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

Anthropic: Claude Opus 5.5: # KiteBridge Update Triggers $2.8 Million in Duplicate Reimbursements **A safeguard meant to prevent double payments failed after a routine software release, leaving 91 companies to recover money they never owed.** Ninety-one companies paid out about $2.8 million they did not owe on Tuesday after a software update from expense-management provider KiteBridge issued 13,400 reimbursements twice. By Wednesday evening, KiteBridge had recovered $2.3 million. The remaining $500,000 was still pending. The affected employers funded the duplicate payments, not their workers. No employees had money withdrawn from their accounts. ## How a delay became a double payment KiteBridge makes software that companies use to manage business spending. Employees submit expenses, and the system reimburses them on the employer's behalf through an outside payment processor that moves the money. The trouble began when the payment processor stopped responding in time, an event known as a timeout. When a timeout occurs, the software cannot immediately tell whether a payment went through, so it tries again. A well-built system makes those retries without paying anyone twice. KiteBridge's system did not. The timeout triggered its fallback system, backup logic meant to keep payments moving when the normal process runs into trouble. The fallback treated each retry as a brand-new reimbursement request. Instead of completing one payment, the software created another. ## A short window, a large bill The problem surfaced quickly. KiteBridge deployed version 4.2 of its software at 8:45 a.m. Tuesday, and the first customer complaint arrived less than half an hour later, at 9:12 a.m. The company did not suspend automated payments until 10:03 a.m., 51 minutes after that complaint and roughly an hour and a quarter after the update went live. The duplicate reimbursements were issued in that window. ## A protection that was supposed to exist The failure is notable because KiteBridge had said its software guarded against this kind of error. A changelog, the public list of changes that accompanies software releases, published in April said version 4.0 had introduced "exactly-once payment protection." The term describes systems designed to process each payment a single time, even when a request is repeated. The available reporting does not explain why that protection failed to stop the duplicates in version 4.2. It also does not say whether the fallback system was covered by it. ## Disclosure lagged behind events KiteBridge's public account of the problem was slow to reflect what had happened. Its public incident page first described the problem as delayed reimbursements. The page did not disclose the duplicate payments until six hours after the company suspended automation, around 4 p.m. Tuesday. For most of that day, customers checking the page would have seen a description of late payments rather than extra ones. ## What comes next CEO Lena Ortiz apologized for the incident and said the company would publish a technical report. That report may answer the central questions left by the episode: - Why did the fallback system treat retries as new requests? - Why did the "exactly-once" protection described in April not apply? - Why did the incident page initially describe the problem as delays? Meanwhile, KiteBridge and its customers are still working to recover the last $500,000.

Qwen: Qwen3.8 2.4T A95B: # KiteBridge’s Faulty Update Sent $2.8 Million in Duplicate Reimbursements A routine software deployment at KiteBridge turned into a costly processing error this week after the expense-management company’s system created 13,400 duplicate reimbursements worth $2.8 million, hitting 91 corporate customers and forcing the company to suspend automated payments. The trouble began Tuesday at 8:45 a.m., when KiteBridge released version 4.2 of its expense-management software. Within minutes, the update encountered a problem outside the core application: a payment processor timed out while handling transactions. In response, KiteBridge’s fallback system attempted to recover by resubmitting the affected payments. But instead of recognizing those resubmissions as retries, the system treated them as new reimbursement requests. For a general reader, the failure is similar to a bill-payment service that loses contact with a bank, tries the payment again, and then mistakenly records the second attempt as a separate charge. In KiteBridge’s case, the duplicate requests were approved and paid, multiplying quickly across customer accounts. The first customer complaint reached KiteBridge at 9:12 a.m., less than half an hour after the release. Over the next hour, the company worked to identify the source of the issue and stop further payments. At 10:03 a.m., KiteBridge suspended automated reimbursements, preventing the system from creating additional duplicates. By then, the damage was already significant. The company said the malfunction generated 13,400 duplicate payments totaling $2.8 million. Ninety-one corporate customers were affected. KiteBridge said employees did not have money withdrawn from their accounts; the duplicate payments were funded by employers, who are now awaiting recovery of the remaining funds. The company’s public communications initially lagged behind the operational response. KiteBridge’s public incident page first described the episode as delayed reimbursements. It did not disclose that payments had been duplicated until six hours after automated payments were suspended. That delay could raise questions for corporate customers relying on KiteBridge for real-time expense processing and financial reconciliation. The incident also draws attention to KiteBridge’s prior claims about payment safeguards. In an April changelog, the company said version 4.0 had introduced “exactly-once payment protection,” a technical control designed to ensure that a payment is processed only one time even when systems encounter errors, timeouts, or retries. Tuesday’s failure is likely to prompt questions about whether that protection covered the path taken by version 4.2, was bypassed by the fallback process, or failed under the specific conditions created by the processor timeout. By Wednesday evening, KiteBridge had recovered $2.3 million of the duplicated funds. The rest remained pending as the company worked with customers and payment partners to reverse or reconcile the remaining transactions. Chief Executive Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report explaining what went wrong. Such a report will likely be closely read by customers seeking assurance that the company can prevent a repeat of the error, especially if the same kind of payment-protection feature was already marketed as a safeguard. For now, the episode underscores how quickly a software failure can move from a technical glitch to a financial mess when automated payment systems are involved. A single timeout, handled incorrectly by a backup process, was enough to create thousands of erroneous payments and leave a company racing to recover millions.

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

Qwen: Qwen3.8 2.4T A95B: 6

Anthropic: Claude Opus 5.5: **C**

Qwen: Qwen3.8 2.4T A95B: C

Anthropic: Claude Opus 5.5: # Arcwell's Relay Uses AI to Coach Support Agents and Spot Customers Who May Leave Arcwell is betting that the answers customer service agents need are already in their companies' files. The problem, it says, is how long they take to find. The company's new platform, Relay, uses artificial intelligence to read customer support conversations and suggest responses to the agents handling them. It also summarizes those exchanges and tries to identify customers who may be preparing to cancel. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Arcwell says the tool produced results in testing. Twelve companies used Relay over six months, and agents working with it closed tickets 14% faster, according to the company. The pilot results leave a notable gap, however. Although Arcwell is pitching Relay as a way to flag customers at risk of leaving, it did not release any customer-retention results from the trial. That makes the speed figure the only performance number the company has disclosed, and it comes from Arcwell itself. Relay costs $65 per agent per month, so the bill grows with the size of a company's support team. For many buyers, the larger consideration may be what the software needs in return. To generate suggestions, Relay requires access to a company's support transcripts and internal help documents. Those records can hold sensitive customer details and proprietary information. Arcwell is trying to address those concerns from the start. The company says customer data is encrypted and is not used to train shared models. Customers can also choose whether their data is retained for 30 or 90 days. How much weight companies give those assurances will likely shape how readily they hand over their support histories. So will the question of whether Relay's cancellation warnings prove useful once more customers begin relying on them. For Arcwell, the launch is part of a broader push to help businesses improve customer service with AI. It is also a test of whether faster ticket times are enough to win them over.

Qwen: Qwen3.8 2.4T A95B: # Arcwell launches Relay, an AI platform that suggests support answers and flags at-risk customers Arcwell has launched Relay, an artificial-intelligence platform that examines customer-support conversations and suggests answers to agents, a product the company says can help close tickets faster. The announcement represents a significant development for Arcwell and for the broader customer service industry, where AI continues to reshape how companies handle support. Relay analyzes support exchanges and recommends responses using a company’s internal help documents. It also identifies customers who may cancel and creates summaries of support interactions. The platform is important because it combines three functions in one system: suggesting answers, flagging possible cancellations and generating summaries. Those features are designed to reduce the time agents spend searching across conversations, transcripts and internal materials, while automating the creation of summaries. CEO Priya Shah said the product addresses a familiar problem in support operations. “Support agents spend too much time searching for information that their companies already have,” Shah said. Arcwell said Relay was tested by 12 companies over six months. Agents using the platform closed tickets 14% faster, according to the company. Arcwell did not release customer-retention results, leaving open whether the cancellation alerts produced measurable results during the testing period. Relay costs $65 per agent each month. The platform also presents potential challenges because it requires access to support transcripts and internal help documents, material that may include sensitive customer information. 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. The launch marks a new chapter for Arcwell as it seeks to help companies improve customer service with AI. By bringing answer suggestions, churn alerts and summaries into day-to-day support workflows, Relay shows how artificial intelligence is being used to speed up customer support work, while also requiring companies to carefully weigh data access and retention decisions.

Anthropic: Claude Opus 5.5: # QueueZero's New AI Meeting Assistant Assigns Tasks to Its Own CEO During Launch **QueueZero on Wednesday launched Proxy, an artificial intelligence assistant designed to attend meetings on a user's behalf, and the product's debut offered an unplanned demonstration of how such tools interpret what they hear.** Proxy can join a meeting, produce notes and assign follow-up work to software agents. The company is positioning it as a way to reduce the manual work that typically follows a meeting, such as summarizing discussions, tracking commitments and routing tasks to the right place. During the launch demonstration, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy treated the remark as an instruction. It created two tasks assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. It was, by any measure, a fast turnaround on a customer request. ## A test of interpretation The episode highlights a central challenge for AI agents that act on spoken conversation. Meetings are full of offhand comments, jokes, hypotheticals and half-formed suggestions. An assistant that converts speech into action items must decide which statements represent real commitments and which do not. In this case, Proxy correctly identified a request for pricing information but also captured a remark that was not meant as a work assignment. The incident also illustrates how such systems may treat input from people outside an organization. The comment came from an audience member, yet it produced tasks assigned to the company's chief executive. ## Safeguards and settings QueueZero says Proxy requires confirmation before it sends an external message. Internal tasks, however, can be created automatically. That distinction means Proxy cannot contact outside parties on its own, but it can add items to internal workflows without a person approving each one first. Administrators can disable automatic task creation, giving organizations the option to require human review before anything Proxy captures becomes an assigned task. For companies evaluating the product, the launch raises practical questions: how Proxy distinguishes instructions from casual conversation, how it handles statements made by meeting guests, and how easily users can review and remove tasks they did not intend to create. Organizations with sensitive workflows may want to examine the default settings before deployment. ## Pricing and availability Proxy costs $30 per user per month. It enters public beta next week, when a broader group of users will be able to test how the assistant performs in their own meetings.

Qwen: Qwen3.8 2.4T A95B: # QueueZero’s AI Meeting Assistant Turns an Audience Remark Into Tasks for Its CEO QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant that can attend meetings, produce notes and assign follow-up work to software agents. The launch demonstration briefly placed its chief executive on the product’s task list. The company demonstrated Proxy during a launch event, showing how the assistant can listen to a discussion, summarize decisions and generate tasks for team members or automated systems. 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 treated the comment as a command rather than a passing remark. It created two tasks assigned to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. The moment provided a live example of both the promise and the caution required with autonomous meeting software. As companies push AI agents beyond chatbots and into workflow tools, small ambiguities in speech can quickly become formal actions inside a system. QueueZero said Proxy requires confirmation before any external message is sent. Internal tasks, however, can be created automatically, depending on how an organization configures the product. Administrators can disable automatic task creation, giving companies a way to limit how aggressively the assistant acts on comments heard during a meeting. That distinction is likely to be central to how enterprises adopt meeting agents. Automated note-taking has become a crowded category, with vendors competing to reduce the administrative burden of calls, status updates and project reviews. The next step is not only recording what happened, but initiating the work that follows. Proxy’s pricing is $30 per user per month. The product enters public beta next week, when more customers will be able to test how it handles real meetings, interruptions, side comments and the occasional instruction aimed at a colleague. For QueueZero, the launch incident was a mild stress test rather than a failure. The company showed that its assistant could act quickly, but also that a human executive still decides which tasks deserve completion. The task “Stop talking” was not among them.

Anthropic: Claude Opus 5.5: # Startup Spotlight: LedgerLark and the Hidden Cost of a Copied Account Number Much of the attention in finance goes to the moment a trade is made. Far less goes to what happens afterward, when confirmations, custodian records and internal ledgers all need to agree. When they don't, someone has to figure out why. LedgerLark, a young company building AI agents for post-trade operations, is betting that this unglamorous work is a good place for automation, especially at smaller investment firms. ## Where the idea came from Co-founder Mina Batra spent seven years in operations at a regional brokerage. The idea for LedgerLark took shape after she watched a six-person team spend three days resolving a single trade mismatch. The cause was a copied account number. The anecdote shows how a small data error can consume a lot of skilled time. Reconciliation problems are rarely complicated in principle. They are tedious because the evidence is spread across several systems, and someone has to trace a discrepancy back to its source. Batra's co-founder, Owen Carr, approaches the problem from the systems side. He previously built reconciliation systems at a large bank. Together, the two cover both the daily experience of handling exceptions and the technical work of building software to catch them. ## What the product does LedgerLark's agents compare trade confirmations, custodian records and a firm's internal ledgers. When they find a discrepancy, they propose a correction, and a human reviews and approves it. The limits are as important as the capabilities. According to the company, the product does not execute trades or move money. That boundary matters in a regulated industry where firms are wary of handing consequential decisions to software. With humans approving each fix, LedgerLark positions its agents as investigators and drafters rather than decision-makers. Whether that design gives firms enough comfort, or simply moves the bottleneck to the approval step, is something early customers will help reveal. ## Early traction LedgerLark says five firms are 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 is promising but should be read carefully. It comes from the company, describes a single customer, and doesn't on its own show how much staff time was saved, how accurate the proposed corrections were, or whether the result holds across firms with different systems and trade volumes. Five customers is a meaningful start for a company selling into financial operations. It is still a small sample. LedgerLark has raised a $3.5 million seed round led by Motive Row. ## The competitive picture LedgerLark is entering a field with established players. Reconciliation vendors have been selling software to financial firms for years, and large banks often run their own internal tools. The company's focus on small investment firms may be partly a response to that landscape. These firms may lack the resources to build custom systems or to justify enterprise-scale vendor contracts. The open question is whether AI agents offer a durable advantage over rules-based reconciliation tools, or whether incumbents can add similar features to products customers already trust. Established vendors have existing relationships and integration histories, which count for a lot in this market. ## The procurement hurdle Selling to financial institutions takes patience. Security and procurement reviews can be lengthy, and a product that touches records and ledgers will face scrutiny even if it never moves money. For a seed-stage company, long sales cycles can strain both runway and momentum. They also test whether a startup can meet the compliance expectations of a cautious customer base. ## What to watch LedgerLark's founders have identified a specific, recognizable pain point and built a product with deliberate guardrails. The next phase will show whether the results reported by one customer repeat across others, whether the human-approval model scales, and whether the company can move through procurement fast enough to grow. A copied account number started the company. Whether it can build a lasting business depends on how many firms decide that problem is worth paying to solve.

Qwen: Qwen3.8 2.4T A95B: # Startup Spotlight: LedgerLark Looks at the Back Office The unglamorous work that happens after a trade is settled can consume hours that small investment firms would rather spend on clients or portfolio decisions. LedgerLark is building artificial intelligence agents for this post-trade operations layer, starting with reconciliation tasks that often rely on spreadsheets, email chains, and manual checks. The insight comes from co-founder Mina Batra’s experience. She spent seven years in operations at a regional brokerage, where she saw how small errors could expand into multi-day projects. In one case that shaped the company, a six-person team spent three days resolving a trade mismatch caused by a copied account number. Co-founder Owen Carr brings a related perspective from building reconciliation systems at a large bank. Together, they are aiming at a problem that is familiar to operations staff but easy for outsiders to overlook: records that should match often do not, and finding the reason can be slow. LedgerLark’s product compares confirmations, custodian records, and internal ledgers, then flags discrepancies and proposes corrections for human approval. The company is careful about the boundaries of that role: the agents do not execute trades or move money. Instead, they sit in the workflow as a review layer, suggesting how mismatched items might be resolved while leaving final decisions to people. That design may matter in financial services, where firms are likely to be cautious about giving autonomous software control over assets or settlement instructions. Early traction is modest but specific. LedgerLark has five paying customers, with annual contracts ranging from $36,000 to $110,000. The company says one customer reduced unresolved daily exceptions from 70 to 18. That figure is a company-reported result, not an independently audited measure, but it points to the kind of operational improvement small firms may find attractive. For an operations team, fewer unresolved exceptions can mean less end-of-day pressure and fewer items carried into the next morning. The startup has also raised a $3.5 million seed round led by Motive Row. The funding should help LedgerLark develop the product further and navigate the sales process typical of financial institutions. That process may be one of the company’s bigger tests. Even small asset managers and brokerages often require security reviews, compliance sign-off, and procurement steps before connecting new software to sensitive data sources. A tool that reads confirmations and custodian records will need to earn trust before it becomes part of daily operations. Competition is another open question. LedgerLark is not entering an empty market. Established reconciliation vendors already serve financial institutions, and larger banks often have internal software built around their own processes. The opportunity may lie with smaller firms that lack the budget or engineering resources for heavy enterprise systems. But those customers may still expect LedgerLark to integrate cleanly with custodians, order management systems, and reporting tools. If integration is difficult, adoption could slow. The broader question is whether LedgerLark’s agents can handle enough edge cases to become dependable. Post-trade exceptions are rarely neat. They involve missing reference numbers, corporate actions, timing differences, partial fills, and instructions that change between parties. A copied account number may be the spark, but the surrounding data can be messy. LedgerLark’s value will depend on whether it can reliably identify the likely cause of a mismatch and suggest a correction that an operations professional trusts enough to approve. For now, the startup appears to have found a practical wedge: human-supervised AI for reconciliation, aimed at firms that need help but may not want a large platform. If LedgerLark can prove reliability and clear security standards, it may turn back-office cleanup into a repeatable software business. If not, it will remain one of many tools trying to make a complicated process slightly less painful.

Anthropic: Claude Opus 5.5: # HelioDock Launches Night Shift to Let Managers Oversee Warehouse Robots Remotely *The autonomous-robot maker also closed a $24 million Series B, though one early customer says the software is only as useful as the network beneath it* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor their robot fleets remotely. The company also announced a $24 million Series B round led by North Coast Ventures. The product is aimed at a specific problem in automated warehouses: a single stalled machine can require someone to be physically present, even in the middle of the night. HelioDock says Night Shift cut unplanned robot stoppages during a 14-month private beta. The product depends on a steady network connection, and at least one beta customer found its value dropped sharply when that connection failed. ## Solving the 2 a.m. problem HelioDock frames Night Shift around the kind of minor glitch that can stall an operation until a person arrives. "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. The beta involved nine logistics companies running a combined 310 robots. According to HelioDock, those customers experienced 17% fewer unplanned robot stoppages. The company has not had that figure independently verified, and it has not said how it measured the reduction. ## Strong signal required Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity. It became less useful, he said, during LTE outages at rural sites. That limitation reflects how Night Shift handles a lost connection. When the link drops, HelioDock's robots revert to their locally configured behavior, and remote control becomes unavailable until connectivity returns. The robots keep operating on their own settings, but the operator loses the ability to step in from afar. A confused robot at a rural facility during an outage could still require the drive across town Velez wants to eliminate. For logistics companies with sites in areas where coverage is unreliable, the software's usefulness may therefore vary from one facility to the next. Pike's experience suggests that buyers will need to weigh how dependable the network is at each location they plan to run it. ## Pricing and competition Night Shift will cost $90 per robot per month and becomes generally available September 15. At that rate, a 100-robot fleet would cost $9,000 a month, or $108,000 a year. A fleet the size of the full beta group, 310 robots, would run about $27,900 a month. Because pricing scales with every robot, the cost rises in step with the size of the fleet. That will put HelioDock's claimed reduction in stoppages at the center of any customer's calculation about whether the subscription pays for itself. HelioDock is entering a market that already has alternatives. Relay Robotics and YardOS both sell competing fleet-management products, giving warehouse operators other options for overseeing automated equipment. ## New funding The $24 million 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 investment. The funding and the product launch arrived together, as HelioDock moves from a closed test with a few hundred robots toward a paid release on a set timetable. The beta's nine participants offer a limited sample for a product the company now plans to sell broadly. Pike's account points to the conditions under which Night Shift may work best: facilities where the network stays up. At sites where it does not, the robots will keep running on their local settings, and the remote controls Velez is promoting will stay out of reach until the connection returns.

Qwen: Qwen3.8 2.4T A95B: # HelioDock Launches Night Shift to Monitor Warehouse Robot Fleets Remotely *The software emerges from a 14-month beta and arrives alongside a $24 million Series B led by North Coast Ventures.* HelioDock on Thursday launched Night Shift, software that lets operators monitor autonomous warehouse robot fleets remotely, giving the company a new product aimed at keeping overnight operations running without requiring managers to be physically present. Night Shift is intended to put remote oversight between autonomous robots and the human managers responsible for their performance. The product arrives as HelioDock moves from a private beta to general availability and reports a new round of funding. The launch pairs HelioDock’s robot business with a software layer intended to reduce the need for physical supervision. The company is selling Night Shift as a subscription tied to each robot. The release follows a 14-month private beta involving nine logistics companies and 310 robots. HelioDock said beta customers experienced 17% fewer unplanned robot stoppages. The company-supplied figure has not been independently verified, so it stands as HelioDock’s account of beta performance rather than a confirmed benchmark. The beta involved nine logistics companies and 310 robots over 14 months, giving the company a record of use across multiple customers before general availability. The results remain those supplied by HelioDock. HelioDock also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. The financing adds to HelioDock’s funding total as Night Shift moves from beta to general availability. The company makes autonomous warehouse robots, and Night Shift extends that hardware business into software that can be sold on a recurring basis. The central problem Night Shift addresses is response time. In a warehouse running robots, a single stopped machine can require attention from a person who is not on site. The software gives operators a remote view of fleet activity and a way to intervene when a robot encounters an obstacle or cannot complete a task. CEO Mara Velez said the product is meant to reduce the need for after-hours trips to a facility. “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. The beta tested the system inside working logistics operations. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity. He said it became less useful during rural LTE outages, a comment that identifies network access as a major condition shaping the product’s usefulness. When connectivity fails, HelioDock’s robots return to their locally configured behavior. Remote control becomes unavailable, leaving the robots to operate according to the settings already in place. In that situation, a remote operator cannot take direct control, and the robot’s response is determined by its local configuration rather than by an off-site intervention. That limitation does not eliminate the software’s function, but it defines where the product can deliver its full value. Facilities with reliable connections can use Night Shift as a remote monitoring layer. Sites affected by rural LTE outages may find that the robots continue operating locally, while the remote-management portion of the system is interrupted. Night Shift will cost $90 per robot per month and become generally available on September 15. The subscription price ties the software’s cost to fleet size. A customer with 10 robots would pay $900 per month, while a customer with 100 robots would pay $9,000 per month. Relay Robotics and YardOS offer competing fleet-management products, giving potential customers alternatives when evaluating remote monitoring. HelioDock’s launch gives logistics companies another option to compare against those systems as they decide how to supervise autonomous equipment.

Anthropic: Claude Opus 5.5: # Foundry Hall to Close After Losing Its Largest Sponsor Foundry Hall, the Milwaukee nonprofit that has served as a first home for hundreds of young companies, will close on October 31. The startup accelerator, which opened in 2018, lost its largest corporate sponsor when Blank River Bank declined to renew its support. The bank had supplied 38% of Foundry Hall's annual budget, a share large enough that its absence could not be absorbed. The organization has enough money to keep operating through October but not enough to run another full program. Executive director Camille Ross said the organization had hoped another funder would step in to cover the shortfall. None did. "We kept waiting for the replacement check that never came," she said. A spokesperson for Blank River Bank said the decision reflected a shift in the bank's community investment priorities, not any judgment about Foundry Hall's performance. By the measures accelerators usually use, that performance was substantial. Over seven years, Foundry Hall supported 212 startups, which together employ 1,140 people. Those jobs will outlast the organization that helped create them. Others will not. Six Foundry Hall employees will lose their positions when the doors close. The most immediate disruption falls on the twelve startups now working out of Foundry Hall's space. They must find somewhere else to go, and their deadline is the same as the organization's. For an early-stage company, a move can mean more than packing boxes. It means new rent and lost time at the stage when both are scarcest. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about space for those companies. She did not say when arrangements might be finalized or whether the city would help cover costs. What the city is unlikely to replace is the thing founders describe when they talk about the building. Dev Malik, who went through the program, said that before Foundry Hall he had been making his pitches from the back table of a coffee shop. Foundry Hall gave him a place where investors could find him. "My first investor meeting happened in that building," he said. Ross did not describe what comes after October 31, and Foundry Hall has announced no plans to reopen in another form. For now, the staff will run the remaining weeks with the money they have left. Twelve companies will begin looking for new desks, and six people will begin looking for new work. The building will empty at the end of the month. The companies it launched, and the 1,140 people they employ, will go on working elsewhere in the city.

Qwen: Qwen3.8 2.4T A95B: # The Last Lights in Foundry Hall The last day will be quiet. On October 31, Foundry Hall, the Milwaukee nonprofit that gave early-stage companies desks, advice, and a sense of place, will close its doors. It opened in 2018 with the simple idea that a city could grow more founders if it gave them somewhere to begin. Since then, it has helped 212 startups and, by its count, 1,140 jobs. The ending did not arrive with a dramatic collapse. It came with a renewal that did not happen. Blank River Bank, Foundry Hall’s largest corporate sponsor, supplied 38% of the organization’s annual budget. When the bank declined to renew, the hall had enough money to operate through October, but not through another full program. Executive director Camille Ross said, “We kept waiting for the replacement check that never came.” That sentence hangs over the final weeks. Six employees will lose their jobs. Twelve startups that have been using the workspace must leave. For a nonprofit built around beginnings, the arithmetic of ending is blunt: rent, salaries, programming, and the gap left by a single large partner. Blank River Bank, in a statement, said its decision reflected a shift in community investment priorities and was not tied to Foundry Hall’s performance. That distinction may matter to the record, but not to the people packing boxes. A budget line can move for reasons no one in the building controls, and still the building empties. Foundry Hall was never just square footage. Dev Malik, a founder who passed through, recalled having his first investor meeting in the building after once pitching from the back table of a coffee shop. It was an address where unproven companies could borrow credibility before they had any of their own. That is the kind of memory a room leaves behind: a door that once opened, a table where someone was taken seriously. The city is trying to soften the landing. Deputy mayor Susan Cole said Milwaukee is talking with building owners and local universities about space for the companies currently inside. Those conversations may help some startups survive the transition. They will not replace the ecosystem of a hall that made room for risk. By Halloween, the desks will be cleared and the last cohort will have gone. What remains is a ledger of companies, jobs, and first meetings, and a warning about how fragile nonprofit infrastructure can be. Foundry Hall is closing not because it failed, but because the support it counted on changed course. The startups it helped will keep going, somewhere else. The room will not.

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: # The Top Two Support Models Scored the Same. One Costs Less Than Half as Much. When five language models were asked to draft replies to real customer-support tickets, the headline result was not which model won. It was that the top two effectively tied, and one of them was far cheaper and faster. Model North posted the highest average score, 84.2. Model Vale came in at 83.8. The researchers found the gap between them was not statistically significant, which means the test cannot tell them apart in overall quality. The costs, however, differ sharply. North is estimated at $1.90 per 1,000 tickets, while Vale costs $0.74, about 39% of North's price. Vale also answered 18% faster. For teams evaluating these tools, that changes the question. The issue is less "which model is best?" and more "what, if anything, justifies paying more than twice as much for North?" ## How the test worked Researchers gave each of the five models the same 200 anonymized customer-support tickets. Each model produced one proposed answer per ticket. Three experienced support managers then graded the answers without knowing which model wrote them, which reduces the risk that reputation or expectations shaped the scores. Each answer received a combined score covering three qualities: - **Factual correctness:** whether the answer was accurate - **Usefulness:** whether it would actually help resolve the customer's problem - **Tone:** whether it read appropriately for a support interaction Because these are blended into one number, a model's average can hide trade-offs. An answer that is warm but slightly wrong and one that is accurate but curt could land near the same score. The overall rankings are useful, but they are not a breakdown of where each model is strong or weak. ## Where North and Vale differ The two leaders were not identical in every category. North performed best on billing disputes, and Vale performed best on technical troubleshooting. That matters more than the 0.4-point overall gap. A company whose support queue is dominated by billing complaints might reasonably test North further. A company fielding mostly technical questions has an even stronger case for Vale, which already holds the advantage on price and speed. The overall average treats all tickets as interchangeable, but real support queues rarely are. ## The cheaper tier involves a real trade-off Below the top two, the scores drop more noticeably. Ember scored 79.1 at $0.31 per 1,000 tickets. Kite scored 76.4 at $0.22, and Harbor scored 71.0 at $0.08. These models are inexpensive, and Harbor costs a small fraction of what the leaders do. But the gap between Harbor and the top pair, about 13 points, is large enough that it likely reflects a meaningful difference in answer quality. Whether that trade is acceptable depends on how the answers are used. A draft that a human agent reviews and edits is a different proposition from a reply sent directly to a customer. At these prices, cost per ticket is unlikely to be the deciding factor for most teams. The more important costs are agent time spent correcting weak drafts and customers who receive poor answers. ## A shared blind spot: images 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, that is a common situation. Teams adopting any of these models should plan to route image-dependent tickets to humans, or at least flag them for closer review, rather than assume the system can handle them. ## What the results don't show The benchmark is a reasonable starting point, but its scope is narrow: - **One company, one language.** All tickets were in English and came from a single software company. Results could differ for other products, industries, customer bases or writing styles. - **Only 200 tickets.** That sample is enough to separate large gaps, but too small to settle close calls. North and Vale's non-significant difference is a direct example. - **No agentic work.** The researchers did not test tool use or agentic actions, such as looking up account details, issuing refunds or updating records. Many modern support deployments depend on exactly those capabilities. - **No multilingual support.** Teams serving customers in other languages will need their own testing. - **One answer per ticket.** Each model got a single attempt, so the test does not capture how consistent a model is across repeated tries. ## The takeaway On this test, paying a premium for the top-scoring model bought no measurable overall improvement. Vale delivered statistically indistinguishable quality at well under half the cost, with faster responses. The more useful next step for most teams is a trial on their own tickets, in their own languages, with their own mix of billing and technical issues. The ranking table matters less than that.

Qwen: Qwen3.8 2.4T A95B: # Vale matches North on support quality at a much lower cost The strongest result in the benchmark is not that Model North beat Model Vale. It is that the two leaders were effectively tied. Across 200 anonymized customer-support tickets, North averaged 84.2 and Vale averaged 83.8 on a score combining factual correctness, usefulness, and tone. The researchers reported that the difference was not statistically significant, so the data does not show that North is truly better overall. What it does show is that Vale costs far less and works faster. Vale’s estimated cost was $0.74 per 1,000 tickets, compared with $1.90 for North. That is about 61% lower. Vale also answered 18% faster. For support teams, this matters because the goal is rarely to maximize a benchmark score by a fraction of a point. The goal is to help agents resolve customer issues reliably at scale. If two models provide similar answer quality, the cheaper and faster one is usually the better default. ## How the benchmark worked The test used 200 English-language tickets from one software company. Each model produced one draft answer per ticket. Three experienced support managers graded the answers blindly, without knowing which model wrote them. Their scores combined factual correctness, usefulness, and tone. That human evaluation is valuable because support is not purely factual: a correct answer that sounds dismissive or fails to move the customer forward may still be a poor reply. Readers should treat the scores as a relative measure, not a guarantee of production performance. Two hundred tickets can reveal broad patterns, but it cannot turn tiny differences into firm conclusions. That caution is especially important for North and Vale, whose gap is small and statistically indistinguishable. ## Where each leader stands out The overall tie hides useful category differences. North performed best on billing disputes, while Vale performed best on technical troubleshooting. That suggests model choice may depend on ticket mix. A team handling many payment, refund, or invoice questions might prefer North. A team facing more setup issues, errors, and product how-to questions might prefer Vale. Some organizations could even route tickets by category, if the added complexity is worth the expected improvement. ## The cheaper models are not close enough Ember, Kite, and Harbor cost less than the leaders, but their scores were lower: 79.1, 76.4, and 71.0 respectively. Harbor is extremely inexpensive at $0.08 per 1,000 tickets, yet its quality gap with the leaders appears substantial. In customer support, a low-cost draft that is often wrong or unhelpful can create extra work for agents and customers. Small differences among models should not be overinterpreted, but the drop from the top tier to the bottom tier looks large enough to matter. ## Limitations that matter The benchmark has important boundaries. It used English-language tickets from one software company, so results may not transfer to other industries, customer expectations, or regulatory contexts. The test also did not include agentic actions, tool use, account lookups, knowledge-base retrieval, escalation decisions, or multilingual support. Real deployments often depend on those surrounding systems as much as on the model itself. The most striking limitation is image attachments. All five models performed poorly when a ticket depended on an image. Support queues frequently include screenshots, receipts, error messages, or annotated diagrams, so this is not a niche weakness. For image-dependent cases, the results argue for human review, multimodal models, or a workflow that extracts the visual information before asking a text model to answer. ## Bottom line Vale is the clearest overall choice from this benchmark: it matches North statistically, costs much less, and answers faster. North may be worth considering where billing disputes dominate, while lower-cost models need stronger evidence before being trusted with customer-facing drafts. The safest reading is not that one model wins decisively, but that Vale offers the best balance of quality, cost, and speed within the test’s limits.

Anthropic: Claude Opus 5.5: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 Sessions *The page was open to anyone for four hours Monday. The company says it has found no sign of access beyond the researcher who reported it, but its investigation is not finished.* A Cloudnote debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday, the company said. Logs show the page held fragments from 1,842 user sessions. Each excerpt could include up to 200 characters of a prompt, enough for a sentence or two of whatever a user had typed. Full documents, passwords and payment information were stored elsewhere and were not on the page, Cloudnote said. The company removed the page 19 minutes after a security researcher notified it. Cloudnote said it has found no evidence that anyone other than the researcher accessed the page. That finding is preliminary. Cloudnote has not completed its investigation, and it is not yet known how the page came to be unprotected or how many individual people the 1,842 sessions represent, since one user could account for several. The affected sessions are a small fraction of activity on a service with 3.2 million registered users. But prompts often contain what people are actively working on, and for those whose words appeared on the page, the question now is whether anyone else saw them.

Qwen: Qwen3.8 2.4T A95B: # Cloudnote Says Debugging Page Exposed Prompt Excerpts From 1,842 Sessions *The page was accessible without authentication for four hours Monday, and Cloudnote says the investigation into possible access is not complete.* Cloudnote said it discovered that a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday, exposing limited fragments of user input at a service with 3.2 million registered users. The company said logs show that excerpts from 1,842 user sessions appeared on the page. Each excerpt could contain up to 200 characters from a prompt. Cloudnote said it removed the page 19 minutes after a security researcher notified it. Cloudnote said it has found no evidence that anyone besides the researcher accessed the page, but it has not completed its investigation. That leaves unresolved whether the excerpts were viewed, copied or retained beyond the researcher’s report. The company also has not said whether the researcher kept copies. The exposure matters because prompts can reveal personal plans, workplace drafts or sensitive questions, even when truncated. Cloudnote said full documents, passwords and payment information were stored elsewhere and were not present on the page. It did not say whether affected users will be notified if later review finds additional access. The incident underscores the governance challenges facing AI tools that process sensitive user inputs.

Anthropic: Claude Opus 5.5: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin startup's software combines drainage maps with rainfall forecasts to show local governments where maintenance crews are needed before a storm.* Raincheck, an Austin company that makes flood-risk software for local governments, announced that it has raised $8 million in a seed round led by Basin Ventures. Existing investor Long Mile also participated, bringing Raincheck's total funding to $10.5 million. The company plans to use the new funding to hire eight engineers and expand into Florida. Raincheck 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 company explains its approach in a [methodology overview](https://example.com/raincheck-methodology). Chief Executive Ana Flores said the software is meant to help cities prioritize work 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." 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. Raincheck competes with other flood-planning vendors, including HydroMap, as well as municipal engineering contractors.

Qwen: Qwen3.8 2.4T A95B: # Raincheck Raises $8 Million to Help Cities Anticipate Floods *The Austin company makes flood-risk software for local governments, combining drainage maps with rainfall forecasts to guide pre-storm maintenance.* Raincheck, an Austin company that makes flood-risk software for local governments, said today it has raised $8 million. The company maps drainage systems and combines those maps with rainfall forecasts, helping cities decide where to send maintenance crews before a storm arrives. The seed round was led by Basin Ventures and included participation from existing investor Long Mile. Raincheck has now raised $10.5 million in total. CEO Ana Flores said: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” More details 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, and 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.