Head to head: MiniMax: MiniMax M3 vs Anthropic: Claude Sonnet 5.5

MiniMax: MiniMax M3: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

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

MiniMax: MiniMax M3: **Cloudnote removes unauthenticated debug page after researcher tip, exposing 1,842 prompt excerpts** *A four-hour window left snippets of user inputs visible online; the company says no evidence suggests access beyond the reporting researcher.* Cloudnote, the AI notetaking service used by 3.2 million registered accounts, took down a debugging page on Monday after learning that it was publicly accessible without a password. The page displayed short excerpts from user sessions and remained reachable for roughly four hours before being removed, the company told reporters on Wednesday. Logs indicate the page contained excerpts from 1,842 user sessions, each capturing as many as 200 characters from a prompt. Cloudnote emphasized that the page did not include full documents, account credentials or payment information, which the company says are stored in separate systems. The exposure ended 19 minutes after an external researcher reported it to Cloudnote, according to a timeline the company provided. Cloudnote said it has not found evidence that anyone other than that researcher viewed the page, though it acknowledged the investigation is incomplete and the full scope of any access may not be knowable. The incident highlights the recurring risk that internal diagnostic tools, often left exposed during development, can become inadvertent data leaks. Cloudnote declined to answer questions about how the page became public or what steps it is taking to prevent a recurrence, citing the continuing review. The company has not said when it expects to complete its investigation or whether affected users will be notified directly.

Anthropic: Claude Sonnet 5.5: # Cloudnote debugging page exposed snippets of user prompts for four hours *The company says it has no evidence anyone but a security researcher viewed the page, but its investigation is not finished.* Cloudnote left an internal debugging page containing excerpts of customer prompts accessible without authentication for four hours on Monday, the company said. A researcher notified Cloudnote of the exposure, and the company removed the page 19 minutes later. Logs show the page contained excerpts from 1,842 user sessions. Each excerpt could include up to 200 characters of a prompt, roughly a sentence or two. Cloudnote has 3.2 million registered users, so the sessions amount to a small share of its base. It is not yet clear how many individual people they represent. The company says it has found no evidence that anyone besides the researcher accessed the page. Because Cloudnote has not completed its investigation, that finding could change. Cloudnote says full documents, passwords and payment information are stored elsewhere and were not on the page. That limits the potential harm, but it does not settle what the excerpts revealed. Prompts can contain personal or work-related details, depending on what users typed. It remains unclear whether any sensitive information appeared in the exposed text, and how the page became publicly reachable has not been explained.

MiniMax: MiniMax M3: # Austin's Raincheck Raises $8 Million to Modernize Flood-Risk Planning for Cities *Seed funding will help the drainage-mapping startup expand into Florida and double down on the messy, human work of storm preparation.* Raincheck, an Austin-based startup that builds flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The company maps municipal drainage systems and layers those maps with rainfall forecasts, giving cities a way to decide where to send maintenance crews before a storm hits. The round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has now raised $10.5 million in total capital. The company's approach reflects a harder truth about urban flooding than most data platforms acknowledge. > "Our customers usually know which streets flood," said CEO Ana Flores. "The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." Raincheck works with seven cities across Texas and Louisiana, though it declined to identify them because some contracts have not been publicly announced. Its forecasts function as recommendations; emergency managers retain control over deployments and public warnings. The company's [rainfall and drainage methodology](https://example.com/raincheck-methodology) outlines how those recommendations are generated. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will support the hire of eight engineers and an expansion into Florida.

Anthropic: Claude Sonnet 5.5: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin software company helps local governments decide where to send crews before a storm hits.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding, it announced today. The software maps drainage systems and combines those maps with rainfall forecasts. Cities use it to decide where to send maintenance crews before a storm. Basin Ventures led the round, and existing investor Long Mile also participated. Raincheck has raised $10.5 million in total. CEO Ana Flores said the software is meant to help with a specific problem: > "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 publishes [details of its methodology](https://example.com/raincheck-methodology) online. The company currently works with seven cities in Texas and Louisiana. It declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. 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.

MiniMax: MiniMax M3: # Arcwell's Relay Aims to Speed Up Customer Support With AI-Generated Answers Arcwell has launched Relay, an artificial intelligence platform that monitors live customer support conversations and recommends responses to human agents in real time. The product is built around a routine but costly problem: support representatives spend a meaningful share of each shift searching knowledge bases, prior tickets and internal documents for the right answer. "Support agents spend too much time searching for information that their companies already have," said CEO Priya Shah. Relay is designed to shorten that hunt. In a six-month beta with 12 companies, agents using the platform closed tickets 14% faster than before, according to Arcwell. The company declined to release customer-retention figures, but said Relay also flags customers showing early signs of churn and produces conversation summaries so agents don't have to take notes after every interaction. The platform is priced at $65 per agent per month. The system's usefulness depends on access. To generate suggestions, Relay ingests support transcripts and internal help documents. Arcwell says the underlying customer data is encrypted and is not used to train shared models. Clients can choose between 30-day and 90-day retention windows. The launch places Arcwell in a crowded field of AI support tools, where vendors are racing to package large language models for contact centers. The company is positioning Relay as a co-pilot for human agents rather than a replacement for them — a distinction likely to resonate with companies that have struggled to deploy fully automated support bots. What prospective buyers won't get from Arcwell, at least not yet, are the retention numbers that would help quantify Relay's most ambitious claim: spotting customers at risk of canceling. The 14% ticket-closure figure comes from the company itself, with no independent validation.

Anthropic: Claude Sonnet 5.5: # Arcwell's Relay Uses AI to Feed Support Agents Answers Arcwell has launched Relay, a platform that uses artificial intelligence to examine customer support conversations and suggest answers to agents. The software can also identify customers who may cancel and create summaries. The idea, according to CEO Priya Shah, is to stop agents from digging for information already sitting in company files. "Support agents spend too much time searching for information that their companies already have," Shah said. Arcwell's early evidence centers on speed. Twelve companies tested Relay over six months, and the company says agents using it closed tickets 14% faster. That number is Arcwell's own, and it speaks to only one of the product's jobs. The company did not release customer-retention results, a conspicuous gap for a tool that also promises to flag customers at risk of leaving. Relay costs $65 per agent each month, which would put a 20-person support team at $1,300 a month. ## What Relay needs in return The product's usefulness depends on access. Relay requires support transcripts and internal help documents, meaning it must see both what customers have said and what the company has written down for its own staff. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period. For buyers, those assurances will matter as much as the speed gains. A tool that reads support conversations is only as acceptable as the safeguards around it, and Arcwell's claims on that front, like its performance figures, come from the company itself. Companies weighing Relay will have to decide whether faster ticket resolution is worth giving software that level of access.

MiniMax: MiniMax M3: # QueueZero's AI Assistant Mishears Audience Quip as Managerial Directive at Launch Event QueueZero unveiled an artificial intelligence meeting assistant called Proxy on Wednesday, billing the tool as a way to automate note-taking and follow-up work in corporate meetings. The launch demonstration, however, produced an unintended moment that underscored both the capabilities and the literal-mindedness of the software. Proxy is designed to join scheduled meetings on a user's behalf, transcribe discussions, summarize decisions and assign subsequent tasks to other software agents. During the company's presentation, an audience member called out that someone should tell Chief Executive Evan Wu to stop talking and send the pricing sheet. Proxy, which was active during the demonstration, registered the comment as an instruction. The assistant generated two follow-up tasks assigned to Wu: "Stop talking" and "Send pricing sheet." Wu, who has led QueueZero since its 2022 founding, completed the second task on stage and deleted the first, drawing laughter from the room. QueueZero characterized the episode as a useful illustration of how Proxy interprets natural language. The company emphasized that the assistant requires explicit human confirmation before any external communication is dispatched, including outbound emails or messages to customers. Internal tasks, by contrast, can be created automatically based on meeting content. Administrators can also disable automatic internal task creation, and the company said it plans to expand its configuration options ahead of general availability. A QueueZero spokesperson described the default behavior as a starting point rather than a fixed setting, and noted that feedback during the beta period would inform additional guardrails. Proxy is priced at $30 per user per month and will enter public beta next week. QueueZero, headquartered in San Francisco, is positioning the product against a growing field of AI notetakers that have drawn scrutiny over recording consent and data handling. The company said Proxy supports standard consent prompts and offers options to redact sensitive segments from transcripts. Industry analysts said the launch incident, while awkward, reflects a broader challenge for vendors of agentic software: balancing automation with predictable behavior. Tools that act on conversational input can mistake casual remarks for directives, and developers must decide how much to filter before execution. QueueZero said it expects to ship general availability later this quarter, pending the outcome of the beta program.

Anthropic: Claude Sonnet 5.5: # QueueZero Launches Proxy, an AI Meeting Assistant That Turns Conversation into Tasks QueueZero on Wednesday launched Proxy, an AI assistant that can attend meetings, produce notes and assign follow-up work to software agents. The company said Proxy enters public beta next week and costs $30 per user per month. The launch demonstration showed both the product's promise and the risk of software acting on what it hears. During the session, 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. The first task did not survive review. ## How the safeguards work The exchange illustrates a basic challenge for meeting assistants. People often make offhand remarks, suggestions and jokes that sound like directives. A system that converts speech into work items has to decide which statements are commitments and which are just conversation. QueueZero said Proxy is designed with limits on what it can do without human approval. The company said the assistant requires confirmation before sending any external message. Internal tasks, such as those assigned to Wu during the demonstration, can be created automatically. Administrators can also disable automatic task creation. That setting would let organizations require review before Proxy adds items to a colleague's list, a choice that may matter in meetings where speakers are less forgiving than a company's own chief executive. ## Product details Proxy's core functions are attending meetings, generating notes and delegating follow-up work to software agents. QueueZero did not say during the demonstration how often the assistant misreads remarks, or what criteria it uses to separate requests from casual commentary. The company set the price at $30 per user per month. The public beta begins next week. ## What to watch The confirmation requirement for external messages addresses the highest-risk scenario, in which an assistant sends something to a customer or partner on the basis of a misunderstood comment. Internal task creation carries lower stakes, but it can still add noise or confusion to a team's workload if the assistant is too eager. For organizations considering the beta, the administrator control over automatic task creation may be the most important setting to evaluate. Teams can choose between convenience and oversight, depending on how much they trust the assistant to interpret their conversations. In the demonstration, the system produced one useful task, one unnecessary task and a correct human decision about which was which. QueueZero's pitch is that Proxy can reduce the manual work that follows meetings. The beta will test how well it makes those distinctions outside a controlled demonstration.

MiniMax: MiniMax M3: # $2.8 Million in Duplicate Payments: A KiteBridge Update Goes Wrong A routine software update at KiteBridge, an expense-management software company, spiraled into a multi-hour incident on Tuesday that produced roughly $2.8 million in duplicate reimbursements across 91 corporate customers. The episode has put a spotlight on the company's much-touted "exactly-once payment protection" feature and raised questions about how the failure remained under-characterized on its public status page for hours after the company had already pulled the plug. ## What Happened KiteBridge rolled out version 4.2 of its platform at 8:45 a.m. Tuesday. Within half an hour, the first customer complaint landed — at 9:12 a.m. — and by 10:03 a.m., the company had suspended its automated payment pipeline entirely. The root cause, according to the company, was a payment-processor timeout. When the external processor failed to confirm a transaction in the expected window, KiteBridge's fallback system was supposed to safely retry. Instead, the fallback treated each retry as a brand-new reimbursement request rather than a repeat of an earlier one. The result: 13,400 duplicate reimbursements totaling $2.8 million, all flowing out to employees of the affected corporate customers. Crucially, the money did not come from employee accounts. Employers funded the duplicate payments, and no workers saw money withdrawn from their personal balances. By Wednesday evening, KiteBridge had recovered $2.3 million of the overpayments, with the remainder still pending. ## A Marketing Promise Meets Reality The incident is especially uncomfortable for KiteBridge because of what the company itself said in April. In the changelog for version 4.0, KiteBridge introduced what it called "exactly-once payment protection" — a safeguard meant to guarantee that each reimbursement could only be processed one time. Tuesday's failure suggests that protection did not survive contact with the real-world timing problems of payment processors. The fallback logic that treated retries as fresh requests is, in essence, the opposite of exactly-once behavior. In distributed systems, achieving exactly-once delivery is notoriously difficult, and engineers typically rely on unique transaction IDs and idempotency keys — small pieces of identifying data attached to each request — to prevent double-processing. KiteBridge has not yet explained why those safeguards failed. ## A Delayed Disclosure The company's public incident page also drew criticism. When KiteBridge first posted about the issue, it described the problem as delayed reimbursements — language that, for several hours, understated what was actually happening. The duplicate-payment figure was not disclosed on the page until roughly six hours after automated payments had been suspended. By that point, several corporate customers had already begun sorting through their own records and realizing the scale of the overpayment. ## What's Next CEO Lena Ortiz has apologized and pledged to publish a full technical report. KiteBridge is also continuing to claw back the remaining roughly $500,000 from the affected employers. For a company that markets its software as a safeguard against exactly this kind of incident, the path back to customer trust is likely to run through that report — and through the next update that promises to make it impossible again.

Anthropic: Claude Sonnet 5.5: # Software Glitch Sends 13,400 Duplicate Reimbursements Worth $2.8 Million **Expense-management firm KiteBridge suspended automated payments Tuesday after a failure hit 91 corporate customers. Most of the money has been recovered, but questions remain about the safeguards and the company's initial public account.** A software failure at KiteBridge, which makes expense-management software, generated 13,400 duplicate reimbursements totaling $2.8 million on Tuesday. The payments affected 91 corporate customers. By Wednesday evening, $2.3 million had been recovered, and the remaining $500,000 was still pending. Employees did not lose money. No one had funds withdrawn from their accounts, because employers funded the duplicate payments. The financial exposure fell on the companies that use KiteBridge. ## What happened KiteBridge deployed version 4.2 of its software at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m. The problem involved a payment-processor timeout, which occurs when the company that moves the money does not respond quickly enough. When that happened, KiteBridge's fallback system, a backup meant to keep payments flowing during disruptions, misread the situation. Instead of recognizing retries as repeat attempts at the same payment, it treated them as new reimbursement requests. Each retry became a fresh payment. KiteBridge suspended automated payments at 10:03 a.m., 78 minutes after the deployment. By then, the duplicates had reached thousands of reimbursements. It is not clear from the available reporting whether version 4.2 caused the timeout or only exposed the fallback flaw. ## A safeguard that was supposed to prevent this The incident is notable partly because KiteBridge had publicly claimed protection against this kind of error. An April changelog said version 4.0 had introduced "exactly-once payment protection." The term describes a design goal in payment systems: each transaction should be carried out one time, even if a request is repeated because of a technical hiccup. Tuesday's duplicates suggest that protection did not hold in this scenario. KiteBridge has not yet explained why. ## Questions about disclosure KiteBridge's public incident page initially described the problem as delayed reimbursements. That description did not match what customers were experiencing, which was extra payments rather than late ones. The page first disclosed the duplicate payments six hours after the company suspended automation. The gap matters for customers deciding whether to halt their own payments, notify finance teams or begin seeking repayment. Incident pages are often the main source of real-time information during a software failure. ## The company's response CEO Lena Ortiz apologized and said KiteBridge would publish a technical report on the incident. The report is expected to address the details still unexplained: how the fallback system came to treat retries as new requests, why the April safeguard did not stop it, and how the incident page came to describe the problem as it did. Until then, the outstanding $500,000 remains the most immediate concern for affected customers. The longer-term concern is whether companies that rely on automated payment software can trust the protections its vendors advertise.

MiniMax: MiniMax M3: 6

Anthropic: Claude Sonnet 5.5: Going through the draft line by line: 1. `#Flood software moves upstream` violates the headline rule, since it lacks the space after `#`. **Violation.** 2. `The company is expanding its pilot.` is meant to be the dek, but it isn't italicized. **Violation.** 3. `##Details` violates the section heading rule, since it lacks the space. **Violation.** 4. `Read [the announcement](www.example.com/news).` has a link that is not an absolute HTTPS URL. **Violation.** 5. `* The first pilot begins Monday.` begins with an asterisk instead of a hyphen. **Violation.** 6. `<div>The second pilot begins Friday.</div>` is raw HTML. **Violation.** The blank lines are not violations. **6**

MiniMax: MiniMax M3: # Startup Spotlight: LedgerLark For small investment firms, the work that happens after a trade is settled often goes unnoticed by clients but can dominate a back office. LedgerLark, a startup, is betting that AI agents can take on a meaningful slice of that work. The company builds software that compares trade confirmations, custodian records and internal ledgers, then proposes corrections for a human operator to review. Crucially, the product does not execute trades or move money; it surfaces discrepancies and recommends fixes, leaving approval and action to staff. The idea traces back to a moment co-founder Mina Batra describes from her time in operations at a regional brokerage. A six-person team spent three days resolving a trade mismatch that ultimately came down to a copied account number. "It was the kind of thing that should not have taken a day, let alone three," Batra said in an interview. The experience stuck with her, and after seven years in operations, she began sketching what an automated first pass might look like. She teamed up with Owen Carr, who had built reconciliation systems at a large bank. The pairing joins an operations insider with someone who has shipped reconciliation infrastructure at scale, a combination that may help explain why the product is built around human review rather than full automation. That choice is also a sensible one for a regulated industry, where automated write-offs of failed trades would be a hard sell to compliance teams. Today, LedgerLark has five paying customers. Annual contracts range from roughly $36,000 to $110,000, depending on firm size and transaction volume. The company says one customer reduced its unresolved daily exceptions from 70 to 18 after deploying the software, though it has not shared the methodology behind that figure or how it would hold up across a longer time window. A single before-and-after number, however flattering, is not the same as a benchmark. LedgerLark has raised a $3.5 million seed round led by Motive Row. The capital is earmarked for engineering hires, security certifications and onboarding new customers. The competitive picture is crowded. Established reconciliation vendors, including legacy platforms and newer SaaS offerings, already serve this market, and many small investment firms rely on internal tools that have been refined over years. LedgerLark's pitch centers on the flexibility of AI agents, which the company argues can handle messy, non-standard data better than rules-based systems. Whether that advantage is durable, or simply a feature that incumbents will eventually copy, is an open question. There are also practical headwinds. Financial institutions typically require lengthy security and procurement reviews, and sales cycles can stretch many months. For a small startup, that timing can be punishing, and a handful of stalled enterprise deals can reshape a year. Hiring engineers who understand both the regulatory landscape and the awkward realities of legacy back-office data is its own slow process. What seems clearer is the underlying problem. Post-trade operations at small firms are often understaffed, and the people doing the work are frequently reconciling by hand across spreadsheets and PDFs. The opportunity for tooling is real, even if the winners are not yet obvious. LedgerLark is at an early stage, with traction that is modest but directional. The product addresses a known pain point, the founders have relevant experience, and the company has paying customers. It also faces entrenched competitors, a slow sales environment and the need to prove that its early results are not just a flattering first case study. The next year, as the seed capital is deployed and more customers go live, will say a lot about whether AI agents can move from interesting demo to durable infrastructure for the back office.

Anthropic: Claude Sonnet 5.5: # Startup Spotlight: LedgerLark Bets That Small Firms Need Their Own Back-Office Help *A young company is applying AI agents to the unglamorous work of fixing trade mismatches.* ## A three-day problem When Mina Batra worked in operations at a regional brokerage, she watched a six-person team spend three days resolving a single trade mismatch. The cause was a copied account number. The error itself was trivial, but finding it meant comparing records held in different places and checking them by hand. Batra spent seven years in brokerage operations, and that episode became the starting point for LedgerLark. She co-founded the company with Owen Carr, who built reconciliation systems at a large bank. Batra brings the view from the operations desk, and Carr brings experience with how such systems are built at scale. ## What the product does LedgerLark builds AI agents for post-trade operations at small investment firms. The agents compare three sources: trade confirmations, custodian records and a firm's internal ledgers. When they find a discrepancy, they propose a correction. A person approves it before anything changes. The boundaries are deliberate. The product does not execute trades or move money. That limits the damage a wrong suggestion can do, since the worst case is a bad proposal that a reviewer rejects. It also leaves open how much time reviewers will save, because a stream of poor proposals could cost as much attention as it saves. ## Early traction Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. By simple arithmetic, that puts annual contract value somewhere between $180,000 and $550,000. It is early revenue, though not yet a large base. The company says one customer cut its unresolved daily exceptions from 70 to 18. That is a notable change if it holds, but it is a single customer's result, reported by the company. Questions remain about how typical the customer is, how the exceptions were counted, and whether the improvement lasted past the first months. Results across all five customers would say more. LedgerLark has also raised a $3.5 million seed round led by Motive Row. ## The competition The field is not empty. Established reconciliation vendors already sell software for matching records across systems, and some large institutions rely on internal bank software built for their own workflows. LedgerLark will have to show what its agents add beyond what these tools already do, and whether smaller firms prefer a newer product to a familiar one. The focus on small firms may be the clearest point of difference. Smaller operations teams have fewer people to absorb a bad week of exceptions, and they may be less well served by tools designed for larger institutions. That is a plausible reading of the market, not a settled fact, and the company's growth will test it. ## The slow part Financial institutions may require lengthy security and procurement reviews. For a company handling confirmations, custodian data and ledgers, that scrutiny is reasonable. Review cycles can stretch sales timelines, and they can strain a small team that must answer detailed questions while still improving the product. The human-approval design may help here, since the software proposes changes without carrying them out. Reviewers may still want to know how the agents reach their conclusions and what data they can see. ## What to watch Several things will show whether LedgerLark's approach holds up: - **Consistency:** whether the exception reductions appear across customers, not only one. - **Reviewer burden:** how often staff accept proposed corrections without rework. - **Renewals:** whether the first annual contracts are renewed, and at what size. - **Sales speed:** how long security and procurement reviews take as the company approaches new firms. - **Differentiation:** whether customers choose LedgerLark over incumbent tools, or use it alongside them. The founding story is a modest one: a slow, costly error in a process that rarely gets attention. If the company can show that its agents reliably shorten the path from mismatch to fix, it will have a useful product. Its first renewals should show whether it does.

MiniMax: MiniMax M3: # Foundry Hall to Close After Seven Years, Citing Loss of Key Sponsor **Foundry Hall, a nonprofit startup accelerator that has helped launch more than two hundred companies in Milwaukee, will close on October 31 after Blank River Bank declined to renew its sponsorship.** The bank had been the organization's largest corporate backer, supplying 38% of its annual budget. Foundry Hall, which opened in 2018, has enough money to operate through the end of October but not through another full program cycle. The closure ends a run that supporters say reshaped the city's early-stage business scene. Over its lifetime, Foundry Hall supported 212 startups that together employ 1,140 people across the region, offering subsidized rent, mentorship, and access to investors through cohort-based programs. Six of its own employees will lose their jobs, and twelve startups currently working out of its shared workspace must find new offices. Executive director Camille Ross said the nonprofit watched its runway shrink in real time. "We kept waiting for the replacement check that never came," she said, describing months of uncertainty as the bank signaled, then withdrew, then signaled again. "You make plans for the year, you commit to cohorts, and then the ground shifts." The bank's decision was framed as strategic. A spokesperson said the move reflected a shift in the bank's community investment priorities and was unrelated to Foundry Hall's performance or outcomes. Bank officials did not respond to follow-up questions about where the redirected funds would go. For the entrepreneurs who passed through Foundry Hall's doors, the loss feels more personal. Dev Malik, who founded his software company after graduating from one of its first cohorts, said the building gave him something the city's startup scene had previously lacked. "My first investor meeting happened in that building," he recalled. "Before Foundry Hall, I was pitching from the back table of a coffee shop." For the twelve founders still working out of the space, the timeline is even tighter. Some have only weeks to secure leases, hire movers, and keep their teams intact while still meeting payroll. City officials said they are trying to soften the immediate blow. Deputy mayor Susan Cole said the city is in active conversations with building owners and local universities about temporary workspace for the companies still operating out of Foundry Hall, though no permanent replacement has been identified. The closure leaves a gap in a small-business support network that has grown more reliant on a shrinking pool of corporate sponsors. With cohorts ending and shared desks emptying out, Ross said the remaining weeks will be spent on transition rather than expansion, helping the twelve companies in residence find their next home. "We built something that worked," she said. "I'd like to think the model survives, even if the building doesn't."

Anthropic: Claude Sonnet 5.5: # Foundry Hall, a Launchpad for Milwaukee Startups, Will Close Oct. 31 Foundry Hall opened in 2018 with a simple purpose: give young companies a place to start. On Oct. 31, the Milwaukee nonprofit accelerator will close its doors. In that time, Foundry Hall has supported 212 startups. Together, they employ 1,140 people. The closure follows the loss of its largest corporate sponsor. Blank River Bank, which supplied 38% of Foundry Hall's annual budget, declined to renew its sponsorship. The organization has enough money to operate through October, but not enough to carry another full program. Executive director Camille Ross said the staff spent months hoping another funder would step in. "We kept waiting for the replacement check that never came." Six employees will lose their jobs. Twelve startups that currently work out of the space will have to leave. A Blank River Bank spokesperson said the decision reflects a shift in the bank's community investment priorities and is unrelated to Foundry Hall's performance. The bank did not point to any shortcoming in the program, and the statement offered little comfort to those who built their work around it. For many founders, the loss is personal as well as practical. Dev Malik said Foundry Hall was where his company first became real to other people. "My first investor meeting happened in that building," Malik said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the companies that must move out. For those twelve startups, the question is no longer only where to find desks, but how to keep the momentum of a community that is being dismantled around them. The numbers describe what Foundry Hall built: hundreds of companies, more than a thousand jobs, and a corner of Milwaukee where an idea could be tested before it had customers or capital. They say less about the daily texture of the place, the introductions made in hallways and the advice passed from one founder to the next, which does not move easily from one address to another. Ross has spent the past several weeks preparing for an ending she did not choose. The work now is practical: closing out the program, supporting the remaining startups, and helping staff look ahead. Foundry Hall will remain open through the end of October. After that, the companies it helped start will have to carry on without it.

MiniMax: MiniMax M3: # The Cheaper Model Tied the Pricier One. The Real Story Is What Comes Next. A benchmark of five language models on 200 real customer-support tickets has produced a finding any operations team will recognize: the most expensive system was not meaningfully better than one costing less than half as much. Model **North** posted the top average score of 84.2 out of 100, narrowly ahead of **Vale** at 83.8. Researchers say the 0.4-point gap is not statistically significant — well within the noise of the evaluation. Yet the two sit at opposite ends of the cost curve. North runs an estimated $1.90 per 1,000 tickets; Vale runs $0.74. The "winner" costs roughly two-and-a-half times as much as Vale, with no measurable quality gain on this test. That is the central finding, and it changes how the rest of the leaderboard should be read. ## How the benchmark worked Researchers collected 200 anonymized tickets from a single English-language software company. Each was answered once by each model, producing a thousand graded responses. Three experienced support managers scored every answer blindly, combining factual correctness, usefulness, and tone into a single composite score out of 100. No grader knew which model produced which answer, and no model had access to the company's internal knowledge base — only the ticket text itself. The blind, multi-grader setup is the study's main strength. It also bounds what the results can tell us. ## Reading the leaderboard If Vale matches North on quality, the question is what the remaining three models are worth. **Ember** (79.1, $0.31) sits about five points behind the leaders. Whether that gap matters depends on the cost of being wrong. For many workflows, the answer is "not much." For others, it is the difference between a resolved ticket and a churned account. **Kite** (76.4, $0.22) and **Harbor** (71.0, $0.08) sit further back. Harbor, the cheapest model, scored roughly 13 points below North — a gap that is almost certainly meaningful. Choosing it is a deliberate trade of quality for budget, not a free lunch. One operational wrinkle: Vale produced answers 18% faster than North. On a high-volume support queue, that latency advantage compounds, and stacks on top of the cost edge. ## Where each model actually wins Aggregate scores flatten a more interesting picture underneath. North performed best on billing disputes, where precise language and references to specific account states seem to matter. Vale performed best on technical troubleshooting, where step-by-step reasoning and a more conversational tone may have helped. A team picking a single model for a single product would do well to look at which category dominates their ticket mix — and possibly route between them. ## What every model got wrong The sharpest signal in the data is not at the top of the leaderboard but at the bottom. All five models performed poorly when a ticket depended on an image attachment. Customers often attach screenshots of error messages, receipts, or broken UI; when the answer hinged on what was in the image, none of the five systems could reliably see it. For any team considering automation, the share of tickets relying on attached images is effectively a hard ceiling on what any model in this study can handle without human help. ## What the study does not tell us The test covered only one company, in one language, with one type of product. Generalizing the absolute scores — Vale is not "83.8 good" in the abstract — would be a mistake. The relative ranking, and especially the cost-versus-quality trade-off, is more robust, but still a single data point. The benchmark also did not evaluate **agentic** behavior: the models could not look up account records, open tickets, issue refunds, or call any external tool. Nor did it test multilingual support, voice transcripts, or anything outside text-in, text-out. A model scoring 84 here could behave very differently once given real tools and real data. ## The practical takeaway For a text-only, English-language support workload similar to the one tested, the most defensible default is Vale. It matches the top model on quality, costs 61% less per thousand tickets, and answers faster. North remains worth considering if billing disputes dominate the queue, or if vendor relationships or specific safety properties justify a premium. Ember is a reasonable cost-down option for lower-stakes categories. Kite and Harbor are real choices only when budget pressure is acute and quality tolerance is real. The headline is not that one model won. It is that the model that "won" did not win by enough to justify what it costs.

Anthropic: Claude Sonnet 5.5: # The Best Support AI Wasn't the Most Expensive One **In a blind test of five language models on customer-support tickets, the two top scorers were statistically indistinguishable. One of them cost about 61% less.** ## The headline finding Researchers asked five language models to draft replies to 200 anonymized customer-support tickets. Model North posted the highest average score, 84.2. Model Vale followed at 83.8. The researchers report that this gap was not statistically significant, so the data gives no good reason to say North wrote better answers than Vale overall. The prices were far apart. North's estimated cost was $1.90 per 1,000 tickets, while Vale's was $0.74. For a team handling a million tickets, that is roughly $1,900 versus $740 in model costs. Vale also answered 18% faster than North. The practical takeaway is that, on this test, paying the highest price did not buy a measurably better result. If quality is the goal, Vale delivered top-tier scores at well under half of North's cost, and faster. ## How the test worked Each model wrote one proposed answer for every ticket. Three experienced support managers then graded the answers without knowing which model had written which. Scores combined three things: factual correctness, usefulness and tone. Blind grading by experienced practitioners is a strength. Graders who don't know the source can't favor a familiar brand, and support managers know what a good reply looks like in practice. ## The rest of the field The remaining models scored lower as they got cheaper. Ember scored 79.1 at $0.31 per 1,000 tickets, Kite 76.4 at $0.22, and Harbor 71.0 at $0.08. The trade-off is real, but its value depends on the job. A model scoring several points lower may be fine for drafts that a human reviews, or for simple queries. It may be a poor choice for sensitive disputes where one wrong answer is costly. The summary reports a significance test only for North versus Vale. The gaps between the other models are larger, but readers shouldn't assume each one is meaningful without the full statistics. ## Different strengths by ticket type The models were not interchangeable within the top pair. North did best on billing disputes, and Vale did best on technical troubleshooting. That could matter to a company whose tickets skew heavily one way, but it should be treated as a lead rather than a rule. Splitting 200 tickets into categories leaves small samples, and the results don't say whether these category-level differences were statistically significant. ## The shared weakness: images The most consistent result across all five models was a failure. Every model performed poorly when a ticket depended on an image attachment, such as a screenshot of an error message or a photo of a billing statement. Price and overall rank made no difference here. Customers often send screenshots, so a deployment that ignores this gap will hit it regularly. Tickets with attachments probably need a human, or a separate process designed for them. ## What the study can't tell you The limits are significant: - **One company, one language.** The tickets were in English and came from a single software company. Results may not carry over to other industries, products or customer bases. - **No multilingual testing.** Performance in other languages is unknown. - **Drafting only.** The researchers did not test agentic actions or tool use, such as issuing a refund, looking up an account or changing a setting. Writing a good reply is different from reliably taking the right action. - **Estimated costs.** The prices are estimates. Real spending will depend on ticket length, prompt design, volume discounts and how often answers need retries or review. - **Unreported details.** The summary doesn't say how closely the three graders agreed, or how the three scoring components were weighted. Both would help in judging how solid the rankings are. Speed also deserves a caveat. Vale's 18% advantage is a relative figure, and the summary doesn't give absolute response times. Whether that difference matters depends on whether replies are customer-facing in real time or queued for agent review. ## What to do with it For a team choosing a model for English-language, text-only support drafting, the results support a shortlist rather than a verdict. North and Vale are effectively tied on quality, and Vale's lower cost and faster responses make it the obvious one to test first, with North a reasonable alternative for billing-heavy workloads. Cheaper models may suit lower-stakes or heavily reviewed work. Before committing, run the same kind of blind comparison on your own tickets. A model's average score matters less than how it handles your customers' questions, including the ones that arrive with a screenshot.

MiniMax: MiniMax M3: C

Anthropic: Claude Sonnet 5.5: C

MiniMax: MiniMax M3: # HelioDock Touts Remote Robot Monitoring After 14-Month Beta With Nine Logistics Operators *Warehouse robotics company closes $24 million Series B as it moves Night Shift software to general availability.* HelioDock, a maker of autonomous warehouse robots, introduced Night Shift on Thursday, a software platform that allows warehouse operators to monitor and, when needed, take control of robot fleets from outside the facility. The release follows a 14-month private beta with nine logistics companies and roughly 310 robots, and comes alongside a $24 million Series B that brings the company's total funding to $39 million. The product targets a specific pain point in 24-hour operations. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," said HelioDock CEO Mara Velez. Night Shift is designed to give operators a live view of each robot, surface alerts when units deviate from expected behavior, and let qualified staff intervene remotely rather than dispatching someone to the site. HelioDock says the beta experience supports the company's pitch. According to the company, participating customers saw 17% fewer unplanned robot stoppages during the trial — a figure that has not been independently verified. HelioDock did not share the methodology behind the number, and the sample of nine operators and 310 robots is small enough that broader results could vary. The new funding round was led by North Coast Ventures. HelioDock did not disclose the round's other participants or a post-money valuation. The company said the capital will support hiring and continued product development. Night Shift enters a market that already includes fleet-management tools from Relay Robotics and YardOS. HelioDock's pricing — $90 per robot per month, with general availability set for September 15 — places it roughly in line with comparable enterprise software, though direct comparisons depend on features each vendor bundles. ## Connectivity as the Constraint The most candid feedback came from one of the beta participants. Aaron Pike, an operations lead at Fulcrum Logistics, said the system performed well in facilities with strong, consistent network coverage but became less useful during LTE outages at rural sites. "When the network drops, the dashboard goes dark with it," Pike said, echoing a limitation HelioDock acknowledged in its own materials. The reason is architectural. When connectivity fails, Night Shift cannot reach the robots, and the machines revert to their locally configured behavior. Remote control, by design, is unavailable in those windows. The result is a product whose value is tightly bound to the quality of the network underneath it — a meaningful caveat for warehouses in areas with patchy cellular service or aging infrastructure. HelioDock says customers can mitigate the gap with redundant connections or by deploying local edge hardware, but those add cost and complexity that the base subscription does not include. The company has not said whether offline dashboards or cached telemetry are on the roadmap. ## What Ships in September The general-availability release will include the alerting, live telemetry, and remote intervention features tested during the beta, according to HelioDock. New additions include multi-site fleet views, role-based access for outside contractors, and an API for integrating Night Shift alerts with existing warehouse management systems. The company is also extending its sales motion beyond direct sales, opening a channel program for systems integrators in North America and parts of Europe. Pricing for the integrator tier was not disclosed. For HelioDock, the launch marks a transition from hardware sales to recurring software revenue — a familiar move for robotics companies seeking smoother growth profiles. Whether Night Shift delivers on that promise will depend as much on cellular coverage maps as on the software itself.

Anthropic: Claude Sonnet 5.5: # HelioDock Launches Remote Monitoring Software for Warehouse Robots and Raises $24 Million *The company's Night Shift product aims to keep robot fleets running without on-site staff, but a beta customer says it falters when connectivity does.* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. The company also announced a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. Night Shift is meant to reduce the number of times a robot stalls and someone has to go to the warehouse to fix it. HelioDock says customers in a private beta had 17% fewer unplanned robot stoppages, a figure that has not been independently verified. The software depends on a network connection, and one beta customer said it was far less useful when that connection dropped. ## Beta results The beta ran for 14 months and involved nine logistics companies and 310 robots. HelioDock attributes the 17% reduction in unplanned stoppages to the software. The company did not provide further detail on how it measured the figure. CEO Mara Velez framed the product around a common overnight problem. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," she said. ## Where connectivity matters Aaron Pike of Fulcrum Logistics, a beta customer, said Night Shift worked well in facilities with strong connectivity. It became less useful during rural LTE outages, he said. HelioDock acknowledges the constraint in how the system behaves when a connection is lost. In that case, robots return to their locally configured behavior, and remote control becomes unavailable. The robots do not stop working, but operators lose the ability to step in from afar until the connection returns. For operators with sites in areas where cellular service is unreliable, that distinction may determine how much of Night Shift's value they can use. The company's notes on the beta did not indicate how many of the nine participating companies operated in rural locations. ## Pricing and competition Night Shift will cost $90 per robot per month and will become generally available September 15. HelioDock is entering a market that already has alternatives. Relay Robotics and YardOS both offer competing fleet-management products. ## Funding The $24 million Series B, led by North Coast Ventures, brings HelioDock's total raised to $39 million. The company did not say in the materials provided how it plans to divide the new money among software development, hardware production or other areas. Night Shift is the company's software layer for the robots it already sells. Its release gives HelioDock a way to charge on a recurring, per-robot basis, as opposed to relying solely on the robots themselves. *Correction note for editors: the reporting did not establish whether Night Shift will be available to operators who use robots from other manufacturers.*