Head to head: Anthropic: Claude Fable 5.1 vs MoonshotAI: Kimi K3

Anthropic: Claude Fable 5.1: # 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 Fable 5.1: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 User Sessions *The page was reachable without a login for four hours on Monday. Cloudnote says it has found no sign anyone but a security researcher viewed it, but its investigation is not finished.* Cloudnote said an internal debugging page containing excerpts of customer prompts was accessible without authentication for about four hours on Monday. The company removed the page 19 minutes after a security researcher notified it of the problem. According to Cloudnote's logs, the page displayed excerpts from 1,842 user sessions. Each excerpt could include up to 200 characters from a prompt a user had typed into the service. Cloudnote has 3.2 million registered users, so the affected sessions represent a small share of its base, though the company has not said how many distinct accounts those sessions belong to. Cloudnote said full documents, passwords and payment information are stored in separate systems and were not present on the page. It also said it has found no evidence that anyone other than the researcher who flagged the issue accessed the page while it was exposed. Several questions remain open. The company has not completed its investigation, and it has not said what the excerpts contained, how the page came to be reachable without authentication, or whether it will notify users whose sessions appeared. Because prompt text can include personal or business details, even a partial, brief exposure of this kind tends to draw scrutiny from users and regulators.

MoonshotAI: Kimi K3: # Cloudnote debugging page exposed excerpts from 1,842 user prompt sessions *The company says it has found no evidence that anyone besides the researcher who reported the flaw viewed the data — but it cautions that its investigation is not yet complete.* Cloudnote said a debugging page containing excerpts of customer prompts was accessible on the public internet without authentication for four hours on Monday. The company's logs show the page held excerpts from 1,842 user sessions, each containing up to 200 characters of a user's prompt. Cloudnote removed the page 19 minutes after a researcher notified it of the exposure. Two hundred characters amounts to roughly two sentences — a small fraction of a typical document, but enough to capture names, business details or other sensitive text. Cloudnote said full documents, passwords and payment information are stored in separate systems and were not present on the page. Even if each session belonged to a different person, the affected users would amount to well under 1 percent of the company's 3.2 million registered users. Because the page required no login, anyone who encountered its address during that window could have read the excerpts. The central unanswered question is whether anyone did. Cloudnote said its review of access logs has so far found no evidence that anyone other than the researcher viewed the page, but it emphasized that its investigation is not complete. It has not explained how the debugging page became publicly reachable. Cloudnote has not said whether it plans to notify the users whose excerpts appeared in the logs, or whether it has disclosed the incident to regulators.

Anthropic: Claude Fable 5.1: # The Top Two Models Tied on Quality. One Costs Less Than Half as Much. A new benchmark of five language models on real customer-support work produced a result that should reshape how teams shop for these tools: the two highest-scoring models were statistically indistinguishable on quality, yet one of them costs about 60 percent less to run and returns answers faster. Model North posted the highest average score, 84.2, with Model Vale close behind at 83.8. That 0.4-point gap sounds like a ranking. It isn't. The researchers report that the difference was not statistically significant—meaning that with a different sample of 200 tickets, Vale could just as easily have landed on top. For practical purposes, the two models tie. The cost columns do not tie. North is estimated at $1.90 per 1,000 tickets; Vale at $0.74. Vale also produced its answers 18 percent faster. If you are choosing between these two on the basis of this benchmark, you are not paying for better answers with North. You are paying more, and waiting longer, for the same answers. ## How the test worked The setup was deliberately close to how support teams actually use these systems. Researchers assembled 200 anonymized tickets from a single software company and had each model draft one proposed answer per ticket. Three experienced support managers then graded every response blindly, without knowing which model wrote it. Scores combined three things managers care about: factual correctness, usefulness to the customer, and tone. That blind, human grading is the benchmark's main strength. Automated metrics tend to reward answers that *look* right. Managers who handle escalations for a living are harder to fool, and they weigh tone—an area where a technically correct but cold reply can still cost a customer. ## Below the top tier, the tradeoffs look different Three cheaper models filled out the field: Ember at 79.1 ($0.31 per 1,000 tickets), Kite at 76.4 ($0.22), and Harbor at 71.0 ($0.08). Here the pattern is more conventional—lower cost generally came with lower quality. But the shape of the curve matters. The drop from Vale to Ember is roughly five points, with cost falling by more than half. The drop from Ember to Harbor is another eight points for a cost reduction from cents to fractions of a cent. Whether those steps are worth taking depends entirely on the volume and stakes of a given team's queue. A company handling millions of low-risk "how do I reset my password" tickets might reasonably accept Ember's score. A company whose tickets routinely involve refunds and contract disputes probably should not. What the data does *not* support is treating any single-point gap between adjacent models as a meaningful separation. Only the North–Vale comparison was explicitly tested for significance, and it failed. Readers should assume similar noise elsewhere in the table. ## Specialization showed up at the margins Within the near-tie at the top, the two leaders had different strengths. North did best on billing disputes; Vale did best on technical troubleshooting. These sub-results come from smaller slices of an already modest sample, so they should be read as hints rather than conclusions. Still, they suggest that a team with a heavily technical queue has one more reason to lean toward Vale, while a team drowning in billing escalations might at least run its own comparison before committing. ## A blind spot shared by every model The single most consistent finding had nothing to do with rankings. All five models performed poorly whenever a ticket depended on an image attachment—a screenshot of an error, a photo of a damaged product, a marked-up invoice. Customers increasingly attach images instead of describing problems in words, and in this test, none of the models handled that well. For deployment planning, this may be more important than the leaderboard. Any workflow that routes tickets to a model should have a path for pulling image-dependent cases back to a human. ## What the benchmark can't tell you The limitations are real and the researchers were clear about them. All tickets were in English and came from one software company; results for a retailer, a bank, or a multilingual support desk could differ substantially. The test measured drafted answers only. It did not evaluate agentic actions—issuing refunds, updating accounts, calling internal tools—which is where many vendors now compete, and where failure modes are more expensive than a poorly worded reply. Two hundred tickets is enough to separate a 71 from an 84. It is not enough to separate an 83.8 from an 84.2, and that, in the end, is the point. Before paying a premium for the top line of a leaderboard, check whether the leaderboard actually says what it appears to say.

MoonshotAI: Kimi K3: # The Best Model on the Leaderboard Isn't the Obvious Choice A new benchmark of five language models on real customer-support work flips the usual narrative: the highest-scoring model is not meaningfully better than a rival that costs roughly 60% less and answers 18% faster. Researchers assembled 200 anonymized customer-support tickets and asked each of five models — North, Vale, Ember, Kite and Harbor — to draft one proposed answer per ticket. Three experienced support managers graded every response blindly, meaning they didn't know which model wrote it. That design choice matters: in model evaluations, a grader who knows an answer came from the prestigious option may unconsciously score it higher. Scores combined factual correctness, usefulness and tone into a single average. ## A tie at the top, a gap at the wallet North posted the highest average score, 84.2. Vale came in at 83.8. That 0.4-point gap was not statistically significant — in practical terms, the graders could not reliably tell the two models apart. | Model | Avg. score | 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 | The economics, however, are easy to tell apart. North costs an estimated $1.90 per 1,000 tickets; Vale costs $0.74. For a team handling 100,000 tickets a month, that's roughly $190 versus $74. Vale also produced answers 18% faster than North, which matters when customers are waiting on a reply. The result suggests that for straightforward ticket drafting, quality at the top of the market has converged enough that price and speed — not score — may be the real differentiators. ## Where the cheaper options fall away Below Vale, the trade-off gets steeper. Ember averaged 79.1, Kite 76.4 and Harbor 71.0, each cheaper than the last. The study didn't report significance tests for these mid-pack gaps, so differences between neighbors like Ember and Kite should be read cautiously. But the overall pattern is hard to miss: the five-point drop from Vale to Ember saves only about $43 per 100,000 tickets, and Harbor's rock-bottom price comes with a 13-point deficit to the leaders — the kind of gap blind graders almost certainly noticed. In other words, the value question isn't "cheapest versus best." It's whether the premium for a top-tier model buys anything measurable. At the very top, this benchmark says it doesn't; one tier down, it apparently does. ## Different strengths, one shared weakness The category breakdown complicates any single-winner story. North performed best on billing disputes, while Vale performed best on technical troubleshooting. For support teams with a predictable mix of ticket types, that points toward routing — sending billing questions to one model and technical ones to another — rather than crowning a single champion. The more sobering finding is what none of the models could do. All five performed poorly on tickets that depended on an image attachment, such as a screenshot of an error message. That's not a ranking problem; it's a capability gap across the board — and for many support teams, screenshots make up a large share of incoming tickets. ## Read the results with caveats The benchmark is narrow in ways that matter. All 200 tickets came from one software company and were in English, so the results may not transfer to retail, finance or healthcare support, or to multilingual queues. Researchers did not test agentic actions — issuing refunds, looking up accounts, calling external tools — which is increasingly where vendors claim their models differ. The grading, while blind and done by three experienced managers, still reflects human judgment of subjective qualities like tone. And 200 tickets is a modest sample for detecting small differences, worth remembering before anyone over-interprets a fraction of a point. ## The takeaway The clearest signal here is not that North is the best model. It's that the top two models are effectively tied on quality while differing sharply on cost and speed — and that every model tested shares the same blind spot on image-based tickets. For buyers, the framework shifts accordingly: benchmark against your own ticket mix, route by category where it pays, and don't pay a premium for a fraction of a leaderboard point that graders can't reliably detect.

Anthropic: Claude Fable 5.1: # Startup Spotlight: LedgerLark Wants to Shrink the Post-Trade Cleanup Pile *An AI-agent startup is betting small investment firms will pay to have mismatched trade records diagnosed, not just flagged.* ## The problem that started it Mina Batra spent seven years in operations at a regional brokerage, and one incident stuck with her. A six-person team spent three days tracking down a single trade mismatch. The cause turned out to be a copied account number — a transcription error that propagated through confirmations, custodian records and internal books until no two sources agreed. That experience led Batra to start exploring a product idea with Owen Carr, who had built reconciliation systems at a large bank. Together they founded LedgerLark, which builds AI agents for post-trade operations at small investment firms. ## What the product does LedgerLark's agents pull in three sets of records that should match but often don't: trade confirmations, custodian statements and the firm's internal ledger. When they disagree, the agents try to work out why and propose a correction. A person then approves or rejects it. The scope is deliberately narrow. The product does not execute trades and does not move money. That limits the agents' potential for harm to recommendations rather than actions, and it may make the software easier to explain to compliance teams wary of automation near client assets. ## Early traction, with caveats Five firms currently pay for LedgerLark. Annual contracts run from $36,000 to $110,000, suggesting pricing scales with customer size or complexity rather than a flat fee. The company reports that one customer reduced its unresolved daily exceptions from 70 to 18. That figure comes from LedgerLark, not an independent audit, and a single customer is a thin basis for generalizing. Still, if the result holds across other installations, it would be a meaningful reduction in the manual work Batra witnessed firsthand. LedgerLark has raised a $3.5 million seed round led by Motive Row. ## The competitive picture Reconciliation is not a new category. Established vendors have sold matching software to financial firms for decades, and large banks typically run their own internal systems. LedgerLark's wager is that the smallest firms are underserved by both: too small for enterprise contracts and too lean to build in-house. Whether AI agents genuinely improve on rules-based matching is an open question. Traditional systems are good at flagging mismatches; the harder work is diagnosing the cause and drafting the fix. If LedgerLark's agents do that second part reliably, the pitch holds together. If they mostly surface the same exceptions with a conversational layer attached, incumbents could add similar features without much difficulty. ## What could slow it down The most obvious obstacle is not technical. Financial institutions, even small ones, often require lengthy security and procurement reviews before adopting software that touches trade data. A sales cycle measured in quarters is hard on a company with five customers and a seed round to stretch. There is also the question of trust in the recommendations. Human approval is a safeguard, but staff who approve everything by reflex would erode the benefit, and staff who distrust the agent would see none at all. Success may depend as much on how the product fits into daily routines as on how accurate it is. ## Worth watching for The useful signals over the next year will be whether additional customers report exception reductions comparable to the first, how long new deals take to close, and whether any established reconciliation vendor responds directly. LedgerLark has identified a real and tedious problem. Whether its approach is durably better than the alternatives is still being tested.

MoonshotAI: Kimi K3: # Startup Spotlight: LedgerLark *The startup builds AI agents that reconcile post-trade records at small investment firms — and leaves the final call to humans.* ## The insight Post-trade operations — the work of confirming that a trade recorded in one system matches the records kept by everyone else — tends to go unnoticed until something breaks. For Mina Batra, something breaking was the starting point. Over seven years in operations at a regional brokerage, Batra saw plenty of routine friction, but one episode stood out: a six-person team spent three days resolving a single trade mismatch traced to an account number copied incorrectly between systems. She began exploring whether software could handle more of that work, and the exploration eventually became LedgerLark. Her co-founder, Owen Carr, came at the problem from the opposite end of the market. At a large bank, he built reconciliation systems — the kind of dedicated tooling that large institutions can afford to develop and maintain, and that smaller firms generally cannot. ## The product LedgerLark's agents compare trade confirmations, custodian records and a firm's internal ledgers, then flag mismatches and propose corrections. Notably, they don't act on those proposals themselves: a human reviews and approves each correction. The product does not execute trades or move money. That boundary looks deliberate. It positions the software as an assistant to operations staff rather than a replacement for their judgment, and it may ease some of the concerns compliance teams raise about autonomous systems touching client assets. Whether it also limits how much time the product can actually save is a fair question — reviewing dozens of proposed corrections is still work. ## Early traction LedgerLark says five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. That implies annual recurring revenue somewhere in the low-to-mid six figures — modest in absolute terms, but a sign that firms will pay for help with this particular headache. The company also reports that one customer reduced its unresolved daily exceptions from 70 to 18 after adopting the product. The figure is self-reported and comes from a single customer, so it's best treated as an encouraging data point rather than a guaranteed outcome. If similar reductions hold across clients, though, the time savings for small operations teams could be substantial. To date, the company has raised a $3.5 million seed round led by Motive Row — enough to fund the next phase, though not so much that the company can afford a long wait for results. ## The harder questions LedgerLark has company in this market. Established reconciliation vendors have sold into financial institutions for years, bringing existing relationships and deep integrations, and some banks run their own internal reconciliation software. A five-customer startup will need a clear reason for firms to switch — or to buy for the first time. Sales cycles are another unknown. Financial institutions often require lengthy security and procurement reviews before adopting new software, which can stretch a young company's runway and slow the conversion of pilots into contracts. There are product questions, too. The value of LedgerLark's agents depends on how accurate their proposed corrections are: if humans override them frequently, trust erodes quickly. And exceptions in post-trade operations are often idiosyncratic — what works at five firms may not transfer neatly to the next fifty. ## What to watch A few things will say a lot about LedgerLark's trajectory over the next year or two: whether early customers renew and expand; whether any are willing to act as public references in an industry that prizes discretion; and whether incumbent vendors respond with agent-style features of their own. The underlying problem Batra observed — skilled people losing days to a copied digit — is real and widespread. The open question is narrower: whether LedgerLark's cautious, human-in-the-loop approach is the one small firms will actually buy.

Anthropic: Claude Fable 5.1: # QueueZero Launches Proxy, an AI Assistant That Turns Meeting Talk Into Tasks QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to sit in on meetings, generate notes and hand off follow-up work to software agents. Within minutes of its public debut, the product demonstrated both its central promise and the question it raises: how literally should software interpret what people say? The launch event featured a live demonstration in which Proxy listened to a staged meeting and populated a task board in real time. Partway through, a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy treated the comment as an instruction. It created two tasks and assigned both to QueueZero chief executive Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. He then continued the presentation. The moment underscored a design choice QueueZero has made deliberately. Proxy is built to turn conversation into action without waiting for a human to transcribe or triage. According to the company, the assistant can create internal tasks automatically whenever it detects a commitment or request during a meeting. Those tasks appear in the team's workspace and can be assigned to people or to other software agents that carry out the work. QueueZero says there are limits. Proxy requires explicit confirmation before it sends any message outside the organization, so the assistant could not, for example, have emailed the pricing sheet to the customer on its own. Internal task creation, by contrast, happens without a prompt by default. Administrators can disable automatic task creation entirely. The distinction between internal and external actions reflects a broader debate in the industry over how much autonomy to grant AI agents. Many vendors have introduced tools that draft documents, schedule meetings or summarize calls, but comparatively few allow software to assign work to colleagues without a human signing off first. QueueZero appears to regard internal tasks as low-risk and easily reversed, as Wu showed when he deleted the one he disagreed with. Critics of agentic systems have argued that even low-stakes errors erode trust. An assistant that misreads an offhand remark as an assignment is harmless once; an assistant that does it several times a week may find its output ignored. QueueZero did not address that concern directly at the event, though the demonstration itself served as an unplanned test of how Proxy handles ambiguous speech. It passed the literal-interpretation portion with distinction. Proxy will cost $30 per user per month. It enters public beta next week.

MoonshotAI: Kimi K3: # QueueZero's AI Meeting Assistant Takes Audience Remark Literally During Launch Demo *Proxy, an agentic meeting tool, assigned its own CEO two tasks after a customer's offhand comment — and only one of them survived the afternoon.* QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend meetings on a user's behalf, produce notes, and assign follow-up work to software agents. The company ended up demonstrating that last capability sooner than planned. During the launch event, with Proxy running live onstage, a customer in the audience remarked: > "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the comment as an instruction. It created two tasks assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed when the tasks appeared on the demonstration screen, completed the second one, and deleted the first. Pricing sheets reached attendees shortly afterward. The episode doubled as a live product demonstration, albeit not the one on the run-of-show. According to QueueZero, the behavior fell within product specifications. Proxy requires explicit user confirmation before any external message — an email to a customer, for example — is sent. Internal actions, including task creation in connected workplace tools, can be performed automatically. Administrators can disable automatic task creation entirely, the company said. The distinction is central to QueueZero's pitch. The company is positioning Proxy less as a note-taker and more as an operator — a system that listens for outcomes rather than summaries. Wednesday's mishap showed both sides of that design. The assistant correctly identified actionable content buried in an offhand remark; it simply had no way to know the remark was not, strictly speaking, a request. Proxy enters public beta next week at $30 per user per month. The launch lands in a crowded market. Transcription-first meeting tools and the major productivity platforms have spent the past year racing toward so-called agentic features — software that does not merely record a meeting but acts on it, drafting follow-ups, updating project boards, and routing work to other systems. QueueZero's wager is that execution, not documentation, is what business customers will pay a premium for. The incident also previews the governance questions such products raise. An assistant permitted to act internally without approval shifts control to configuration: which tools it may touch, whose tasks it can create, and when a human must sign off. QueueZero's answer — automated internally, gated externally, adjustable by administrators — reflects permission frameworks now taking shape across the category. For his part, Wu appeared unbothered. He completed the pricing-sheet task onstage and deleted the other before the demonstration ended. QueueZero said the public beta will proceed as scheduled.

Anthropic: Claude Fable 5.1: C

MoonshotAI: Kimi K3: **C** This lede is the strongest because it's accurate and specific: it states what happened (product pulled), the scope (4,860 accounts), the nature of the data (workspace titles), and the duration (11 hours). The others fail: - **A** falsely claims "private messages" leaked — only titles were exposed, not message contents. - **B** buries the security incident in PR language and omits the actual news. - **D** falsely claims message contents were visible to hackers — contents weren't sent, and no malicious access was found.

Anthropic: Claude Fable 5.1: # Software Flaw at KiteBridge Sends $2.8 Million in Duplicate Reimbursements to Employees of 91 Companies A faulty software update at KiteBridge, a maker of expense-management software, caused the company's systems to issue 13,400 duplicate reimbursement payments totaling $2.8 million on Tuesday morning, affecting ninety-one corporate customers and exposing a gap between the company's safety promises and what its software actually did. As of Wednesday evening, KiteBridge said it had recovered $2.3 million. The remaining roughly $500,000 was still pending. ## A fallback that made things worse The trouble began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its platform. KiteBridge's software sits between employers and their workers: when an employee submits an expense, the system approves it and instructs a payment processor to send the money. On Tuesday, that payment processor began timing out — failing to confirm, within the expected window, whether a payment had gone through. Under normal conditions, a system handling money is designed to retry a stalled payment carefully, checking first whether the original transaction actually completed. KiteBridge's fallback system did not do that. Instead, it interpreted each retry as a brand-new reimbursement request and paid it again. The result was that employees who had submitted a single expense received the money multiple times. Because employers fund reimbursements through KiteBridge, it was companies, not workers, who paid for the duplicates. KiteBridge said no employee had money withdrawn from a personal account. ## A protection that was supposed to exist The failure is notable because KiteBridge had told customers this exact problem was solved. In an April changelog accompanying version 4.0, the company said the release introduced "exactly-once payment protection" — a term borrowed from software engineering that means a given payment should be executed one time and only one time, no matter how many retries or network hiccups occur along the way. It is not yet clear whether version 4.2 broke that protection, bypassed it in the fallback path, or whether the protection never covered this scenario in the first place. KiteBridge has not said. ## Twenty-seven minutes to the first complaint, hours to full disclosure The first customer complaint reached KiteBridge at 9:12 a.m., less than half an hour after the deployment. Automated payments continued running for another fifty-one minutes. At 10:03 a.m., the company suspended them entirely. What customers were told, however, lagged well behind what the company knew. KiteBridge's public incident page initially described the problem as *delayed* reimbursements — the opposite of what was happening. It was not until roughly 4 p.m., six hours after automation had been shut off, that the page first acknowledged duplicate payments had been sent. That gap is likely to draw scrutiny from affected employers, some of whom may have spent Tuesday afternoon fielding questions from employees about unexpected deposits without knowing the cause. ## Recovery and an apology By Wednesday evening, KiteBridge reported it had clawed back $2.3 million of the $2.8 million, leaving several hundred thousand dollars outstanding. The company did not say how the remaining funds would be recovered or how long that process might take. Chief executive Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report explaining what went wrong. The company has not given a timeline for that report, nor has it said when automated payments will resume. For KiteBridge's customers, the questions now extend beyond the money. A payment platform's core promise is that it moves exactly the funds it is told to move — no more, no less — and that it tells its customers the truth when something breaks. On Tuesday, KiteBridge fell short on both counts.

MoonshotAI: Kimi K3: # KiteBridge Bug Triggers $2.8 Million in Duplicate Reimbursements *A payment-processor timeout turned a routine software update into a costly cleanup for 91 corporate customers.* A faulty software update at expense-management company KiteBridge unleashed more than 13,400 duplicate reimbursement payments totaling $2.8 million on Tuesday, forcing the company to halt automated payments and spend the next two days clawing the money back. By Wednesday evening, KiteBridge said it had recovered $2.3 million; the remaining $500,000 was still pending. ## A routine update, then a cascade KiteBridge's software sits between employees and their employers' finance teams: workers submit expenses, and the platform approves and reimburses them automatically. On Tuesday at 8:45 a.m., the company deployed version 4.2 of that platform. The first customer complaint arrived at 9:12 a.m., and by 10:03 a.m. — roughly 80 minutes after the update went live — KiteBridge had suspended automated payments entirely. By then, the damage was done. The failure began with a timeout at KiteBridge's payment processor, the outside service that actually moves money. When a payment request times out, software typically retries it — and well-designed systems recognize those retries as repeats of the same instruction, so a stalled payment isn't sent twice. KiteBridge's fallback system, the backup logic that takes over when the primary process falters, did the opposite: it treated each retry as a brand-new reimbursement request and paid it out again. That single misreading, repeated thousands of times, produced the $2.8 million in duplicate payments. Ninety-one corporate customers were affected. Because employers fund reimbursements, the duplicate payouts drew on company accounts; no individual employees had money withdrawn from their personal accounts. ## A safeguard that didn't hold The incident is especially awkward for KiteBridge because the company had advertised protection against precisely this scenario. An April changelog said version 4.0 introduced "exactly-once payment protection" — industry shorthand for guaranteeing that a payment instruction executes one time only, no matter how often it is sent or retried. KiteBridge has not explained how Tuesday's failure slipped past that safeguard — whether version 4.2 weakened it or whether the fallback system simply operated outside it. The answer matters to customers deciding whether the bug was a one-off or a flaw in the platform's core design. ## Slow to say "duplicate" The company's public communication lagged behind its internal response. KiteBridge's incident page initially described the problem as delayed reimbursements — a characterization at odds with reality, since money was moving too many times, not too late. The page did not disclose the duplicate payments until six hours after automation was suspended, leaving finance teams to discover unexpected debits on their own. ## Recovery and response KiteBridge spent Wednesday reversing the errant payments. By evening, $2.3 million of the $2.8 million had been recovered, with the rest still pending. The company has not said how long the remaining recoveries will take. CEO Lena Ortiz apologized to customers and said KiteBridge will publish a technical report on the incident. For a company whose software sits in the middle of corporate spending, that report — and whether it explains the gap between the promised "exactly-once" protection and Tuesday's cascade — will be closely watched.

Anthropic: Claude Fable 5.1: # Raincheck raises $8M to help cities fix drains before the storm hits *The Austin startup pairs drainage maps with rainfall forecasts so local governments can decide where to send maintenance crews first.* Raincheck, an Austin company that makes flood-risk software for local governments, said today it has raised $8 million in seed funding. The round was led by Basin Ventures, with participation from existing investor Long Mile. It brings Raincheck's total funding to $10.5 million. The company maps a city's drainage system and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm arrives. CEO Ana Flores said the problem is not knowing where flooding happens, but what to do about it in the hours before it does: > "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." The company has published [an explanation of how its forecasting methodology works](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. It declined to identify them, saying some of the contracts have not yet been publicly announced. The software's output is advisory. Its forecasts are recommendations, and emergency managers retain control over crew deployments and public warnings. Raincheck competes with other flood-planning vendors, including HydroMap, as well as municipal engineering contractors. The company said the new funding will go toward hiring eight engineers and expanding into Florida.

MoonshotAI: Kimi K3: # Raincheck Raises $8 Million to Help Cities Get Ahead of Floods *The Austin startup's software tells local governments which drains, pumps and repairs need attention before a storm arrives.* Raincheck, an Austin-based startup that makes flood-risk software for local governments, has raised $8 million in seed funding, the company announced today. The company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. The seed round was led by Basin Ventures, with participation from existing investor Long Mile. The round brings Raincheck's total funding 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." The company details how its software prioritizes repairs in a [public overview of its methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. The company's forecasts are recommendations; emergency managers retain control over deployments and public warnings. ## Competition 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 Fable 5.1: # Arcwell Bets on AI to Cut Support Agents' Search Time Arcwell has introduced Relay, a platform that uses artificial intelligence to analyze customer support conversations and hand agents suggested replies. It also flags customers who appear likely to cancel and generates summaries of support interactions. The problem Arcwell is targeting is one most support teams will recognize: answers exist somewhere in the company's systems, but agents burn time finding them. "Support agents spend too much time searching for information that their companies already have," said CEO Priya Shah. Arcwell has some early data to back the claim. Twelve companies piloted Relay over six months, and the company says agents using the tool closed tickets 14% faster. That is the only outcome Arcwell is sharing. It has not released figures on customer retention — a conspicuous gap, given that identifying customers at risk of cancelling is one of Relay's three headline capabilities. Until those results appear, buyers will have to take the churn-prediction feature on faith. Relay costs $65 per agent per month. The platform's value rests on how much it is allowed to see. To draft useful responses, Relay needs access to support transcripts and a company's internal help documents — exactly the kind of material that makes security and legal teams uneasy about third-party tools. Arcwell's response is a set of data-handling commitments. Customer data is encrypted, the company says, and is not used to train models shared across its client base, meaning one customer's support logs should not shape the suggestions another customer's agents receive. Companies also control how long Arcwell holds their data, choosing between a 30-day and a 90-day retention period. Those assurances address the most common objections, though they stop short of letting customers keep data entirely on their own systems. Relay arrives in a market where AI-assisted support tooling is increasingly common, and the differences between products tend to come down to accuracy, price and a buyer's comfort with sharing transcripts. Arcwell's pilot figure gives it something concrete on the first point, and its encryption and retention terms speak to the third. On the churn-prediction feature, the company has described what Relay does but not yet shown what it delivers — and that is the number prospective customers are most likely to ask for.

MoonshotAI: Kimi K3: # Arcwell Launches Relay, an AI Assistant for Customer Support Agents Arcwell has launched Relay, an artificial intelligence platform that analyzes customer support conversations and suggests answers to the agents handling them — a significant step for the company as customer service operations continue to adopt AI. The software does more than draft replies. Relay also flags customers whose messages suggest they may be at risk of canceling, and it generates summaries of support conversations, letting teams review and hand off cases without reading through every exchange. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. It's a frustration familiar to anyone who has watched an agent toggle between a chat window and a maze of internal documents. Relay is the company's attempt to close that gap, surfacing answers at the moment agents need them. Arcwell has early data to support the pitch. Twelve companies tested Relay over six months, and agents using the platform closed tickets 14% faster, according to the company. Notably absent from the announcement were customer-retention results — a conspicuous omission given that spotting would-be cancelers is one of Relay's selling points. Arcwell did not release those figures. Relay costs $65 per agent per month. Using it means opening up a good deal of internal material. Relay requires access to a company's support transcripts and internal help documents, a requirement that could give some prospective customers pause. Arcwell says customer data is encrypted and is not used to train shared models, and companies can choose a 30-day or 90-day retention period. Whether those assurances prove sufficient may shape Relay's prospects. AI tools are reshaping customer service, but they also concentrate some of a company's most sensitive conversations in a single system. For now, the 14% figure is the only hard number Arcwell is offering — and the company is betting that the hours agents save hunting for information will outweigh any discomfort over granting a machine access to it.

Anthropic: Claude Fable 5.1: **6**

MoonshotAI: Kimi K3: 6

Anthropic: Claude Fable 5.1: # HelioDock Launches Remote Fleet Software After 14-Month Beta, Raises $24 Million *Night Shift lets operators oversee warehouse robots off-site, though a beta customer says its usefulness drops when connectivity fails* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that allows operators to monitor and manage robot fleets from outside the facility. The company announced the product alongside a $24 million Series B led by North Coast Ventures. The launch moves HelioDock into fleet management, a category where Relay Robotics and YardOS already sell competing products. Night Shift is priced at $90 per robot per month and becomes generally available September 15. HelioDock is pitching it as a way to keep robots running overnight without putting a manager on-site, but the system's reach ends where the network does: when connectivity drops, remote control becomes unavailable and robots fall back to locally configured behavior. ## The pitch CEO Mara Velez framed the product around a specific inconvenience. "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 spent 14 months in a private beta with nine logistics companies and 310 robots before Thursday's release. According to HelioDock, beta customers saw 17% fewer unplanned robot stoppages during that period. The figure comes from the company and has not been independently verified. ## What a beta customer found Aaron Pike of Fulcrum Logistics, one of the beta participants, said the system worked well in facilities with strong connectivity. In locations that relied on rural LTE, he said, Night Shift became less useful during outages. That tracks with how the software is built. Remote monitoring and control depend on a live connection between the operator and the fleet. When that connection fails, the robots do not stop; they revert to whatever behavior has been configured locally on the machines themselves. The operator, however, loses the ability to intervene until the link is restored — which, in the two-in-the-morning scenario Velez described, is precisely the moment intervention would matter. HelioDock has not described any offline mode or local fallback for remote operators beyond the robots' preconfigured behavior. ## The funding The $24 million Series B brings HelioDock's total funding to $39 million. North Coast Ventures led the round. The company did not break out how the money will be allocated between its robot hardware business and the new software line. The raise lands as HelioDock shifts from selling only robots to selling a recurring software subscription on top of them. At $90 per robot per month, a fleet the size of the one in the beta — 310 robots — would generate roughly $27,900 a month in Night Shift revenue if every unit were enrolled. ## Where it sits in the market Fleet-management software for warehouse robots is not a new idea. Relay Robotics and YardOS both offer products in the space, and HelioDock will be competing with them for operators who already have mixed or single-vendor fleets. HelioDock's angle is its own hardware: Night Shift is built by the same company that makes the robots it monitors. Whether that integration translates into an advantage over standalone fleet-management vendors is something the September rollout will begin to test. The 17% stoppage reduction is the central number HelioDock is putting forward, and it is worth reading carefully. It reflects nine customers over 14 months, it was measured by the company, and it has not been checked by anyone outside HelioDock. Pike's account suggests the gains were not uniform across sites — strong where the network was strong, weaker where it was not. General availability begins September 15. Operators considering the software will have to weigh the per-robot cost against the connectivity profile of their own facilities, since the product's value, by HelioDock's own design, is tied to a connection staying up.

MoonshotAI: Kimi K3: # HelioDock Launches Night Shift, Software for Managing Warehouse Robots From Afar *The remote-monitoring platform will be generally available September 15 at $90 per robot per month; HelioDock also announced a $24 million Series B led by North Coast Ventures.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely and step in from a distance when a machine runs into trouble. The release takes aim at one of the more thankless routines in automated logistics: the late-night drive to a facility to deal with a single stalled robot. The launch moves HelioDock beyond the machines themselves and into the software used to oversee them, and the company is pairing the expansion with new capital. It also announced Thursday a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. The new round is larger than all of the company's previous funding combined. Chief executive Mara Velez framed the product around the human cost of keeping robots running. "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. The remark points to a gap between automation in theory and automation in practice: robots handle the physical work, but people have remained the backup plan when a machine falters. Night Shift is HelioDock's attempt to make the backup plan remote. Under normal conditions, the software gives operators a remote view of their fleets and the ability to take control of a machine from off site. That model assumes a reliable connection between the operator and the warehouse floor — an assumption the product's beta put to the test across nine companies' facilities. Before Thursday's release, Night Shift spent 14 months in a private beta involving nine logistics companies and a combined 310 robots, an average of roughly 34 machines per company. HelioDock says customers in the program experienced 17 percent fewer unplanned robot stoppages, a company-supplied figure that has not been independently verified. One beta customer offered a more qualified account. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. The caveat reflects how the software behaves when a connection drops. When connectivity fails, a robot returns to its locally configured behavior — it keeps operating under its on-site configuration — and remote control becomes unavailable until service resumes. In practice, an operator watching a fleet from a distant office loses the ability to intervene at exactly the moment the network goes down, which is often the kind of moment the software is meant for. The robot keeps working; the remote link does not. For warehouses in rural areas with inconsistent cellular coverage, Night Shift's usefulness will hinge on local infrastructure as much as on the software itself. Night Shift will cost $90 per robot per month when it becomes generally available on September 15, a structure that scales directly with fleet size. At that rate, a 100-robot deployment would cost an operator $9,000 a month, and a fleet the size of the entire beta program — 310 machines — would run $27,900. HelioDock will not have the category to itself. Relay Robotics and YardOS already sell fleet-management products built for the same basic job, giving warehouse operators established alternatives for overseeing robots from a distance — and for handling a stalled machine without the drive across town.

Anthropic: Claude Fable 5.1: # Foundry Hall to Close After Losing Its Largest Backer The lights at Foundry Hall will go out on October 31. The nonprofit startup accelerator, which opened in Milwaukee in 2018, announced this week that it will shut down at the end of the month after its primary corporate sponsor, Blank River Bank, declined to renew its support. The bank had supplied 38 percent of the organization's annual budget. Foundry Hall has enough cash to keep the doors open through October. It does not have enough to run another full program, and its leadership decided not to try to limp partway through one. "We kept waiting for the replacement check that never came," said executive director Camille Ross. Ross said the organization spent months approaching other sponsors after learning the bank would not renew. None stepped forward at a scale that could cover the gap in time. The board concluded that a planned closing was more responsible than an abrupt one midway through a cohort. In seven years, Foundry Hall supported 212 startups. Those companies now employ 1,140 people, a figure that has become the organization's shorthand for what it meant to the city. The accelerator offered mentorship, introductions to investors, and, perhaps most practically, a place to work that was not a kitchen table or a borrowed conference room. Dev Malik, who founded his company out of the program, remembers the difference it made. "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 closure will cost six employees their jobs. Twelve startups currently working out of the building must find somewhere else to go by the end of October. A spokesperson for Blank River Bank said the decision came from a shift in the bank's community investment priorities and was not a judgment on how Foundry Hall had performed. The bank did not elaborate on what those new priorities are. City officials have begun looking for landing places for the displaced companies. Deputy mayor Susan Cole said the city is in conversations with building owners and local universities about available space. No arrangements have been finalized, and the timeline is tight. Ross has not said what she will do next. For now, she is focused on the calendar. There are founders to notify, files to archive, and a lease to wind down. The organization is compiling records of each company it worked with so that the history does not disappear along with the office. On a recent afternoon, the workspace looked much as it always has. Whiteboards still carried half-finished diagrams. Someone's coffee cup sat cooling by a window. The twelve companies still there are working, because that is what they came to do, and because the deadline for leaving has not quite arrived. It will, at the end of the month. After that, 1,140 jobs will remain in Milwaukee. The building that helped create them will not.

MoonshotAI: Kimi K3: # Foundry Hall to Close at Month's End, Seven Years After It Opened *The Milwaukee accelerator backed 212 startups. It could not replace the one sponsor it lost.* MILWAUKEE — Foundry Hall, the nonprofit startup accelerator that has anchored this city's early-stage scene since 2018, will close October 31, ending a seven-year run as one of the few dedicated homes for new companies here. In a small startup scene, it had become the default first address — the place a founder went before there was a product, a payroll, or a reason for anyone to return a phone call. It offered the unglamorous infrastructure of beginnings: desks, introductions, and a room where an investor would agree to take a meeting. The legacy it leaves is not small. In seven years, 212 startups came through its programs; together they now employ 1,140 people, a workforce the size of a midsize factory. Twelve of those companies still work in the building. They have until the end of the month to go. The closure was less a collapse than a subtraction. Blank River Bank, Foundry Hall's largest corporate sponsor, supplied 38 percent of its annual budget — until it declined to renew. A bank spokesperson said the decision reflects a change in community investment priorities and is unrelated to the accelerator's performance. Whatever the reason, the hole was the same size. Losing one sponsor might have been survivable. Losing the sponsor that covered more than a third of the budget was not. Executive director Camille Ross spent months trying to fill it. Smaller donors gave, she said, but no one writes a check the size of a bank's. "We kept waiting for the replacement check that never came," she said. What remains is enough to operate through October — enough to wind down with care, not enough to fund another full program. The plan, such as it is, is to keep the doors open and the lights on until the money runs out. Six employees will lose their jobs when it ends. For the founders who passed through, the loss is harder to itemize. Dev Malik remembers where his company began. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." City officials are trying to cushion the landing. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the twelve displaced companies, though nothing has been announced, and it is unclear whether any arrangement will be ready by November. The 212 companies that already passed through will keep going; their 1,140 jobs remain. What Milwaukee loses is the next 212 — and the room where their first meetings would have happened. After October 31, the city's newest founders may find themselves pitching once more from the back tables of coffee shops.