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

Anthropic: Claude Opus 5.5: # Raincheck Raises $8 Million to Help Cities Prepare Drainage Systems Before Storms *The Austin startup's software combines drainage maps with rainfall forecasts to help local governments decide where to send maintenance crews ahead of heavy rain.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding. The software maps cities' drainage systems and combines those maps with rainfall forecasts. Cities use it to decide where to send maintenance crews before a storm. Basin Ventures led the round, with participation from existing investor Long Mile. The funding brings Raincheck's total raised to $10.5 million. CEO Ana Flores said the challenge for cities is not identifying flood-prone areas but prioritizing the work that can prevent flooding: > "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 outlines its approach in [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 serve as recommendations. Emergency managers retain control over crew deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says it will use the new funding to hire eight engineers and expand into Florida.

MoonshotAI: Kimi K3: # Raincheck Raises $8 Million to Help Cities Get Ahead of Flood Risk *The Austin startup's software maps drainage systems and pairs them with rainfall forecasts, helping local governments decide where to send maintenance crews before a storm hits.* Raincheck, an Austin-based company that makes flood-risk software for local governments, has raised $8 million in seed funding, the company announced today. Raincheck maps municipal drainage systems and combines those maps with rainfall forecasts, [according to the company's published methodology](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews before a storm. The seed round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has raised $10.5 million to date. CEO Ana Flores said the software is designed to solve a problem that goes beyond simply identifying flood-prone areas: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations; emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Anthropic: Claude Opus 5.5: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 Sessions *The page was publicly accessible for four hours Monday. Cloudnote says it has found no sign that anyone other than the researcher who reported it viewed the data, but its investigation is not finished.* Cloudnote left a debugging page containing excerpts of customer prompts accessible without authentication for four hours on Monday, exposing fragments of text from 1,842 user sessions, according to the company's logs. Each excerpt could include up to 200 characters from a prompt, enough in some cases to capture names, work details or personal questions, depending on what a user had typed. Full documents, passwords and payment information were stored elsewhere and were not on the page, the company said. Cloudnote removed the page 19 minutes after a researcher notified it of the problem. The company says it has found no evidence that anyone besides that researcher accessed the page. That conclusion is provisional. Cloudnote has not completed its investigation, and it has not said how many individual users the affected sessions represent or whether any excerpts were copied before the page came down. The company has 3.2 million registered users. The incident shows how a tool built for internal troubleshooting can surface text that people assume stays private. Cloudnote's final findings will determine whether the exposure ended with a single researcher or reached further.

MoonshotAI: Kimi K3: # Cloudnote Left Excerpts of Customer Prompts on an Unauthenticated Page for Four Hours *The company says a security researcher appears to have been the only person to view the data — but its investigation is not finished.* Cloudnote, an AI note-taking service with 3.2 million registered users, exposed excerpts of customer prompts on a debugging page that anyone could reach without logging in for roughly four hours on Monday, the company said. The page contained prompt fragments from 1,842 user sessions, according to Cloudnote's review of its access logs. Each excerpt was capped at 200 characters — a limit still large enough to capture a name, a health concern, a business detail or anything else a user happened to type. The exposure ended quickly once it was flagged. A security researcher discovered the page and notified Cloudnote, which removed it 19 minutes later. The company says its logs show no evidence that anyone besides the researcher accessed the page during the hours it was live. That assurance comes with a caveat: Cloudnote has not completed its investigation, and it has not said how the debugging page became publicly accessible in the first place. Its finding that the researcher was the only visitor is based on an ongoing review and could change. Cloudnote said full documents, passwords and payment information were stored elsewhere and were not present on the page. Still, even short prompt excerpts can be revealing: people routinely paste meeting notes, drafts and personal reflections into AI tools, making fragments of their input sensitive regardless of length.

Anthropic: Claude Opus 5.5: # KiteBridge Glitch Sends $2.8 Million in Duplicate Reimbursements Ninety-one companies that use KiteBridge to repay employees' business expenses ended up paying many of those expenses twice on Tuesday, after a failure in the company's payment system issued 13,400 duplicate reimbursements worth a combined $2.8 million. By Wednesday evening, $2.3 million had been recovered. The remaining roughly $500,000 was still pending. Employees were not left out of pocket. No money was withdrawn from their accounts, and the duplicate payments were funded by their employers. ## How it unfolded KiteBridge makes expense-management software, which companies use to track spending such as travel and meals and to reimburse workers for it. At 8:45 a.m. Tuesday, the company deployed version 4.2 of its software. The first customer complaint arrived less than half an hour later. KiteBridge suspended automated payments at 10:03 a.m., about an hour and 20 minutes after the deployment and roughly 50 minutes after that first complaint. The available information does not establish whether the version 4.2 update played a role in the failure. ## What went wrong The duplicates stemmed from how KiteBridge's system handled a delay. When the software issues a reimbursement, it sends the request to a payment processor, the outside service that actually moves the money. On Tuesday, the processor timed out, meaning it did not respond within the expected window. Payment software often reacts to a timeout by retrying the request, because it cannot be sure whether the first attempt went through. KiteBridge's fallback system, a backup mechanism designed to keep payments moving when the normal path stalls, treated those retries as brand-new reimbursement requests rather than repeat attempts at existing ones. As a result, the same expense could be paid more than once. ## A promise of "exactly-once" payments The failure is the kind of error KiteBridge had said its software was built to prevent. An April changelog, the running list of changes a software company publishes with its releases, said version 4.0 had introduced "exactly-once payment protection." In payments engineering, "exactly once" refers to safeguards meant to ensure a transaction is carried out a single time, even if the request behind it is sent repeatedly. It is not yet clear why that protection did not stop Tuesday's duplicates, or whether it extended to the fallback system. ## A delayed disclosure KiteBridge's public explanation lagged behind events. Its incident page, where the company posts updates on service problems, initially described the issue as delayed reimbursements. It did not disclose that duplicate payments had been made until six hours after automated payments were suspended, around 4 p.m. Tuesday. For the employers whose accounts funded the extra payments, that framing understated the problem. A delay suggests money arriving late. The actual error sent millions of dollars out the door that companies then had to recover. ## What comes next CEO Lena Ortiz apologized and said the company would publish a technical report on the incident. That report will likely face several questions. Customers may want to know why the fallback system did not recognize retries, how a safeguard advertised as preventing duplicate payments failed to do so, whether the morning's software update was connected, and why the company's first public description did not mention duplicate payments. The status of the roughly $500,000 still pending will also remain a concern for affected customers until it is recovered.

MoonshotAI: Kimi K3: # KiteBridge Update Triggers $2.8 Million in Duplicate Reimbursements Expense-management software maker KiteBridge is working to recover $2.8 million in erroneous payments after a flawed software update caused its platform to reimburse employee expense claims multiple times, generating 13,400 duplicate payments across 91 corporate customers. By Wednesday evening, the company said it had recovered $2.3 million, with roughly $500,000 still pending. Recovering mistaken payments is often slow work: once a transfer settles in a recipient's account, reversing it generally requires cooperation from banks or the recipients themselves. KiteBridge said no employees had money withdrawn from their personal accounts; the duplicate reimbursements were funded entirely by the employers. KiteBridge's platform handles the reimbursement process for corporate finance teams — approving employees' out-of-pocket expenses, such as travel and meals, and paying the money directly into workers' bank accounts. That central role in moving company money is what made Tuesday's failure so costly. ## Seventy-eight minutes of duplicates The failure began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its software. The first customer complaint arrived 27 minutes later, and the company suspended automated payments at 10:03 a.m. By then, in a window of less than an hour and a half, the system had already generated thousands of duplicate transactions. The trigger, according to the company, was a timeout at its payment processor — the outside service that actually transfers approved reimbursements to banks. When a transfer request takes too long to confirm, payment software ordinarily retries it, a standard defense against momentary network failures. KiteBridge's fallback system — the backup pathway meant to keep payments moving when the primary process falters — mishandled those retries. Rather than recognizing each one as a repeat of an existing request, it treated every retry as a new reimbursement, and paid it. ## A protection that didn't protect The incident is an especially uncomfortable one for KiteBridge. In an April changelog, the company announced that version 4.0 had introduced "exactly-once payment protection" — a reference to a long-standing challenge in financial software: guaranteeing that a transaction executes precisely one time, even when connections fail or requests are resent. Tuesday's events suggest that guarantee did not hold under the 4.2 update. The company has not said whether the protection failed, was bypassed by the fallback system, or was removed. The company's public communication also trailed events. KiteBridge's incident page initially described the problem as delayed reimbursements — language suggesting money was moving slowly, not twice. The distinction matters: a customer reading "delayed" might simply wait, while "duplicate" signals that funds have already left the account. The page first disclosed the duplicate payments six hours after automation was suspended, leaving some finance teams unaware for much of Tuesday that they had been charged for thousands of extra claims. CEO Lena Ortiz apologized to customers and said the company would publish a technical report examining the failure. KiteBridge has not said when that report will appear, how it is recovering the outstanding funds, whether version 4.2 remains in production, or whether affected customers will be compensated for costs stemming from the error.

Anthropic: Claude Opus 5.5: **6** 1. `#Flood software moves upstream` is missing the space after `#`. 2. The dek is not italicized. 3. `##Details` is missing the space after `##`. 4. The link `www.example.com/news` is not an absolute HTTPS URL. 5. The list item begins with `*` instead of a hyphen. 6. `<div>…</div>` is raw HTML.

MoonshotAI: Kimi K3: 6

Anthropic: Claude Opus 5.5: # QueueZero's AI Meeting Assistant Turns Audience Remark Into Tasks for Its CEO *Proxy, which can attend meetings and hand off follow-up work to software agents, enters public beta next week at $30 per user per month.* QueueZero launched Proxy, an AI meeting assistant, on Wednesday. The product's debut offered an unscripted look at how the software handles ambiguous language. Proxy is designed to attend meetings, produce notes and assign follow-up work to software agents. During the launch demonstration, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the remark as an instruction. It created two tasks, both assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. The customer, at least, got the pricing sheet. ## Automatic tasks, confirmed messages The episode illustrates a distinction in how Proxy is built. According to QueueZero, the assistant requires confirmation before an external message is sent. Internal tasks, however, can be created automatically, which is what happened on stage. Administrators can disable automatic task creation, the company says. Under that design, a misread comment is more likely to produce an errant item on an internal task list than an unintended message to a client or partner. The demonstration also showed that Proxy acted on an informal remark from an audience member, not only on direct requests from the meeting's host or presenter. ## A broader challenge for AI assistants Meeting software has long offered transcription and summaries. Proxy belongs to a newer group of tools that go further by taking action on what they hear. That shift raises a practical problem: in a live conversation, people mix requests, jokes, complaints and asides, often in the same sentence. For a tool that converts speech into assigned work, telling the difference between an instruction and an expression of frustration is central to its usefulness. A system that is too literal can clutter task lists or misattribute requests. A system that is too cautious may miss genuine follow-ups. Organizations evaluating Proxy during the beta may want to look at how it decides which statements become tasks, how it attributes work to specific people, and how easily automatically created tasks can be reviewed or reversed before software agents act on them. ## Pricing and availability Proxy costs $30 per user per month. It enters public beta next week. QueueZero's confirmation requirement for external messages and the administrator option to turn off automatic task creation give companies some control over how much the assistant does on its own. Wednesday's demonstration showed why some customers may want to review those settings before deploying it in their own meetings.

MoonshotAI: Kimi K3: # QueueZero's Proxy AI Takes Audience Remark Literally, Assigns CEO Task to 'Stop Talking' *The launch of QueueZero's meeting assistant included an unplanned test of its automatic task-creation rules.* QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant that the company says can attend meetings, produce notes and assign follow-up work to software agents. The product's debut also delivered an unscripted demonstration of how literally it can take what it hears. During the launch event, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was running alongside the demonstration, interpreted the remark as an instruction. It created two tasks assigned to QueueZero chief executive Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first, according to the company. ## Guardrails The episode drew attention to a central question for this class of software: how much authority should an assistant have to act on what it hears? QueueZero's answer is a split set of rules. The company says Proxy requires user confirmation before any external message is sent, a safeguard intended to keep the tool from contacting customers or partners on its own. Internal tasks, by contrast, can be created automatically, which is how two assignments for Wu appeared without anyone's approval. Administrators who want tighter control can disable automatic task creation entirely. ## Pricing and availability Proxy costs $30 per user per month and enters public beta next week. QueueZero is entering a crowded market. Large tech companies and a long list of startups already sell AI note-takers for virtual meetings, and the industry's newer pitch is that such tools should do more than transcribe — they should schedule follow-ups, draft messages and delegate work to other software. That shift from recording to acting raises the cost of misreading a room. A garbled transcript is an inconvenience; an automated action can email the wrong person or, as on Wednesday, instruct a chief executive to be quiet. Proxy had, in a narrow sense, done its job — it found action items and assigned them to the person named. What it did not do was distinguish between a speaker in the meeting and a voice from the audience. The launch did, at minimum, establish that the software was listening. Whether customers will trust Proxy to act on what it hears may depend on how carefully those administrative controls are set — and on how clearly the company can show that Wednesday's misfire was a configuration question rather than a habit.

Anthropic: Claude Opus 5.5: # Foundry Hall to Close After Losing Its Largest Sponsor Foundry Hall, the Milwaukee nonprofit that has helped launch more than 200 startups since 2018, will close October 31 after its largest corporate backer declined to renew its sponsorship. Blank River Bank had supplied 38% of the accelerator's annual budget. Without that money, Foundry Hall can operate through October but cannot fund another full program. Six employees will lose their jobs, and the twelve startups now working out of its space will need to find somewhere else to go. Executive director Camille Ross said the organization had hoped another funder would step in to cover the shortfall before the money ran out. "We kept waiting for the replacement check that never came," she said. In a statement, Blank River Bank said its decision reflected a shift in its community investment priorities and had nothing to do with how Foundry Hall performed. By most measures, that performance was substantial. The 212 startups Foundry Hall has supported over the past seven years now employ a combined 1,140 people. For a city that has long worked to build an entrepreneurial base beyond its industrial past, that is no small figure. The numbers do not fully capture what the space meant to the people who passed through it, though. Dev Malik, a founder who came through the accelerator, recalled how he worked before he had access to it. He took meetings and made his case to potential backers wherever he could find a seat. Foundry Hall gave him somewhere to bring them. "My first investor meeting happened in that building," Malik said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." For the twelve companies still in residence, the more pressing question is where they will be in November. Deputy Mayor Susan Cole said the city is in talks with building owners and local universities about finding space for them. No arrangement has been announced, and the deadline is fixed. The closure also leaves a gap that is harder to measure than square footage. An accelerator offers more than desks and meeting rooms. It offers a place where early founders can find one another, along with the mentors and investors they might not otherwise reach. Whether that network holds together once the building empties remains uncertain. Foundry Hall's staff will spend the coming weeks winding down operations and helping the current companies plan their move. When the doors close at the end of the month, the startups it helped build will carry on. The six people who ran it will be looking for work, and the next founder with an idea and a laptop may find that the nearest open table is back at the coffee shop.

MoonshotAI: Kimi K3: # Foundry Hall to Close After Seven Years, Leaving Twelve Startups Without a Home *The Milwaukee nonprofit accelerator backed 212 companies but could not replace the bank that supplied more than a third of its budget.* MILWAUKEE — Foundry Hall, the nonprofit startup accelerator that has anchored this city's early-stage business community since 2018, will close on October 31, its leaders confirmed this week, ending a months-long effort to survive the loss of its biggest funder. The accelerator's largest corporate sponsor, Blank River Bank, declined to renew its sponsorship, withdrawing the 38 percent of Foundry Hall's annual budget it had long supplied. A bank spokesperson said the decision reflected a change in the bank's community investment priorities and was unrelated to the accelerator's performance. That explanation did little to ease the blow. Executive director Camille Ross said her team spent months pursuing new funding to close the gap. "We kept waiting for the replacement check that never came," she said. Foundry Hall has enough money to operate through October but not enough to run another full program. Six employees will lose their jobs, and the twelve startups currently working in its space must move out by month's end. There will be no next cohort. News of the closing moved quickly through Milwaukee's startup community, where Foundry Hall had become a default first stop for new founders. Since opening in 2018, the accelerator has supported 212 startups, which collectively employ about 1,140 people — a record that made it one of the city's most visible efforts to grow companies at home rather than watch them leave for bigger markets. For founders, though, the loss is less about numbers than about a place. Founder Dev Malik said the building gave him legitimacy before he had a track record. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." The city is trying to soften the landing for the companies still inside. Deputy Mayor Susan Cole said officials are "speaking with building owners and local universities" about space for the displaced startups, though no arrangement has been announced and the deadline is fixed. Space, in any case, was only part of what Foundry Hall offered. Its programs paired young companies with mentors, sharpened their pitches, and put them in rooms where investors actually appeared — the kind of infrastructure that is easy to overlook until it is gone. For now, the lights stay on. Twelve companies still come to work each day, and six employees still show up to help them. The founders who once traded coffee shop tables for desks at Foundry Hall will spend their final weeks there doing what founders always do: searching, again, for someone willing to take a chance on them.

Anthropic: Claude Opus 5.5: # Startup Spotlight: LedgerLark Bets That AI Can Untangle the Back Office Most conversations about AI in finance focus on trading desks and research. LedgerLark is looking at what happens after a trade is done. The startup builds AI agents for post-trade operations at small investment firms. That's the work of making sure that what a firm thinks happened, what its counterparties confirmed and what its custodian recorded all match. When those records don't line up, the result is an "exception," and someone has to chase it down. At larger institutions, that work is supported by dedicated teams and expensive software. At smaller firms, it often falls to a handful of people working through spreadsheets and email. ## Where the idea came from Co-founder Mina Batra spent seven years in operations at a regional brokerage. She began exploring the idea after watching a six-person team spend three days resolving a single trade mismatch. The cause was a copied account number. The anecdote points to a familiar problem in operations. The original error may be trivial, but finding it can be expensive. The data needed to diagnose the problem usually exists, but it sits in different systems and formats, and people have to compare it by hand. Batra's co-founder, Owen Carr, approaches the problem from the systems side. He previously built reconciliation systems at a large bank. That gives the pair a mix of experience: one has done the investigative work firsthand, and the other has engineered tools meant to reduce it. ## What the product does LedgerLark's agents compare three sources of record: - trade confirmations - custodian records - a firm's internal ledgers When the agents find a discrepancy, they propose a correction. A person then reviews it and decides whether to approve it. The limits matter as much as the capabilities. LedgerLark's software does not execute trades or move money. That narrower scope may make the product easier for cautious compliance teams to accept, since the AI's role is to suggest rather than to act. It also means the product's value depends on whether its proposals are accurate enough to save reviewers time. If people end up double-checking every suggestion from scratch, much of the benefit disappears. ## Early traction LedgerLark says five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. The company has raised a $3.5 million seed round led by Motive Row. The most striking figure comes from the company itself. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. If that result holds up, it would be a meaningful change for a small operations team. But it is a single, company-reported example. It doesn't yet show how consistent results are across customers, how long the improvement took or what kinds of exceptions remain unresolved. ## The competitive picture LedgerLark is not entering an empty market. Established reconciliation vendors already sell to financial firms, and large banks often build their own software in-house. The startup's apparent bet is that smaller investment firms are underserved, and that AI agents can handle messy, investigative cases that rules-based tools struggle with. Whether that distinction holds is an open question. Incumbent vendors can add AI features to products customers already trust and have already approved. LedgerLark will need to show that its approach is different in practice, not just in description. ## What could slow it down Selling to financial institutions is rarely quick. Even small firms may require lengthy security and procurement reviews, especially for software that touches records of trades and client accounts. Long sales cycles can strain a seed-stage company. They can also make it harder to know whether slow growth reflects demand or just paperwork. ## What to watch The next useful signals will be whether LedgerLark's early customers renew and whether the exception reductions it reports are repeated elsewhere. Another signal will be how often reviewers accept the agents' proposed fixes. The underlying problem is real and familiar to anyone who has worked in operations. The harder question is whether LedgerLark can solve it better than the tools already on the market.

MoonshotAI: Kimi K3: # Startup Spotlight: LedgerLark *Five paying customers, one company-reported win, and a long road through procurement.* The idea for LedgerLark traces back to a copied account number. At the regional brokerage where co-founder Mina Batra spent seven years in operations, that single error once triggered a trade mismatch that took a six-person team three days to resolve. That kind of failure is common in post-trade operations, the unglamorous work of confirming that trades settled as intended and that every record agrees. Batra's co-founder, Owen Carr, built reconciliation systems at a large bank, and the pair's shared insight is straightforward: the exceptions — the small fraction of trades that don't match cleanly — consume a disproportionate share of operations staff time, and small investment firms have the least capacity to absorb that work. LedgerLark's product is a set of AI agents aimed at those firms. The agents compare trade confirmations, custodian records and internal ledgers, then propose corrections for human approval. The company is careful about scope: its software does not execute trades or move money. That boundary matters. If an agent suggests a bad fix, a person can catch it before anything moves; the trade-off is that the efficiency gains depend on how much reviewers come to trust the suggestions. ## Early traction, with caveats Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000 — implying annual revenue somewhere between $180,000 and $550,000, real income but still early. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. That figure comes from the company and reflects a single deployment; whether it generalizes across firms with different systems and data quality is an open question. The company has raised a $3.5 million seed round led by Motive Row — a modest sum for selling into financial services, where trust is expensive to earn. ## The harder questions LedgerLark does not have the reconciliation problem to itself. Established reconciliation vendors have sold matching and exception-management software for years, and many banks run internal systems built for this work. The company's bet is that small firms are underserved — too small for enterprise contracts, too busy to build in-house — and that agents proposing fixes, rather than just flagging discrepancies, are meaningfully different from what incumbents offer. Both claims are plausible; neither is proven. Reconciliation tooling is also sticky: once a firm closes its books a certain way, switching means re-validating the whole process. Procurement may be the bigger obstacle. Financial institutions often require lengthy security and procurement reviews, and a startup with five customers and a short track record will face skeptical committees. Long sales cycles can strain a small company's finances even when the product works. There are also product questions LedgerLark has yet to answer publicly: how often do its agents propose the wrong correction, how is that measured, and what happens as customers ask the software to handle more complex asset classes? None of this diminishes the observation that drew Batra to the problem. Post-trade operations remain full of repetitive, high-stakes manual work, and the people doing that work — as Batra once did — know exactly where the time goes. The next year or two should show whether agent-proposed corrections can earn durable trust in a corner of finance where mistakes are remembered longer than efficiencies.

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

Anthropic: Claude Opus 5.5: # HelioDock Launches Night Shift Remote Robot Monitoring and Raises $24 Million *The warehouse robotics maker says its new fleet software cut unplanned stoppages during a 14-month beta, though one customer found its value shrank when rural cell connections dropped.* HelioDock, a maker of autonomous warehouse robots, on Thursday released Night Shift, software that lets operators monitor their robot fleets without being on site. The company also said it had closed a $24 million Series B round led by North Coast Ventures. The launch puts HelioDock into a fleet-management market where Relay Robotics and YardOS already sell competing products. Its pitch rests on a practical promise: fewer middle-of-the-night emergencies for the people who run automated warehouses. A beta customer's experience suggests that how well Night Shift delivers may depend heavily on the strength of a facility's network connection. ## Built for the overnight call HelioDock frames the product around a familiar frustration for warehouse managers. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," said Chief Executive Mara Velez. Velez's example points to a hidden cost of automation. When a robot halts, someone has to figure out why, and outside business hours that can mean a trip to the building. Night Shift is designed to let operators see what their fleet is doing from elsewhere and, when the connection holds, take remote control. HelioDock tested the software for 14 months in a private beta with nine logistics companies running a combined 310 robots, an average of about 34 machines per company. According to HelioDock, those customers experienced 17% fewer unplanned robot stoppages. The figure has not been independently verified. ## Only as reliable as the network Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity. During LTE outages at rural sites, he said, it became less useful. The limitation follows from how Night Shift handles a lost connection. When connectivity fails, HelioDock's robots revert to the behavior configured locally at the site, and remote control becomes unavailable. The robots keep operating under their existing on-site rules rather than being left without instructions. But the remote oversight at the center of the product is gone until the connection returns. For operators whose facilities sit in areas with dependable networks, that tradeoff may rarely come into play. For those relying on rural cellular service, an outage could leave a manager facing the same late-night choice Velez described. ## Pricing and competition Night Shift will cost $90 per robot per month, or $1,080 per robot per year, and becomes generally available on September 15. At that list price, a fleet the size of the entire beta program, 310 robots, would cost $27,900 a month. The per-robot model means costs scale directly with fleet size. Buyers weighing the software will be able to compare it against fleet-management products from Relay Robotics and YardOS. For companies that already run HelioDock robots, the comparison may come down to whether the claimed reduction in stoppages, and the overnight trips it could save, justifies the monthly fee. ## Fresh capital The Series B brings HelioDock's total funding to $39 million, which means the company had raised $15 million before this round. North Coast Ventures led the new financing. The funding arrives as HelioDock moves from selling robots toward also selling recurring software for managing them. Night Shift's monthly per-robot price gives the company a revenue stream tied to the size of the fleets it supports, beginning when the product opens to all customers in mid-September.

MoonshotAI: Kimi K3: # HelioDock Takes Warehouse Robot Oversight Remote With Night Shift *The autonomous-robot maker's fleet-monitoring software reaches general availability September 15 at $90 per robot per month, backed by a new $24 million Series B.* 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. The round brings the company's total funding to $39 million and accounts for the majority of the capital it has raised to date. The twin announcements mark an expansion for HelioDock, from a company that builds warehouse robots into one that also sells the software used to watch over them. The wager behind Night Shift is that as warehouses deploy larger autonomous fleets, the constraint shifts from the machines to the people responsible for keeping them running — and that operators will pay to centralize that oversight rather than staff it facility by facility. Chief executive Mara Velez described the problem the software is meant to eliminate. "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. Night Shift's answer is to move that intervention off the highway and onto a screen, giving a single operator visibility across an entire fleet without a commute. The software arrives behind an extended real-world trial: 14 months in a private beta involving nine logistics companies and 310 robots among them, an average of roughly 34 machines per operator. HelioDock says beta customers experienced 17 percent fewer unplanned robot stoppages during that period, a company-supplied figure that has not been independently verified. Unplanned stoppages are the metric most directly tied to Velez's two-in-the-morning scenario; whether the improvement holds beyond the beta's nine-company sample is a question the wider release will begin to answer. The beta defined the product's limits as clearly as its benefits. Because Night Shift operates over a network connection, its performance tracks the quality of that 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. When connectivity fails entirely, remote control becomes unavailable and the robots return to their locally configured behavior — a fallback that keeps the machines under their existing instructions but cuts the operator out of the loop. Pike's account gives the limitation a geography: rural operations dependent on LTE coverage are, by the drive-time logic of Velez's pitch, the very sites where remote oversight should be most valuable. HelioDock is not the only company selling this layer of software. Relay Robotics and YardOS both offer competing fleet-management products, so Night Shift launches into a market where buyers already have alternatives and where HelioDock's beta-derived claims will be measured against what rivals can show. The company's counterweight is its testing record: 14 months and 310 robots of operational data gathered before general release. Night Shift will cost $90 per robot per month — $1,080 per robot annually — when it becomes generally available on September 15. At the scale of the beta's combined 310-robot deployment, that works out to roughly $28,000 a month. The per-robot structure ties cost directly to fleet size, so a small facility pays the same unit price as the largest operator. September 15 is the date the claims meet the open market.

Anthropic: Claude Opus 5.5: C

MoonshotAI: Kimi K3: **C** Option C is the strongest lede because it is accurate, specific, and complete — it includes the key facts (the product, the price, the number of affected accounts, the duration, and what data was exposed) without exaggeration. The other options fail: - **A** falsely claims private messages were leaked (only workspace titles were exposed) and editorializes with "massive privacy disaster." - **B** buries the news, omitting the data exposure entirely and reading like corporate PR. - **D** falsely claims message contents were visible to hackers (contents were not sent, and no malicious access was found).

Anthropic: Claude Opus 5.5: # The Cheaper Model Kept Pace: What a Customer-Support Benchmark Shows When companies choose a language model to draft customer-support replies, the assumption is often that the best answers come from the most expensive option. A new benchmark of five models challenges that assumption. The top two performers were statistically indistinguishable in quality, yet one cost roughly 60% less to run and responded faster. ## The headline: a tie at the top, at very different prices Model North earned the highest average score, 84.2. Model Vale scored 83.8. The researchers found that this 0.4-point gap was not statistically significant. In practical terms, the test cannot say which of the two produces better answers. The costs were not close. North's estimated cost was $1.90 per 1,000 tickets, while Vale's was $0.74. Vale also answered 18% faster than North. For a team processing a million tickets, that works out to roughly $1,900 versus $740. Depending on scale, the dollar savings may or may not be decisive. The pattern is the more useful lesson: in this test, paying more did not buy measurably better answers. ## How the test worked Researchers gave each model the same 200 anonymized customer-support tickets. Each model produced one proposed answer per ticket. Three experienced support managers then graded the answers without knowing which model wrote them. Each score combined three qualities: - **Factual correctness:** whether the answer was accurate. - **Usefulness:** whether it would actually help the customer. - **Tone:** whether it sounded appropriate for a support interaction. The blind grading is a real strength. Graders could not favor a model whose name or reputation they recognized. Using managers with support experience also means the scores reflect the judgment of people who evaluate this kind of work for a living. The design has limits that matter for interpreting the results. Each model answered each ticket only once, so the results do not capture how much a model's output varies from one attempt to the next. The scores also blend three qualities into one number. Two models with similar totals could reach them differently, for example one stronger on accuracy and the other on tone. ## Strengths differ beneath the averages The two leaders had different specialties. North performed best on billing disputes. Vale performed best on technical troubleshooting. Readers should treat these as signals rather than verdicts. The benchmark summary does not say how many tickets fell into each category or whether these differences were statistically significant. Even so, they suggest a sensible step for buyers. A company whose support queue is dominated by billing questions may want to test North and Vale on its own billing tickets before assuming they are interchangeable. ## A clear drop-off further down Below the top two, quality declined as prices fell. Ember and Kite formed a middle tier, scoring in the high 70s at a fraction of Vale's cost. Harbor was the cheapest model by a wide margin, at $0.08 per 1,000 tickets, but also the weakest, with an average of 71.0. The researchers reported significance testing only for the North–Vale comparison. As a result, it is unclear whether the smaller gaps among the lower-ranked models are meaningful. The broad tiers are easier to defend than a strict ranking of each model. ## A shared blind spot: images One weakness appeared across the entire field. All five models performed poorly when a ticket depended on an image attachment, such as a screenshot of an error message. Customers often send exactly this kind of ticket, and none of the tested models handled it well. Teams deploying any of these models should consider routing image-dependent tickets to human agents or to a separate process. ## What the benchmark doesn't tell you The scope of the test was narrow. All tickets were in English and came from a single software company. A retailer, a bank or a telecom provider may see different results, as could any team supporting customers in other languages. The researchers also did not test agentic actions or tool use. Every model drafted a reply, but none was asked to look up an account, issue a refund or update a record. Many real support deployments depend on those capabilities. Finally, the costs are estimates. Actual spending will depend on ticket length, pricing changes and how each company integrates the model. ## The takeaway Within its limits, this benchmark supports a simple point: the most expensive model is not automatically the best value. Vale matched North's quality within the margin the test could detect, at well under half the cost and with faster responses. The findings are best used as a starting point for evaluation. Before committing to a model, a company should run a comparable blind test on its own tickets, in its own languages, and include the attachments its customers actually send.

MoonshotAI: Kimi K3: # The Top-Scoring Support Model Isn't the Obvious Choice A new benchmark of five language models on customer-support work produces a familiar-looking leaderboard — and a finding that undercuts it. Model North posted the highest average score, 84.2 out of 100. But Model Vale, at 83.8, was statistically indistinguishable from North on quality, answered tickets 18% faster, and did so at roughly 40% of the cost. For teams deciding what to deploy, the real story is less "North wins" than "the premium tier buys less than it used to." ## How the test worked Researchers assembled 200 anonymized customer-support tickets and had each model produce one proposed answer per ticket. Three experienced support managers graded every response blindly — without knowing which model wrote it — using a combined score covering factual correctness, usefulness, and tone. | Model | Avg. score | Est. cost per 1,000 tickets | |---|---|---| | North | 84.2 | $1.90 | | Vale | 83.8 | $0.74 | | Ember | 79.1 | $0.31 | | Kite | 76.4 | $0.22 | | Harbor | 71.0 | $0.08 | ## A flat top and a steep price curve Because the 0.4-point gap between North and Vale wasn't statistically significant, the fairest reading is a tie at the top. What genuinely differs is the economics. At these estimates, processing 100,000 tickets costs about $190 with North versus $74 with Vale — a gap that compounds at scale — and Vale returns its answers faster to boot. Below the leaders, quality declines in steps rather than collapsing. Ember (79.1) and Kite (76.4) form a middle tier at a fraction of Vale's price, while Harbor trails at 71.0 but costs just $0.08 per 1,000 tickets. Whether Harbor's savings justify a roughly 13-point drop from the leaders depends on the stakes of the conversation — a password reset and a disputed invoice are not the same interaction. ## Category strengths complicate the ranking The averages hid meaningful specialization. North performed best on billing disputes; Vale performed best on technical troubleshooting. A company whose queue skews technical has even less reason to pay North's premium, while a billing-heavy team might see a case for it — though the overall gap remains small enough that cost is likely to dominate the decision either way. More interestingly, the split suggests routing different ticket types to different models could beat crowning a single winner. ## One weakness none of them escaped The clearest shared failure was visual: all five models performed poorly when a ticket depended on an image attachment, such as a screenshot of an error message. For support organizations — whose queues are full of screenshots — this is arguably the most actionable finding in the study. Chasing small gains in text-only quality does nothing to address it. Teams should plan to keep humans on image-dependent tickets, or evaluate models specifically on vision before deploying them. ## Read the results with care Several limitations shape what this benchmark can actually say. All 200 tickets came from one software company and were in English, so the rankings may not transfer to other industries, customer expectations, or languages. The test measured answer quality only: researchers did not examine agentic actions, tool use, or multilingual support — precisely the capabilities vendors now compete on. Each model answered each ticket once, so the results reveal nothing about consistency. The composite score blends correctness, usefulness, and tone, which can mask trade-offs (a response can be warm and wrong). And while the grading panel was experienced, three raters applying one rubric is a narrow lens, and the costs are estimates rather than billed figures. ## The takeaway The benchmark's real lesson is about diminishing returns at the top of the market. The best-scoring model and the best-value model here are different products — and the quality difference between them is noise while the cost difference is not. For most support teams, the sensible default is the cheaper near-tie, with spending reserved for the ticket categories where a specialist actually earns it. The bigger worry isn't which model tops the table. It's what every model in the table still can't see.

Anthropic: Claude Opus 5.5: # Arcwell's Relay Puts AI Inside the Support Queue Arcwell is betting that the answers customer service agents need are already sitting in their companies' files. Its new platform, Relay, uses artificial intelligence to read support conversations as they unfold and suggest responses to the agents handling them. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Suggested answers are only part of the pitch. Relay also flags customers who may be about to cancel and generates summaries of support conversations, a combination meant to help agents work faster while also spotting accounts at risk. Arcwell tested the platform with 12 companies over six months. According to the company, agents using Relay closed tickets 14% faster. That figure comes from Arcwell itself, and it covers only one side of the product. The company did not release customer-retention results, leaving open whether Relay's cancellation warnings actually helped any of the test companies hold on to customers. Relay costs $65 per agent each month, which works out to $780 per agent a year. For a team of 50 agents, the bill comes to $3,250 a month. That pricing means the cost grows in step with the size of a support operation. Larger teams will pay more, and they will have more reason to weigh the speed gains Arcwell reports against the retention benefits it has not yet documented. The software also needs a great deal of company data to function. Relay requires access to support transcripts and internal help documents, the same material it draws on to produce suggestions and summaries. Giving an outside vendor that access is likely to be a sticking point for some buyers, and Arcwell has moved to address the concern. The company says customer data is encrypted and is not used to train shared models, meaning one client's conversations should not inform the suggestions another client receives. Companies can also limit how long their data is kept, choosing either a 30-day or a 90-day retention period.

MoonshotAI: Kimi K3: # Arcwell Says Its New AI Platform Sped Up Support Tickets by 14%. Whether It Keeps Customers From Leaving Is Another Question Arcwell has launched Relay, an artificial intelligence platform that follows customer support conversations as they unfold and suggests answers to agents along the way. The company says agents using the software during a six-month trial closed tickets 14% faster — though it released no data on whether those quicker interactions kept customers from walking away. Relay is built around three capabilities. It proposes answers to agents mid-conversation, drawing on information the company already holds. It flags customers whose messages suggest they may be about to cancel. And it generates summaries of support interactions, giving teams a record of what was discussed and decided. The through line, according to CEO Priya Shah, is wasted time. "Support agents spend too much time searching for information that their companies already have," she said. Relay is Arcwell's attempt to give that time back. Twelve companies piloted Relay over the six months leading up to launch, and Arcwell says the 14% improvement comes from that trial. The platform is now generally available, priced at $65 per agent per month. Getting value from Relay means handing over the keys. The platform requires access to a company's support transcripts and internal help documents — exactly the kind of material most businesses guard closely. Arcwell says customer data is encrypted and is never used to train models shared across its customer base, a stance likely aimed at companies wary of feeding proprietary information into someone else's AI. Companies can choose a data-retention period of either 30 or 90 days. The conspicuous gap in Arcwell's pitch is retention itself. Flagging at-risk customers is among Relay's headline features, yet the company released no customer-retention results from the trial. That leaves the 14% speed improvement as the only hard number in the launch — and whether Relay's early warnings actually save accounts as something its first paying customers will have to discover for themselves.