decision inputs, not final verdicts. They describe inference pricing per 1,000 tickets, not the full cost of deployment. Support automation also requires integration, monitoring, policy controls, and quality assurance. If a cheaper model creates more escalations, agents may spend more time correcting mistakes than they save on inference. If a slightly more expensive model needs less review, it may deliver better net productivity. The benchmark provides useful evidence about draft quality, but it does not capture these operational effects. The most important weakness cut across all five models: they performed poorly when a ticket depended on an image attachment. Many support issues rely on screenshots, error pages, invoices, or configuration panels. If a model cannot interpret the relevant image, it may produce a confident but incomplete or incorrect reply. Until multimodal performance improves, human review remains essential for attachment-heavy tickets. The practical takeaway is that support teams should not rank models by average score alone. North and Vale form the top tier,

Anthropic: Claude Opus 5.5: # Arcwell's Relay Suggests Answers to Support Agents and Flags Customers Likely to Cancel Arcwell is betting that customer support agents already have most of the answers they need. The trouble is finding them. The company has announced Relay, an artificial intelligence platform that analyzes customer support conversations and suggests responses to agents as they work. Relay also tries to identify customers who may be at risk of canceling and produces summaries, a feature aimed at cutting the routine work that fills an agent's day. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Arcwell tested Relay with 12 companies over six months. According to the company, agents who used the platform closed tickets 14% faster. That is a meaningful gain in a field where speed is closely tracked, but it is Arcwell's own figure, and the company has not published retention results from the trial. That leaves open one of the most commercially interesting questions about the product: whether flagging at-risk customers actually helps companies keep them. Relay costs $65 per agent each month. For a large support operation, that per-seat price adds up quickly, and buyers will likely want to weigh it against the efficiency gains Arcwell reports. The platform also depends on access to sensitive material. To generate its suggestions, Relay needs a company's support transcripts and internal help documents, records that can contain customer details and proprietary information about products and policies. Arcwell is trying to get ahead of those concerns. The company says customer data is encrypted and is not used to train shared models, so one client's conversations should not shape the suggestions another client receives. Companies can also choose how long their data is kept, with retention periods of either 30 or 90 days. Those assurances address some of the most common objections to bringing AI tools into customer service, where businesses are often wary of handing conversation logs to outside vendors. Whether they are enough will depend on how much risk each company is willing to accept. For now, Relay's pitch rests on a familiar promise: help agents find answers faster so they can spend more of their time on the customer in front of them.

Qwen: Qwen3.8 2.4T A95B: # Arcwell Launches AI Support Platform Relay Arcwell has launched Relay, a new artificial intelligence platform that reads customer support conversations and suggests answers to agents, describing the release as a significant development for the company and the broader customer service industry. Relay’s features go beyond answer recommendations. The platform can identify customers who may cancel and create summaries of support conversations, signs

Anthropic: Claude Opus 5.5: # KiteBridge Update Sends $2.8 Million in Duplicate Reimbursements Ninety-one companies paid out $2.8 million they did not owe on Tuesday after a software update at expense-management firm KiteBridge issued 13,400 duplicate reimbursements. Months earlier, the company had told customers its system was built to prevent exactly that kind of error. By Wednesday evening, $2.3 million had been recovered. The remaining $500,000 was still pending. ## A routine update, then a cascade KiteBridge makes software that businesses use to track and repay employees' work expenses. The company deployed version 4.2 of its platform at 8:45 a.m. Tuesday. The trouble started when a payment processor, the outside service that actually moves money between bank accounts, timed out. In other words, it failed to respond within the expected window. When a payment request goes unanswered, software commonly sends it again. KiteBridge's fallback system is meant to keep payments flowing when the primary route falters. Instead, it treated those repeated attempts as brand-new reimbursement requests, so each retry became another payment. The first customer complaint reached KiteBridge at 9:12 a.m., 27 minutes after the release. The company suspended automated payments at 10:03 a.m., roughly an hour and 20 minutes after the update went live. In that window, the duplicate payments averaged about $209 each. Employers bore the cost. The duplicate payments were funded by the affected companies, and no employees had money withdrawn from their own accounts. ## A promise of "exactly-once" payments The failure stands out because KiteBridge had said this problem was already solved. A changelog published in April said version 4.0 had introduced "exactly-once payment protection." A changelog is the running list of changes a software company publishes alongside its updates. In payment systems, "exactly-once" describes a guarantee that each transaction is processed one time and only one time, no matter how often a request is retried after a network hiccup or timeout. It is a safeguard against the scenario that unfolded Tuesday. It is not yet clear why the protection introduced in version 4.0 did not stop the duplicates created under version 4.2. ## A slow disclosure KiteBridge's public incident page is the status site where it reports outages and service problems to customers. The page initially described the issue only as delayed reimbursements. It did not mention duplicate payments until six hours after automation was suspended, around 4 p.m. Tuesday. By then, the duplicate payments had stopped for nearly the entire business day, and affected companies had been given an incomplete picture of what had gone wrong. ## Apology and next steps Chief executive Lena Ortiz apologized and said the company would publish a technical report on the incident. For the affected companies, the immediate concern is the $500,000 that has not yet been recovered. For KiteBridge, the report will need to answer a harder question: how a system advertised as guaranteeing one payment per request ended up making thousands of extra ones in little more than an hour.

Qwen: Qwen3.8 2.4T A95B: # Software Failure Turns Expense Reimbursements Into $2.8 Million Duplicate-Payment Incident A routine software update at KiteBridge turned a basic business process into a costly error, as the expense-management company’s system issued 13,400 duplicate reimbursements totaling $2.8 million before automated payments were shut down. The failure affected 91 corporate customers, triggered a recovery effort, and raised questions about a payment safeguard KiteBridge had previously promoted. KiteBridge makes software that helps companies manage employee expenses and reimbursements. The trouble began Tuesday at 8:45 a.m., when the company deployed version 4.2. The first customer complaint arrived at 9:12 a.m., indicating that something had gone wrong shortly after the update reached production. The technical problem involved the interaction between KiteBridge’s software and a payment processor. When the processor timed out, KiteBridge’s fallback system did not simply wait or resume the original request. Instead, it treated retries as new reimbursement requests. For a general reader, that means the system mistook a second attempt to complete one payment for a separate instruction to pay again. The result was duplicate reimbursements, not just delayed ones. Because expense platforms connect approval workflows, payment processors, and corporate accounts, a retry error can multiply quickly once each new attempt is accepted as valid. The episode also underscores the difference between stopping a failure and unwinding its financial consequences. KiteBridge suspended automated payments at 10:03 a.m., a little more than an hour after the release. The company said the affected payments were funded by employers, not drawn from employees’ accounts. That distinction matters because workers were not debited personally; the financial exposure fell on the corporate customers whose reimbursement accounts were used. The company’s public messaging did not immediately reflect the full scope of the incident. KiteBridge’s incident page initially described the problem as delayed reimbursements. It first disclosed duplicate payments six hours after automation had been suspended, a gap that is likely to be a point of concern for customers watching how quickly the company identified and communicated the issue. KiteBridge also faces questions about an earlier product promise. In an April changelog, the company said version 4.0 had introduced “exactly-once payment protection.” In payment systems, that language usually describes controls intended to prevent a transaction from being processed more than once, even when errors force software to retry. Version 4.2 followed that release, so customers will likely want to know whether the protection failed, was not applied to the affected payment path, or was overcome by the timeout fallback. By Wednesday evening, KiteBridge had recovered $2.3 million of the duplicate payments. The rest remained pending, meaning the company was still working to reconcile and retrieve outstanding funds. Chief Executive Lena Ortiz apologized and said KiteBridge would publish a technical report. The report will need to explain not only when each stage occurred, but why a fallback designed to handle payment failures created thousands of duplicate payments instead.