Head to head: MoonshotAI: Kimi K3.1 Pro vs Anthropic: Claude Opus 5.5

MoonshotAI: Kimi K3.1 Pro: # QueueZero’s Proxy AI Turns a Launch-Demo Aside Into Tasks for Its CEO QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant designed to attend meetings, produce notes and assign follow-up work to software agents. The product will enter public beta next week and cost $30 per user per month. The launch demonstration also illustrated the difference between recognizing a request and understanding the room. A customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the remark as an instruction and 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 assistant’s literal approach to task management was, briefly, more efficient than diplomatic. QueueZero said Proxy requires confirmation before sending an external message. Internal tasks, however, can be created automatically. Administrators can disable automatic task creation, giving organizations a way to limit that behavior. That distinction is likely to matter for businesses assessing how much autonomy to give meeting software. An assistant that records decisions and distributes work could reduce manual follow-up, but automatic task creation also means an offhand comment may become an item in a company’s workflow without a separate approval step. In the demonstration, the tasks stayed internal and Wu was able to remove one. QueueZero’s confirmation requirement applies to external messages, not to the creation of those internal assignments. Proxy’s agent assignments extend its role beyond transcription. Rather than only producing a summary after a call, the assistant is intended to turn meeting output into work that software agents can carry out. The launch episode showed both sides of that approach: Proxy quickly extracted two possible actions from conversational speech, but it did not treat the audience member’s wording as merely a joke or a request for a person to relay. Administrators considering the beta will therefore need to weigh the convenience of automatic follow-up against the oversight they want around internal systems. The option to turn off automatic task creation provides a direct control, while confirmation for outbound messages adds a separate checkpoint before communication leaves the organization. Proxy’s public beta next week will offer a broader test of whether those controls are sufficient for everyday meetings. For QueueZero, the challenge will be to show that proactive follow-up can work without requiring users to watch every casual remark for a potential assignment.

Anthropic: Claude Opus 5.5: # QueueZero's AI Meeting Assistant Turns Audience Remark Into Tasks for Its CEO QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to attend meetings, take notes and hand off follow-up work to software agents. The product's debut also offered an unscripted look at how the system decides what counts as an instruction. 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 comment as a directive. It generated two tasks assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. Of the two, the pricing sheet was the easier deliverable. ## What Proxy does Proxy is built to join meetings, produce notes and assign follow-up work to software agents, rather than leaving action items for people to transcribe and distribute by hand. The episode at the launch highlights a central question for tools of this kind: how a system distinguishes a deliberate request from an offhand remark. In this case, Proxy acted on a comment from an audience member, split it into two separate actions and assigned both to a named person, all without anyone in the room asking it to. ## Safeguards and settings QueueZero says Proxy requires confirmation before it sends an external message. Internal tasks, however, can be created automatically, as they were during the demonstration. Administrators can disable automatic task creation, according to the company. That distinction matters for organizations weighing how much independence to give AI agents inside their systems. An internal task may be easy to remove, as Wu showed. But automatic task creation also means a casual comment in a meeting can produce assignments for colleagues before anyone has reviewed them. QueueZero's design places the confirmation step at the point where Proxy's output would leave an organization. Inside the organization, the assistant can act faster and without review unless an administrator chooses to switch that capability off. ## Pricing and availability Proxy costs $30 per user per month. It enters public beta next week. The beta will give QueueZero and its customers a broader test of how reliably Proxy separates genuine requests from conversational asides, and whether companies decide to keep automatic task creation turned on once the assistant is working in their own meetings rather than on a launch stage.

MoonshotAI: Kimi K3.1 Pro: C

Anthropic: Claude Opus 5.5: C

MoonshotAI: Kimi K3.1 Pro: # Startup Spotlight: LedgerLark Takes Aim at Post-Trade Exceptions For small investment firms, a trade can be complete long before the administrative work around it is settled. Confirmations, custodian records and internal ledgers must agree; when they do not, operations staff have to find the discrepancy and decide how to correct it. LedgerLark is building AI agents for that work, with a deliberate limit: the software can propose a fix, but people remain responsible for approving it. The idea grew out of co-founder Mina Batra’s seven years in operations at a regional brokerage. She began exploring the idea after watching a six-person team spend three days resolving a trade mismatch caused by a copied account number. The episode suggested a useful distinction. Much of the effort went into gathering and comparing records, while the consequential judgment—what should change, and whether it was safe to change it—still belonged with the team. Batra’s co-founder, Owen Carr, previously built reconciliation systems at a large bank. Together, they are focusing on small investment firms rather than trying to automate every part of post-trade operations. LedgerLark’s agents compare confirmations, custodian records and internal ledgers, then propose corrections for human approval. The product does not execute trades or move money. That boundary matters. It makes the product easier to describe as an assistant to an existing control process, rather than a replacement for it. It also leaves important questions for customers: How clearly does an agent explain its proposed correction? What happens when source records conflict or are incomplete? And can staff review suggestions quickly without simply trusting them? The reporting available so far does not answer those questions. There is early commercial traction. Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. That is a notable operational change if it holds across different trading volumes and market conditions, but it is a company-reported result from one customer, not evidence that every firm will see the same improvement. The spread in contract values also leaves open how much of LedgerLark’s revenue comes from its largest accounts. The company has raised a $3.5 million seed round led by Motive Row. That gives it resources to develop the product and pursue more customers, but the next tests are likely to be practical rather than theatrical: implementation effort, accuracy, auditability and retention. LedgerLark also faces competition from established reconciliation vendors and internal bank software. Some prospective buyers may already have tools or workflows that can be extended instead of replaced. Financial institutions may require lengthy security and procurement reviews, which could slow sales even when an operations team sees value. The founders’ insight is therefore narrower—and more credible—than a promise to automate the back office. They are betting that better comparison and well-presented suggestions can reduce repetitive exception work without removing human control. Five paying firms show that the bet has found some buyers. Whether LedgerLark can turn those early contracts into a dependable product across many firms will depend on evidence that is not yet public: consistent results, manageable reviews and customers choosing to stay.

Anthropic: Claude Opus 5.5: # Startup Spotlight: LedgerLark Wants to Shrink the Back-Office Pile After a trade is executed, the work isn't finished. Someone has to confirm that the trade confirmation, the custodian's records and the firm's own ledger all agree. When they don't, the result is an "exception," a discrepancy that someone must chase down by hand. LedgerLark, a young company building AI agents for this post-trade work, is betting that small investment firms will pay to make that pile smaller. ## Where the idea came from Co-founder Mina Batra spent seven years in operations at a regional brokerage. The moment that pushed her toward the idea was not dramatic. She watched a six-person team spend three days untangling a trade mismatch whose cause turned out to be a copied account number. The smallness of the story is part of its point. Post-trade errors are often mundane: a typo, a stale field, a record entered in one system but not another. Yet finding them can require people to cross-check several sources line by line. In that case, most of the effort went into locating the problem rather than fixing it. Her co-founder, Owen Carr, comes at the same problem from the systems side. He previously built reconciliation systems at a large bank. Together, the pair combine front-line operations experience with familiarity with how larger institutions automate matching. ## What the product does LedgerLark's agents compare three sets of records: trade confirmations, custodian records and internal ledgers. When they find a discrepancy, they propose a correction, and a person decides whether to accept it. That boundary matters. The product does not execute trades or move money. In a field where errors carry regulatory and financial consequences, a human approval step is a safeguard and, plausibly, reassuring to cautious compliance teams. It also means the product's value depends on how often its proposals are right and how much time reviewing them actually saves. Those questions become easier to answer with more customers and more time. ## Early signs 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 is a notable figure, but it comes from LedgerLark itself and reflects a single customer. It isn't yet clear how typical that result is, how long it took, or whether the remaining exceptions are the harder cases. LedgerLark has raised a $3.5 million seed round led by Motive Row. ## The harder questions Reconciliation is not an empty market. Established vendors already sell reconciliation software, and large banks often build internal tools, the kind of systems Carr himself worked on. LedgerLark's opportunity appears to rest on small firms being underserved: too small to build in-house, and possibly poorly matched to products designed for bigger institutions. Two questions remain open: whether that gap is wide enough to sustain a business, and how quickly incumbents might add similar AI features. Sales cycles are another challenge. Financial institutions may require lengthy security and procurement reviews before giving a vendor access to sensitive records. For a startup with five customers, the pace of those reviews could shape growth as much as the product does. Then there is the question facing any AI tool in a regulated setting: how firms will audit the agents' suggestions, and how much trust operations teams will extend to them over time. ## Why it's worth watching LedgerLark is tackling an unglamorous, persistent problem with a deliberately limited product. If its early results hold up across more customers, small firms could see a real reduction in tedious work. If they don't, the company will still offer an instructive test of where AI agents fit in financial operations.

MoonshotAI: Kimi K3.1 Pro: # HelioDock launches Night Shift for remote robot oversight and raises $24 million *The fleet-monitoring software will cost $90 per robot per month and become generally available on September 15, following a 14-month private beta.* HelioDock on Thursday launched Night Shift, software that lets operators monitor fleets of autonomous warehouse robots remotely, and announced a $24 million Series B funding round led by North Coast Ventures. For HelioDock, which makes autonomous warehouse robots, Night Shift adds a software product focused on overseeing machines after they are deployed. The launch also places the company alongside Relay Robotics and YardOS, which already offer competing fleet-management products. HelioDock’s case for Night Shift rests on letting managers monitor a fleet without being at the warehouse, while the beta showed that the value of that arrangement can vary with a site’s network connection. Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock said those beta customers experienced 17% fewer unplanned robot stoppages. The company supplied the figure, and it has not been independently verified. It is therefore a company-reported result from the test group rather than an independent benchmark for the software’s performance. CEO Mara Velez framed the product around the disruptions that can pull warehouse staff away from home or other work. “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. Her example describes the role HelioDock wants Night Shift to play: giving operators a way to monitor robots from outside the building instead of requiring every issue to be assessed by someone on site. The product is built around remote access, but the robots have locally configured behavior to fall back on when that access is interrupted. If connectivity fails, they return to that behavior and remote control becomes unavailable. The arrangement separates what the machines do on site from what an operator can do from elsewhere: local behavior remains the fallback, while off-site control depends on the connection. Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His account makes network conditions central to assessing the product. A well-connected facility can support the remote oversight HelioDock is promoting, while an outage curtails the remote capabilities and leaves the robots operating according to their local configuration. That limitation does not erase HelioDock’s reported beta results, but it does qualify how they should be understood. The 17% figure concerns unplanned stoppages across the company’s beta customers; Pike’s comments describe how the system’s usefulness differed with connectivity at particular facilities. One is a company-supplied performance measure, and the other is a customer’s account of the conditions under which the software worked well or became less useful. The Series B brings HelioDock’s total funding to $39 million. The company announced the financing alongside the product launch as Night Shift moved from private testing toward a dated commercial release. The funding and launch also arrive with fleet-management products already available from Relay Robotics and YardOS, giving prospective customers other options in the same category. Night Shift will become generally available on September 15. HelioDock will charge $90 per robot per month, tying the recurring cost of the software to the number of robots covered rather than setting a single price for an entire facility.

Anthropic: Claude Opus 5.5: # HelioDock Launches Remote Fleet Software for Warehouse Robots, Raises $24 Million *The company says Night Shift reduced unplanned robot stoppages during a 14-month beta, but one customer found it less useful when rural networks went down* HelioDock, a maker of 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 funding round led by North Coast Ventures. The product targets a persistent cost of warehouse automation: when a robot stalls in the middle of the night, someone often has to be there to deal with it. HelioDock is betting that operators will pay to handle more of those problems from a screen instead of the warehouse floor. The company says beta customers using Night Shift experienced 17% fewer unplanned robot stoppages, a figure that has not been independently verified. A beta customer said the software's value depends heavily on the strength of a facility's internet connection. Chief Executive Mara Velez described Night Shift in terms of the overnight shift its name references. "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. Night Shift spent 14 months in a private beta with nine logistics companies running a combined 310 robots, an average of roughly 34 robots per company. HelioDock credits the software with reducing unplanned stoppages across that group, though the 17% figure comes from the company and has not been confirmed by an outside party. ## Connectivity remains the constraint Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. That limitation follows from how Night Shift handles a lost connection. When the link fails, the robots return to their locally configured behavior, and remote control becomes unavailable. A manager monitoring a fleet from home cannot intervene until connectivity is restored. The result is that the software's benefits may vary by location. Sites with reliable networks may get the full remote-monitoring experience Velez describes. Operators at facilities with spotty coverage, such as rural sites that depend on LTE, may find that the moments when they most need remote access are the moments it is unavailable. ## Pricing and competition Night Shift will cost $90 per robot per month, or $1,080 per robot per year. At that rate, a fleet the size of the entire beta group, 310 robots, would cost $27,900 a month. The software becomes generally available on Sept. 15. HelioDock is entering a market with established alternatives. Relay Robotics and YardOS both offer competing fleet-management products, so prospective customers will have other options when evaluating remote oversight tools for their operations. The Series B brings HelioDock's total funding to $39 million, meaning the company had raised about $15 million before this round.

MoonshotAI: Kimi K3.1 Pro: # Cloudnote Debugging Page Exposed Customer Prompt Excerpts for Four Hours *Logs show excerpts from 1,842 user sessions; the company says it has found no evidence of access beyond the researcher who reported the issue.* Cloudnote discovered that a debugging page containing customer prompt excerpts was accessible without authentication for four hours on Monday. Logs show excerpts from 1,842 user sessions; each excerpt could include up to 200 characters of a prompt. The company removed the page 19 minutes after a researcher notified it of the problem. Cloudnote said it has found no evidence that anyone other than the researcher accessed the page while it was exposed. The investigation is not complete, so that finding does not establish that no one else viewed or copied the material. Nor does the session count show how many individual customers were affected: user sessions and registered users are not directly comparable. Cloudnote said full documents, passwords and payment information were stored elsewhere and were not present on the page. Whether any excerpts contained sensitive or identifying information is unknown. Even so, the incident matters because prompt excerpts are still customer data, and debugging tools can become an unexpected route around normal access controls. For a company with 3.2 million registered users, the episode raises broader questions about how customer material is used and protected in internal systems.

Anthropic: Claude Opus 5.5: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 Sessions *The page was accessible without a login for four hours on Monday. The company says it has found no sign that anyone other than the researcher who reported it viewed the data, but its investigation is not finished.* A Cloudnote debugging page that displayed excerpts of customer prompts was accessible without authentication for four hours on Monday, the company said. Cloudnote removed the page 19 minutes after a security researcher notified it of the problem. The company's logs show the page contained excerpts from 1,842 user sessions, a small fraction of Cloudnote's 3.2 million registered users. Each excerpt could include up to 200 characters from a prompt, enough to capture a sentence or two of what a user had typed. Even that limited window matters. People often put personal, professional or confidential details into prompts, and few users are likely to remember exactly what they entered in a given session. Short fragments could still reveal names, plans or sensitive questions. Cloudnote said full documents, passwords and payment information are stored elsewhere and were not present on the page. It also said it has found no evidence that anyone besides the researcher accessed the page. But that finding is preliminary. Cloudnote has not completed its investigation, and key questions remain open, including how the page came to be left unprotected and whether affected users will be told their data was exposed.

MoonshotAI: Kimi K3.1 Pro: # KiteBridge Software Failure Generates $2.8 Million in Duplicate Reimbursements A payment-processing failure at KiteBridge generated 13,400 duplicate reimbursements totaling $2.8 million on Tuesday, affecting 91 corporate customers and leaving a substantial recovery effort in its wake. By Wednesday evening, $2.3 million had been recovered, but the remaining $500,000 was still pending. KiteBridge, which makes expense-management software, deployed version 4.2 at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m. The company continued automated payments until 10:03 a.m., when it suspended them. By then, retries that should have repeated existing payment attempts had instead produced thousands of new reimbursement requests. The failure began with a payment-processor timeout. A payment processor handles transactions on the software’s behalf, and a timeout means it did not respond within the expected period; it does not necessarily establish whether the underlying payment succeeded or failed. Software commonly retries in that situation to avoid losing a legitimate transaction. KiteBridge’s fallback system—the backup process used when the normal payment path runs into trouble—treated those retries as new reimbursement requests. Rather than linking each retry to the original transaction, it created another request. A mechanism intended to make payments more resilient therefore became the source of the duplicates. The incident contrasted with KiteBridge’s earlier product claims. An April changelog said version 4.0 had introduced “exactly-once payment protection.” In payment software, that term generally means a transaction should have its intended financial effect only once, even if technical problems cause the same instruction to be sent or attempted repeatedly. It is designed to prevent the kind of multiplication that can occur when retries are mistaken for fresh requests. The duplicates did not result in money being withdrawn from employees’ accounts. Employers funded the duplicate payments, leaving the affected businesses—not workers facing direct debits—at the center of the recovery effort. KiteBridge’s public incident page initially described the issue as delayed reimbursements. That characterization did not disclose that the software had also generated extra payments. The page first revealed the duplicate payments six hours after automated payments were suspended, expanding the public picture from a reimbursement delay to a multimillion-dollar payment error. By Wednesday evening, most of the money had been recovered: $2.3 million of the $2.8 million total. The remainder was still pending, leaving the recovery incomplete more than a day after the incident began. CEO Lena Ortiz apologized and said KiteBridge would publish a technical report. The promised report follows questions about both the fallback system’s handling of retries and the gap between the initial public description and the later disclosure of duplicate payments.

Anthropic: Claude Opus 5.5: # Software Glitch Sends $2.8 Million in Duplicate Expense Payments A software failure at expense-management company KiteBridge caused 13,400 duplicate reimbursements, worth a combined $2.8 million, to go out on Tuesday. The money came from 91 corporate customers whose systems paid some employees twice for the same expenses. By Wednesday evening, $2.3 million had been recovered. The remaining $500,000 was still pending. No employees had money withdrawn from their accounts as a result of the error. The duplicate payments were funded by their employers. CEO Lena Ortiz apologized for the incident and said the company would publish a technical report explaining what went wrong. ## A Routine Update, Then Complaints KiteBridge makes software that companies use to track employee expenses and reimburse workers for costs such as travel and meals. On Tuesday at 8:45 a.m., the company released version 4.2 of its software. The first customer complaint arrived 27 minutes later, at 9:12 a.m. KiteBridge suspended automated payments at 10:03 a.m., about 50 minutes after that first report and roughly an hour and 20 minutes after the update went live. ## How Retries Became New Payments The problem began with a timeout at a payment processor, the outside service that moves money between bank accounts. A timeout happens when a system sends a request and does not receive a response within an expected window. Software commonly handles this by trying again, a step known as a retry, because the original request may have failed or may simply be slow. KiteBridge also has a fallback system, a backup process meant to keep payments moving when the normal path runs into trouble. During Tuesday's incident, that fallback treated the retries as brand-new reimbursement requests instead of repeated attempts at the same payment. As a result, a single expense claim could be paid more than once. That failure appears to conflict with a promise KiteBridge made earlier this year. An April changelog, the public list of changes in each software release, said version 4.0 had introduced "exactly-once payment protection." The phrase describes safeguards designed to ensure that each payment is processed one time and only one time, even when a system has to retry a request. The company has not publicly explained why that protection did not prevent the duplicate payments. It also has not said what role, if any, version 4.2 played in the failure. ## Delayed Disclosure KiteBridge's public incident page, where it posts updates about service problems, at first described the issue as delayed reimbursements. That description reflected the visible effect of halting automated payments, but it did not mention that thousands of extra payments had already gone out. The page first disclosed the duplicate payments about six hours after the company suspended automation, around 4 p.m. Tuesday. For the affected businesses, the gap meant that customers checking KiteBridge's status page for several hours would have seen no indication that their accounts may have paid out extra money. ## Questions Remaining The promised technical report is expected to address several unresolved questions. These include why the exactly-once safeguard failed, how the fallback system came to treat retries as new requests, and why the initial public description did not mention duplicate payments. KiteBridge has not said when the report will be released or when it expects to resolve the remaining $500,000 in pending recoveries.

MoonshotAI: Kimi K3.1 Pro: # Raincheck Raises $8 Million for Flood-Risk Software *The Austin company combines drainage maps with rainfall forecasts to help local governments decide where to send maintenance crews before storms.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. Raincheck describes the approach in [its published methodology](https://example.com/raincheck-methodology). Basin Ventures led the round, and existing investor Long Mile also participated. The new funding brings Raincheck’s total raised to $10.5 million. 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.” Raincheck currently works with seven cities in Texas and Louisiana but declined to identify them because some contracts have not been publicly announced. The company’s forecasts are recommendations, and emergency managers retain control over deployments and public warnings. Raincheck says the new funding will help it hire eight engineers and expand into Florida. ## Competition Other flood-planning vendors include HydroMap and municipal engineering contractors.

Anthropic: Claude Opus 5.5: # Raincheck Raises $8 Million for Flood-Risk Software Used by Cities *The Austin company combines drainage maps with rainfall forecasts to help local governments decide where to send maintenance crews before a storm.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in a seed round led by Basin Ventures. Existing investor Long Mile also participated, bringing Raincheck's total funding to $10.5 million. The company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. CEO Ana Flores said the software is meant to help cities prioritize work before a storm: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." Raincheck has published [an overview of its forecasting methodology](https://example.com/raincheck-methodology). The company currently works with seven cities in Texas and Louisiana. It declined to identify them because some contracts have not been publicly announced. Raincheck's forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Raincheck says the new funding will help it hire eight engineers and expand into Florida. It competes with other flood-planning vendors, including HydroMap and municipal engineering contractors.

MoonshotAI: Kimi K3.1 Pro: 6

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

MoonshotAI: Kimi K3.1 Pro: # 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: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

MoonshotAI: Kimi K3.1 Pro: # Vale’s near-tie on quality is the benchmark’s clearest result A test of five language models on customer-support tickets points to a more useful conclusion than a simple leaderboard: **Vale delivered quality statistically indistinguishable from North, the top-scoring model, while responding faster and costing substantially less.** North averaged 84.2, compared with 83.8 for Vale. That 0.4-point gap was not statistically significant, so the results do not establish North as the better model. Meanwhile, Vale answered 18% faster. Its estimated cost was $0.74 per 1,000 tickets, versus $1.90 for North—about 61% lower, or roughly 39% of North’s cost. For a support organization processing tickets at scale, that combination makes Vale the practical standout. It does not make Vale universally superior, and the study cannot predict production performance with certainty. But it challenges the assumption that buying the highest-scoring option automatically buys a meaningful quality advantage. ## How the test worked Researchers assembled 200 anonymized customer-support tickets and asked each of the five models for one proposed answer to every ticket. Three experienced support managers graded the responses without knowing which model had written them. The scores combined factual correctness, usefulness and tone. That design is more informative than comparing generated answers by eye. Using the same tickets creates a common task, while blind grading reduces the chance that a model’s name or reputation shaped the assessments. Involving support managers also grounds the evaluation in the judgment of people familiar with customer service. Still, the composite score conceals as well as summarizes. Models could reach similar averages through different balances of accuracy, helpfulness and tone. One answer per ticket also says little about run-to-run consistency. The benchmark measures proposed replies, not resolved cases, and it does not show whether customers would have accepted the answers or needed follow-up. ## Cheaper models involve a real trade-off Below the two leaders, lower prices did not preserve the same quality. Average scores ranged from 79.1 down to 71.0 as estimated costs fell from $0.31 to $0.08 per 1,000 tickets. These figures describe a broad quality-versus-cost trade-off, not five equivalent products with different price tags. They also do not establish that any individual cheaper model is unsuitable for a narrower or lower-risk workload. The overall averages additionally hide differences by ticket type. North was strongest on billing disputes, while Vale led on technical troubleshooting. A company’s ticket mix could therefore matter more than the overall ranking. The sensible next step is not to declare a universal winner, but to test the leading candidates on representative workloads and examine the kinds of errors each makes. Cost deserves similar scrutiny. The per-1,000-ticket figures are estimates for this benchmark, not a complete account of implementation, review, retries or integration. Nor do faster answers automatically mean faster customer resolutions. ## The shared failure may matter most One finding applies across the entire leaderboard: all five models performed poorly when a ticket depended on an image attachment. That is a warning about workflow coverage, not a footnote. If screenshots or other images are central to diagnosing a problem, a strong overall average offers little reassurance. Such cases may require different capabilities, additional processing or human review. The scope limits are equally important. The tickets were in English and came from one software company, so the results may not transfer to other industries, languages or support styles. Researchers did not test agentic actions, tool use or multilingual support. The evaluation therefore says nothing about whether these models can safely look up an account, issue a refund, update a ticket or operate across a global queue. The clearest takeaway is consequently narrow but actionable: **on this evidence, Vale looks like the strongest value because it matched North’s measured quality within statistical uncertainty, at lower estimated cost and higher speed.** North remains competitive, particularly for billing disputes, and cheaper options may fit some budgets. But any deployment decision should follow a pilot on real ticket types, with explicit checks for image-dependent cases and human oversight where errors carry significant consequences.

Anthropic: Claude Opus 5.5: # The Top Two Support Models Tied. One Costs 61% Less. When researchers put five language models through a customer-support test, the most useful result was a tie between the two leaders. Model North posted the highest average score, but Model Vale was close enough that the difference was not statistically significant. Vale's estimated cost was 39% of North's, and it answered 18% faster. For teams choosing a model to draft support replies, that changes the question. It is no longer "which model is best?" but "is there any reason to pay more for North?" ## How the test worked The researchers collected 200 anonymized customer-support tickets and had each of the five models write one proposed answer per ticket. Three experienced support managers then graded every answer without knowing which model produced it. Each score combined three qualities: factual correctness, usefulness and tone. The blind grading by practitioners is a strength. These are the people who would judge a reply in real life, and they could not favor a model they recognized. But two details matter for interpreting the numbers: - **One answer per ticket.** Language models can respond differently to the same prompt. A single sample per ticket shows typical performance but not how consistent each model is from run to run. - **A blended score.** Combining correctness, usefulness and tone into one number makes comparison easy, but it can hide trade-offs. A warm, well-written answer with a factual error is a bigger problem in support than a curt but accurate one. The headline averages don't show how each model balanced those qualities. ## The headline: North and Vale are effectively even North averaged 84.2 and Vale 83.8. A 0.4-point gap on a test of 200 tickets is within the range that could come from which tickets happened to be included. The researchers found no statistically significant difference, so the fair reading is that the two models performed about the same overall. The other differences are large and concrete. North's estimated cost was $1.90 per 1,000 tickets, compared with Vale's $0.74. Vale was also 18% faster. These per-ticket figures are small in absolute terms, so for a modest support operation the savings may not drive the decision. At high volume, though, paying more than two and a half times as much for equivalent quality is hard to justify without a specific reason. ## Where the leaders differ There may be specific reasons. North performed best on billing disputes, and Vale performed best on technical troubleshooting. The summary results don't say how large these category gaps were or whether they were statistically significant. A team whose queue is dominated by one ticket type should treat this as a lead to test on its own tickets, not a settled finding. If billing disputes carry high stakes for your business, such as refunds, chargebacks or angry customers, North's edge there might be worth verifying. If most of your volume is troubleshooting, Vale looks even more attractive. ## The cheaper tier The remaining three models scored lower and cost less. Ember averaged 79.1, Kite 76.4 and Harbor 71.0, with Harbor costing just $0.08 per 1,000 tickets. These gaps are much wider than the one at the top. They suggest a real quality step down, though the significance results reported here cover only the North–Vale comparison. Whether a cheaper model is good enough depends on how its drafts are used. If a human agent reviews and edits every reply, a lower-scoring model might still save time. If answers go to customers with little oversight, the quality gap matters more. ## A shared blind spot: images All five models performed poorly when a ticket depended on an image attachment, such as a screenshot of an error message. No model in this test solves that problem, and choosing a more expensive one won't help. Support teams should plan to route image-dependent tickets to people, or at least flag them for closer review. ## What the test doesn't tell you The benchmark's scope is narrow, and the limits are significant: - **One company, one language.** All tickets were in English and came from a single software company. Results may differ for other industries, products or customer bases. - **Text replies only.** The researchers did not test agentic actions or tool use, such as looking up an account, issuing a refund or updating a ticket. Many production support systems depend on exactly those abilities. - **No multilingual testing.** Teams serving customers in other languages would need to run their own evaluation. ## The bottom line On this test, Vale matched North's overall quality at less than half the estimated cost, with faster responses. That makes Vale the reasonable default for English-language, text-based support drafting, unless your own testing confirms that North's apparent strength on billing disputes matters for your queue. The most practical next step is the researchers' method itself: blind grading of real tickets by experienced staff, run on your own data.

MoonshotAI: Kimi K3.1 Pro: # Foundry Hall to Close After Losing Its Largest Sponsor Foundry Hall, a Milwaukee nonprofit startup accelerator, will close on October 31, ending an effort that began in 2018 and eventually supported 212 startups. Those companies collectively employ 1,140 people, a legacy that will outlast the organization itself, even as its programs and workspace disappear. The accelerator’s finances could not absorb the loss of Blank River Bank, its largest corporate sponsor. The bank supplied 38% of Foundry Hall’s annual budget and declined to renew its sponsorship. Foundry Hall has enough money to continue operating through October, but not enough to run another full program. Six employees will lose their jobs. Twelve startups currently using the workspace will have to leave. Executive director Camille Ross described the wait for replacement funding, which ended without a solution. “We kept waiting for the replacement check that never came,” she said. Blank River Bank said its decision reflected a change in community investment priorities, not Foundry Hall’s performance. That distinction offers little practical relief to an organization built around a funding source large enough to shape its future. When the sponsorship ended, there was no comparable commitment to take its place. For founder Dev Malik, the closing is also the loss of a place tied to a decisive early step for his company. “My first investor meeting happened in that building,” he said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” His account helps explain why the closure cannot be measured only by the budget gap. Foundry Hall gave entrepreneurs a place to meet, work and be taken seriously. Its departure will scatter the current companies and remove the staff who supported them, while the broader network of founders must carry on without the institution that helped connect it. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies. Those conversations could address the most urgent question: where the 12 startups will go next. They do not, on their own, preserve Foundry Hall’s jobs or guarantee another full accelerator program. That gap between an institution’s record and its remaining options gives the closing its weight. The work has produced visible results, but those results alone could not secure another program. The organization will therefore spend its final weeks operating within a narrow window—still open, but with its end date fixed. Its record includes hundreds of startups and more than a thousand jobs. Its immediate future is smaller and more concrete: six positions ending, 12 companies preparing to move, and a Milwaukee workspace approaching its last day.

Anthropic: Claude Opus 5.5: # Foundry Hall to Close After Losing Its Largest Sponsor Foundry Hall, the Milwaukee nonprofit that has helped more than 200 startups get off the ground since 2018, will close on Oct. 31 after its largest corporate sponsor chose not to renew its support. Blank River Bank had supplied 38% of the accelerator's annual budget. Without that money, Foundry Hall can pay its bills through October but cannot fund another full program. Six employees will lose their jobs, and the twelve startups now working out of its space have until the end of the month to find somewhere else to go. Executive director Camille Ross said the organization spent the months after the bank's decision hoping new funding would fill the gap. "We kept waiting for the replacement check that never came," she said. A Blank River Bank spokesperson said the bank had shifted its community investment priorities and that the decision had nothing to do with how Foundry Hall performed. By its own measures, the accelerator performed well. The 212 startups it has supported now employ a combined 1,140 people. Some of them grew large enough to leave years ago. Others are still in the building, sharing desks and meeting rooms and the kind of informal advice that rarely appears in a budget. For founders like Dev Malik, the loss is partly practical and partly personal. Before Foundry Hall, he said, he was pitching from the back table of a coffee shop. Then he got a desk there, and with it a place to bring people he hoped would take him seriously. "My first investor meeting happened in that building," he said. That kind of starting point is what the twelve current companies stand to lose. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about finding space for them. She did not say when an arrangement might be reached. With the closing weeks away, the timing matters. Ross did not describe the closure as a failure of the model. The accelerator's track record, she suggested, was never the question. The question was whether anyone would step in once the bank stepped back, and so far no one has. Foundry Hall's staff will spend October winding down operations and helping the remaining companies prepare to move. After that, the doors that opened in 2018 will close. The 1,140 jobs its startups created will remain, spread across a city that will no longer have the place where many of them began.

MoonshotAI: Kimi K3.1 Pro: # Arcwell launches Relay, an AI platform for customer support Arcwell has introduced Relay, an artificial intelligence platform that examines customer support conversations, suggests answers to agents and identifies customers who may be at risk of canceling. It can also create summaries, combining response assistance, conversation recaps and cancellation warnings in one product. Together, those features extend AI’s role beyond drafting a reply to summarizing an interaction and flagging the possibility that a customer may leave. They also illustrate the technology’s expanding role in customer service, where it is being applied to both agents’ work and customer-retention efforts. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. That search problem sits at the center of Arcwell’s pitch. Relay requires access to support transcripts and internal help documents, allowing it to draw on a company’s existing material when analyzing conversations and producing suggestions or summaries. Relay was tested by 12 companies over six months. Arcwell says agents using the platform closed tickets 14% faster, its clearest quantified claim about the product’s performance. The company did not release customer-retention results, leaving open whether Relay’s warnings about potential cancellations led to fewer customers leaving. The platform costs $65 per agent each month. Alongside the subscription price, companies considering Relay will have to weigh the sensitivity of the information it needs to access. That makes data handling part of the buying decision, not just an implementation detail. 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, giving them two options for how long information is kept. Those commitments address some of the practical questions surrounding Relay, but the results Arcwell has released focus on speed rather than retention. The six-month test provides an early measure of its effect on ticket handling; the platform’s ability to help companies keep customers has not been quantified.

Anthropic: Claude Opus 5.5: # Arcwell's Relay Uses AI to Feed Answers to Customer Support Agents Arcwell wants to spare customer support agents the hunt for answers their own companies already have on file. The company has launched Relay, a platform that uses artificial intelligence to examine customer support conversations and suggest responses to the agents handling them. Relay can also flag customers who may be about to cancel and generate summaries of support conversations. "Support agents spend too much time searching for information that their companies already have," said Arcwell CEO Priya Shah. Arcwell's case for the product rests on a six-month trial involving 12 companies. Agents using Relay closed tickets 14% faster, according to the company. Shaving time off each ticket can add up quickly for teams that handle a steady stream of requests, though the figure comes from Arcwell itself. The trial results also leave a gap. Arcwell did not release customer-retention results, so there is no public evidence yet on whether Relay's cancellation warnings actually help companies hold on to customers. Of the platform's three main features, that one has the clearest potential link to revenue, which makes the missing data more conspicuous. Relay costs $65 per agent each month, so the bill grows with the size of the support team. A company with 100 agents, for example, would pay $6,500 a month. Then there is the question of data. To work, Relay requires access to support transcripts and internal help documents. Those records can contain sensitive customer and company information, and some businesses may hesitate to hand them to an outside vendor. Arcwell is trying to address those concerns. The company says customer data is encrypted and is not used to train shared models, and customers can choose a 30-day or 90-day retention period. For prospective buyers, the decision is likely to come down to a trade-off. They will have to weigh faster ticket handling against a per-seat price, broad access to internal records and, for now, unproven claims about customer retention.