Head to head: Anthropic: Claude Fable 5.2 vs MoonshotAI: Kimi K3

Anthropic: Claude Fable 5.2: # Arcwell launches Relay, an AI assistant for support agents, at $65 a seat Arcwell has launched Relay, an AI platform that reads customer support conversations and suggests answers to the agents handling them. The company's pitch is that the information support teams need is usually already somewhere inside their own organizations, and that finding it takes too long. "Support agents spend too much time searching for information that their companies already have," said chief executive Priya Shah. Relay is meant to shorten that search. Alongside suggesting answers, the platform can identify customers who may cancel and create summaries. It costs $65 per agent per month. Arcwell tested the platform with 12 companies over six months and says agents using it closed tickets 14% faster. That figure comes from the company itself, and it speaks to only one of Relay's three functions. The second function, spotting a customer on the way out, is harder to judge. Arcwell did not release customer-retention results from the test, leaving open whether flagging an at-risk customer made any difference to whether that customer stayed. Relay also asks a good deal of the businesses that adopt it. The platform requires access to support transcripts and internal help documents, which gives it a view of both what customers tell a company and what the company tells its own staff. Arcwell says customer data is encrypted and is not used to train shared models. Companies can also decide how long their data is kept, choosing between a 30-day and a 90-day retention period. Those assurances, like the speed figure, are the company's own. For support managers weighing the product, one side of the calculation is clear: $65 a month for each agent, against a claimed 14% gain in ticket-closing speed. What Relay does for customer retention is a question that Arcwell's released test results don't answer.

MoonshotAI: Kimi K3: # Arcwell Launches Relay, an AI Assistant for Customer Support Agents Arcwell has launched Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to the agents handling them. The company says Relay can also identify customers who appear likely to cancel and distill conversations into summaries. "Support agents spend too much time searching for information that their companies already have," said Arcwell CEO Priya Shah. Relay is built around that frustration. Rather than leaving agents to hunt through internal documents for answers, the platform reads each exchange and proposes responses drawn from material the company already has on file. When a conversation points to a customer who might walk away, Relay flags the risk. The pitch comes with early numbers. Twelve companies tested Relay over six months, and agents using the platform closed support tickets 14% faster, according to Arcwell. That figure is self-reported, and the company did not release customer-retention results from the pilot, so it remains an open question whether Relay's warnings about at-risk customers translate into accounts actually saved. Relay costs $65 per agent per month. Using it means granting the platform access to support transcripts and internal help documents — the raw material Relay needs to generate its suggestions, and the kind of data that can make security teams nervous. Arcwell's answer to that scrutiny comes in three parts: customer data is encrypted, it is not used to train shared models, and companies can choose how long their information is stored, with retention options of 30 or 90 days. Together, those policies are meant to assure buyers that their support conversations stay private and under their control. For now, Arcwell's clearest evidence is speed: a measurable gain in how quickly agents resolve tickets. Whether Relay also delivers on its quieter promise — spotting unhappy customers in time to keep them — is a claim the company has yet to back with data. For a product that sells retention as a core feature, that absence is conspicuous.

Anthropic: Claude Fable 5.2: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 User Sessions *The company took the page down 19 minutes after a researcher's warning. Whether anyone else saw it during its four hours online is still under investigation.* Cloudnote left a debugging page containing excerpts of customer prompts accessible without authentication for four hours on Monday, the company said. It removed the page 19 minutes after a researcher alerted it to the problem. Logs show the page held excerpts from 1,842 user sessions, each up to 200 characters of a prompt. Cloudnote has 3.2 million registered users, but the logs count sessions, not people, so the number of individuals affected is not yet clear. Full documents, passwords and payment information were stored elsewhere and were not on the page, the company said. The exposure matters because of what a prompt can hold. Two hundred characters is roughly a sentence or two, enough room for a name, an address or a detail a user would not want a stranger to read. Whether any of the exposed excerpts held such details is not known. The larger open question is whether anyone saw them. Cloudnote says it has found no evidence that anyone besides the researcher accessed the page. But its investigation is not complete, and no evidence so far is not the same as a final finding.

MoonshotAI: Kimi K3: # Cloudnote Exposed Customer Prompt Excerpts on an Unauthenticated Page for Four Hours *A debugging page held prompt fragments from 1,842 user sessions. Cloudnote says it has found no evidence that anyone besides the researcher who flagged the issue viewed it — but its investigation is not finished.* Cloudnote, a note-taking app with 3.2 million registered users, left excerpts of customer prompts visible on a debugging page that required no login for four hours on Monday, the company said. Company logs show the page contained excerpts from 1,842 user sessions, with each excerpt capturing up to 200 characters from a prompt. Cloudnote said it removed the page 19 minutes after a researcher notified it of the exposure. What those excerpts contain is not yet known. Prompts can include whatever a user types into an app — from meeting notes to personal details — so even short fragments can carry sensitive information. Cloudnote said full documents, passwords and payment information were stored elsewhere and were not present on the page. The company said it has found no evidence that anyone besides the researcher accessed the page during the four-hour window, but it cautioned that its investigation is still underway. It has not said how the page became publicly accessible, or whether it plans to notify the users whose sessions were logged. The exposure is modest in scale compared with major data breaches — the logged sessions represent a small fraction of Cloudnote's user base — but it arrives as companies face mounting scrutiny over how they handle customer prompts, including on internal engineering tools that were never meant to be public.

Anthropic: Claude Fable 5.2: # QueueZero's Proxy meeting assistant assigns its CEO a task to "stop talking" during launch demo *The AI tool, which enters public beta next week, treated an audience member's remark as an instruction, giving an unscripted look at how it handles automatic task creation.* QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend a meeting, produce notes, and assign follow-up work to software agents. The launch demonstration also showed how the product decides what counts as an instruction. During the 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 assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task, and deleted the first. The demonstration then continued, with Wu still talking. ## What the assistant got right and wrong The customer's comment contained a real request, and Proxy captured it. The pricing sheet was asked for, logged as a task, and delivered. The assistant also extracted the action item from an offhand sentence rather than a formal directive, which is the kind of work it is built to do. The first task is the harder case. Meeting conversation is full of asides, jokes, and figures of speech, and a system that turns spoken remarks into assigned work has to separate those from real instructions. Here, Proxy treated both halves of the sentence as equally actionable. The consequences were minor, since the task was visible to its assignee and deleted in seconds. But the episode illustrates a question buyers of meeting assistants will likely weigh: how much should such a tool do on its own, and how much should wait for a person's approval? ## Confirmation rules and admin controls QueueZero says Proxy draws that line between internal and external actions: - **External messages:** Proxy requires confirmation before one is sent. - **Internal tasks:** These can be created automatically. - **Admin controls:** Administrators can disable automatic task creation. The tasks assigned to Wu fell into the internal category, consistent with the company's description of how the product works. Organizations that would rather review every task before it reaches an employee's queue can turn the automatic behavior off. ## Pricing and availability Proxy costs $30 per user per month and enters public beta next week. The beta will put the assistant in meetings less controlled than a launch event, where ambiguous remarks are routine. How often automatically created tasks need to be deleted, and how many administrators choose to disable the feature, will indicate whether QueueZero's default settings suit how its customers work.

MoonshotAI: Kimi K3: # QueueZero's AI Meeting Assistant Takes Audience Comment Literally, Assigns CEO Two Tasks QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant that can attend calls, produce notes and assign follow-up work to software agents. The product's first public demonstration also became an unscripted lesson in how literally such tools can interpret human speech. During the launch presentation, 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. The assistant created two tasks and assigned both to QueueZero chief executive Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed when the tasks appeared. He completed the second — the pricing sheet was sent — and deleted the first. The demonstration then continued as planned. Proxy, for its part, did not editorialize. The exchange offered the audience an unusually candid look at the central tension in so-called agentic AI: how much authority software should have to act on ambiguous, overheard speech. An assistant that converts offhand remarks into assignments can save time, but it can also misread intent, and the difference between the two is not always obvious in the moment. QueueZero says it has drawn that boundary deliberately. According to the company, Proxy requires user confirmation before any external message is sent, so a stray comment made during a meeting cannot trigger an email to a client or partner. Internal tasks, by contrast, can be created automatically — the behavior on display Wednesday. Administrators who prefer tighter control can disable automatic task creation entirely. The company is entering a crowded market. Meeting transcription and summarization tools have proliferated over the past two years, and several large platform vendors have folded similar capabilities into their productivity suites. QueueZero is wagering that Proxy's ability to route follow-up work to software agents, rather than simply recap what was said, will distinguish it from competitors that stop at notes. Enterprise buyers, meanwhile, will have to decide how much automation they actually want. A task created by mistake inside a company is mostly harmless; a misinterpreted request that reaches a customer is not. That distinction is likely to make Proxy's external-message confirmation requirement the first feature IT administrators evaluate during the beta. Proxy is priced at $30 per user per month and enters public beta next week. QueueZero said administrators will be able to tune the assistant's automation settings from the first day of the beta, including turning automatic task creation off.

Anthropic: Claude Fable 5.2: # HelioDock launches remote monitoring for its warehouse robots, raises $24 million *Night Shift arrives after a 14-month private beta. One tester says it worked well until rural LTE dropped out.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely, and announced a $24 million Series B round led by North Coast Ventures. The two announcements push HelioDock into a software category where Relay Robotics and YardOS already sell fleet-management products. Night Shift's premise is that a robot problem should not require someone to travel to the warehouse. By the company's own account of how the product works, and that of a customer who tested it, the premise holds only while the facility stays online. Chief executive Mara Velez framed the product around the after-hours callout. "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 involving nine logistics companies and 310 robots. HelioDock says those customers experienced 17% fewer unplanned robot stoppages. The figure comes from the company and has not been independently verified. One beta participant offered a more qualified assessment. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. That matches how the software is designed to behave. When connectivity fails, robots return to their locally configured behavior and remote control becomes unavailable. The robots fall back on their on-site settings, but an operator loses the ability to step in from somewhere else, which is the capability Night Shift is sold on. For prospective buyers, that makes network quality part of the purchasing decision. A well-connected site could expect something like the experience Pike described in Fulcrum's stronger facilities. A rural warehouse that depends on LTE would have remote control only while the link holds. If a robot stops during an outage, the manager is back to the situation Velez described. The fallback also bears on HelioDock's headline number. The 17% reduction in unplanned stoppages is a single figure drawn from a beta in which, by Pike's account, usefulness varied with connectivity. The Series B brings HelioDock's total funding to $39 million. The new round accounts for more than 60% of all the capital the company has raised, and it lands as HelioDock adds a monthly software charge to its robot business. Night Shift will cost $90 per robot per month, or $1,080 per robot per year. At that rate, a fleet of 310 robots, the size of the entire beta pool, would run $27,900 a month. Operators weighing that cost have alternatives in the Relay Robotics and YardOS products. They will also have to take the company's word on performance for now. The beta was private, limited to nine companies, and its results have been reported only by HelioDock. The beta's size puts the stoppage figure in context as well. Spread evenly, 310 robots across nine logistics companies comes to roughly 34 robots per customer, and at least one of those customers was running the software in locations where LTE coverage was unreliable. Night Shift becomes generally available Sept. 15.

MoonshotAI: Kimi K3: # HelioDock Launches Remote Oversight for Warehouse Robots, Raises $24 Million *Night Shift will cost $90 per robot per month when it reaches general availability September 15, following a 14-month beta that the company says cut unplanned robot stoppages by 17 percent — a figure that has not been independently verified.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor and control robot fleets remotely, and announced a $24 million Series B round led by North Coast Ventures. The round brings the company's total funding to $39 million — meaning the new money accounts for more than half of everything HelioDock has raised to date. Together, the product debut and the funding reposition the company. HelioDock has made its name selling robots; now it wants to sell the software layer that runs on top of them, too, billed as a subscription attached to every machine a customer operates. A subscription business also gives HelioDock a continuing relationship with customers after the hardware sale. The launch also moves HelioDock into more direct competition with Relay Robotics and YardOS, both of which already offer fleet-management products to warehouse operators. Night Shift is pitched squarely at the hours automated warehouses keep and their managers would rather not — an idea baked into the product's name. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," chief executive Mara Velez said. In the company's framing, that late-night drive is the problem the software removes. When a machine stalls or meets something it cannot handle — a misplaced pallet, in Velez's example — an operator can see what is happening and respond from off site instead of discovering the problem in person. HelioDock arrives at launch with a long test run behind it. Night Shift spent 14 months in a private beta involving nine logistics companies operating a combined 310 robots, an average of roughly 34 machines each. According to the company, customers in the beta saw 17 percent fewer unplanned robot stoppages — the headline figure in HelioDock's pitch. The result comes from HelioDock's own reporting and has not been independently verified, so the number should be read as the company's account of its product rather than an audited one. The length of the trial does mean HelioDock is launching with data from live warehouse operations rather than lab conditions. The beta also mapped the boundary of what remote oversight can do. Night Shift depends on a facility's network connection, and 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 a connection drops, the robots return to their locally configured behavior and remote control becomes unavailable. An outage, in other words, does not stop the machines — it stops the oversight, leaving robots to run as they were configured on site until service is restored. For warehouses in areas with spotty coverage, that distinction sits at the center of the buying decision. HelioDock is not entering an empty field. Relay Robotics and YardOS both sell competing fleet-management products, so Night Shift will have to win customers against software that is already on the market rather than define a category of its own. Night Shift will cost $90 per robot per month when it becomes generally available on September 15 — about $1,080 per robot annually, with the per-machine price scaling directly with fleet size. For an operation running roughly 34 robots, the average across the beta, that comes to about $3,060 a month.

Anthropic: Claude Fable 5.2: # 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 5.2: **6** Every non-blank line in the draft breaks a rule: 1. `#Flood software moves upstream` has no space after `#`, so it is not in `# Headline` form. 2. `The company is expanding its pilot.` is the dek but is not italicized. 3. `##Details` has no space after `##`, so it is not in `## Heading` form. 4. `[the announcement](www.example.com/news)` is not an absolute HTTPS URL. 5. `* The first pilot begins Monday.` begins with an asterisk instead of a hyphen. 6. `<div>The second pilot begins Friday.</div>` is raw HTML.

MoonshotAI: Kimi K3: 6

Anthropic: Claude Fable 5.2: # Startup Spotlight: LedgerLark Wants AI Agents to Untangle Post-Trade Errors A copied account number is a small mistake. Mina Batra once saw one occupy a six-person team for three days as they worked to resolve the trade mismatch it caused. The episode prompted her to explore whether software could take on more of the tedious work that follows a trade. The result is LedgerLark, which builds AI agents for post-trade operations at small investment firms. The company has five paying customers and a $3.5 million seed round led by Motive Row. It has also set itself a hard problem: selling new technology into a cautious industry that already has reconciliation tools. ## The insight The founders know the problem from two sides. Batra spent seven years in operations at a regional brokerage, close to the daily work of chasing exceptions. Her co-founder, Owen Carr, built reconciliation systems at a large bank. Their premise is that much of the effort in resolving a mismatch goes into finding where records disagree and working out why, rather than into the fix itself. If software can do that tracing reliably, operations staff could spend their time reviewing proposed fixes instead of hunting for errors. Whether it can do so reliably enough is the question the company has to answer. ## What the product does LedgerLark's agents compare three sets of records: trade confirmations, custodian records and a firm's internal ledgers. When those sources disagree, the agents propose a correction, which a person must approve before anything changes. The product does not execute trades or move money. That boundary limits the damage a wrong suggestion could cause, and it may make the product easier for compliance teams to accept. It also means the product's value depends on the quality of its proposals. If staff have to recheck every suggestion from scratch, the time savings shrink. ## Early traction Five firms are paying for the product, with annual contracts ranging from $36,000 to $110,000. For a seed-stage company, that is a meaningful signal that some buyers see enough value to commit budget. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. That figure deserves both interest and caution. It comes from the company, describes a single customer and has not been independently verified. It is also unclear over what period the drop occurred or what else may have changed at the firm. Results across the other four customers would give a fuller picture. ## Competition and open questions LedgerLark is not entering an empty field. Established reconciliation vendors already sell to financial firms, and banks often rely on software built in-house, the kind of system Carr once worked on. Incumbents have existing customer relationships and long track records, and they could add similar AI features to their own products. The sales process is another hurdle. Financial institutions may require lengthy security and procurement reviews before adopting new software, particularly software that touches trade and account data. For a young company with limited funding, long sales cycles can strain finances, even when buyers are interested. Other questions remain open: - **Accuracy:** How often do human reviewers reject the agents' proposed corrections? - **Generalization:** Do results hold across firms with different custodians, systems and trade volumes? - **Durability:** Will the first five customers renew, and will larger contracts follow? ## What to watch LedgerLark's founders have identified a real and costly problem, and they have experience on both the operational and technical sides of it. Paying customers suggest the idea has some pull. The next stage will test whether early results repeat across more firms and whether the company can get through procurement processes without running short of cash. It will also show whether a narrow, human-in-the-loop product can hold its ground against vendors that already have customers' trust.

MoonshotAI: Kimi K3: # Startup Spotlight: LedgerLark Takes Aim at the Back Office's Error Pile *The startup's AI agents reconcile post-trade records for small investment firms — and leave the final decision to humans.* The trade that pushed Mina Batra toward founding a company was not a complicated one. An account number, copied incorrectly from one system into another, set off a mismatch that kept a six-person operations team at her regional brokerage occupied for three days. Batra had spent seven years in operations there, and the episode crystallized an observation she'd been circling for a while: much of post-trade work is comparison, not judgment — matching one record against another and finding where they diverge. Yet at small firms, that comparison is still done largely by hand, by people who are simultaneously settling trades and fielding calls. She teamed up with Owen Carr, who had built reconciliation systems at a large bank and understood both what that software does well and where it strains. The result is LedgerLark, which builds AI agents for post-trade operations at small investment firms — the regional brokerages, advisers, and boutique funds that can't justify a big-bank operations stack. ## What the agents actually do LedgerLark's agents ingest trade confirmations, custodian records, and a firm's internal ledgers, then compare them. When the records disagree, an agent drafts a proposed correction — match against this confirmation, adjust that field — and queues it for a human to approve or reject. The boundaries are deliberate. The product does not execute trades and does not move money. That limits the damage an errant agent could cause and may simplify compliance conversations, though it won't end them: any system touching client data at a financial institution should expect lengthy security and procurement reviews. ## Early, and narrow, traction LedgerLark says it has five paying customers, with annual contracts ranging from $36,000 to $110,000. That implies a business still measured in the hundreds of thousands of dollars of annual revenue — real, but early. The company's headline metric comes with caveats. LedgerLark says one customer cut its unresolved daily exceptions from 70 to 18 after deploying the agents. The figure is company-reported, and a single customer's experience says little about typical results across firms with different volumes and systems. The company recently raised a $3.5 million seed round led by Motive Row. For a startup selling to regulated buyers, the capital's most obvious use is patience — surviving sales cycles that can stretch for quarters while security teams do their work. ## The crowded middle Reconciliation is not an empty market. Established vendors have sold reconciliation software for years, and large institutions often run internal systems of the kind Carr once worked on. LedgerLark's bet is that smaller firms are underserved by both: vendor tools can be heavy to implement, and in-house builds are out of reach. Two questions hang over the company. First, will operations teams at small firms trust agent-proposed corrections enough to rework processes built around spreadsheets and email — even with a human approving every change? Second, can LedgerLark build a durable position before incumbents add similar agent features to products already embedded in back offices? There is a reasonable case in its favor. The founders' combined experience maps closely onto the problem, the product's scope is conservative in a way regulated buyers tend to appreciate, and five contracts suggest at least some firms find the pitch persuasive. Batra's founding insight is modest, and that may be its strength: the error pile is unglamorous, expensive, and real. Whether an agent can shrink it faster than competitors can copy the idea remains an open question. --- **LedgerLark at a glance** - **Product:** AI agents for post-trade reconciliation - **Customers:** 5 paying firms - **Pricing:** $36,000–$110,000 per year - **Funding:** $3.5M seed, led by Motive Row

Anthropic: Claude Fable 5.2: **C** It is the only option that sticks to the stated facts: the product pulled, the day, the 4,860 accounts, the public endpoint and the 11-hour exposure. - **A** is wrong because message contents were not sent, and "massive privacy disaster" is unsupported. - **B** is vague and promotional, and leaves out the exposure entirely. - **D** is false on two counts: only workspace titles were exposed, and there is no evidence of malicious access. One caveat: C's "after discovering" implies the exposure caused the withdrawal, which the facts suggest but do not state outright.

MoonshotAI: Kimi K3: C

Anthropic: Claude Fable 5.2: # KiteBridge error sends $2.8 million in duplicate reimbursements, charged to employers A software failure at expense-management company KiteBridge generated 13,400 duplicate reimbursements worth $2.8 million on Tuesday, drawing the money from 91 corporate customers before the company shut off automated payments. By Wednesday evening, roughly $500,000 had yet to be recovered. The cost fell on employers rather than workers. No employees had money withdrawn from their accounts. The duplicate payments were funded by the businesses that use KiteBridge to repay staff for work expenses. ## A retry mistaken for a new request The trouble began shortly after KiteBridge deployed version 4.2 of its software at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m., 27 minutes later. The cause was a timeout at KiteBridge's payment processor, meaning the processor did not respond to a request within the expected window. Payment software usually handles a timeout by sending the request again, known as a retry. The risk is that the original request may have gone through, so a well-behaved system must recognize the retry as a repeat of the same payment rather than a fresh one. KiteBridge's fallback system did not do that. It treated retries as new reimbursement requests, and each one produced another payment. The duplicates averaged about $209 apiece. KiteBridge suspended automated payments at 10:03 a.m., 51 minutes after the first complaint and 78 minutes after the new version went live. ## A safeguard announced in April The failure is awkward for KiteBridge because the company had advertised a defense against this type of error. An April changelog said version 4.0 had introduced "exactly-once payment protection." The term describes a guarantee that a payment will be processed a single time, however many times the underlying request is sent. It is not yet clear why that protection did not stop Tuesday's duplicates, or what role the version 4.2 release played in the breakdown. ## Disclosure trailed the shutdown KiteBridge's public account of the incident also lagged behind events. Its incident page initially described the problem as delayed reimbursements, which suggested payments were arriving late rather than being sent twice. The page did not disclose the duplicate payments until six hours after automation was suspended, or around 4 p.m. Tuesday. Until then, a customer relying on the page had no indication from it that company funds had been paid out in error. ## Most of the money is back By Wednesday evening, $2.3 million of the $2.8 million had been recovered, about 82 percent of the total. The remainder was still pending. Chief executive Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report. ## Timeline | Time | Event | |---|---| | Tuesday, 8:45 a.m. | KiteBridge deploys version 4.2 | | Tuesday, 9:12 a.m. | First customer complaint arrives | | Tuesday, 10:03 a.m. | Automated payments suspended | | Tuesday, around 4 p.m. | Incident page first discloses duplicate payments | | Wednesday evening | $2.3 million recovered, with the rest pending |

MoonshotAI: Kimi K3: # KiteBridge Software Bug Triggers $2.8 Million in Duplicate Reimbursements *A payment-processor timeout turned routine retries into 13,400 unwanted payouts — and 91 corporate customers are still waiting for some of the money back.* A routine software update at expense-management company KiteBridge went badly wrong Tuesday morning, generating 13,400 duplicate reimbursements worth $2.8 million and leaving 91 corporate customers to clean up the mess. KiteBridge, whose software companies use to process employee expense claims, said no employees had money withdrawn from their accounts — the duplicate payments were funded by employers. By Wednesday evening, the company said it had recovered $2.3 million, with the remaining $500,000 still pending. ## An 80-minute failure The incident began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its platform. The first customer complaint arrived within half an hour, and by 10:03 a.m. the company had suspended automated payments altogether. In the roughly 80 minutes between deployment and shutdown, the duplicate payments were already flowing. According to KiteBridge, the trigger was a timeout at its payment processor — the outside service that actually moves money between accounts. When a payment request times out, systems like KiteBridge's are built to retry rather than drop the transaction. But the company's fallback system failed to recognize those retries as repeats. Instead, it treated each retry as an entirely new reimbursement request — and paid it again. ## A protection that failed to protect The bug is especially awkward because of what KiteBridge had promised customers. An April changelog for version 4.0 announced "exactly-once payment protection" — industry shorthand for safeguards that guarantee a transaction is processed a single time, even when networks fail or requests are sent twice. Achieving true exactly-once behavior is a notoriously difficult problem in payment engineering, and Tuesday's incident suggests KiteBridge's safeguard did not hold up in a real-world outage. ## Disclosure lagged The company's public communications may draw as much scrutiny as the bug itself. KiteBridge's incident page initially described the problem as delayed reimbursements — a characterization that substantially understated what had happened. The page did not disclose the duplicate payments until roughly six hours after the company had suspended its payment automation, leaving customers without a full official account of the incident for much of the day. ## Apology and a promised report CEO Lena Ortiz apologized for the incident and said KiteBridge will publish a technical report explaining what went wrong. The company has not said whether it will compensate customers for costs stemming from the error, or how long recovery of the outstanding funds will take. For the affected employers, the immediate concerns are cash flow and accounting cleanup. But the episode raises a broader question for the fintech industry: as companies hand more authority to automated payment systems, a single flawed deployment can move millions of dollars in minutes. The thoroughness of KiteBridge's promised report — and the fixes that follow — will determine how much customer trust the company can rebuild.

Anthropic: Claude Fable 5.2: # Raincheck Raises $8 Million for Flood-Risk Software for Cities *The Austin company, which helps local governments decide where to send maintenance crews before storms, plans to hire eight engineers and expand into Florida.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in a seed round announced today. The company [maps drainage systems and combines those maps with rainfall forecasts](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews before a storm. Basin Ventures led the round, and existing investor Long Mile also participated. Raincheck has raised $10.5 million in total. The company says the new funding will help it hire eight engineers and expand into Florida. Ana Flores, Raincheck's CEO, said: > Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives. Raincheck currently works with seven cities in Texas and Louisiana. 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.

MoonshotAI: Kimi K3: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin startup's software pairs drainage-system maps with rainfall forecasts, helping local governments decide where to send maintenance crews before a storm.* Raincheck, a company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The Austin-based 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. Basin Ventures led the round, and existing investor Long Mile also participated. Raincheck has now raised $10.5 million in total. CEO Ana Flores said: > Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives. Raincheck outlines its approach in [its published forecasting methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana, though it declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations; emergency managers retain control over deployments and public warnings. ## The 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 5.2: # Foundry Hall to Close After Losing Its Largest Sponsor *The Milwaukee startup accelerator, open since 2018, will shut down October 31.* Foundry Hall, the Milwaukee nonprofit that helped more than 200 startups get off the ground, will close October 31 after its largest corporate sponsor declined to renew its support. Blank River Bank had supplied 38% of the accelerator's annual budget. Without that money, Foundry Hall can pay its way through October but cannot afford to run another full program. The loss of a single sponsor was enough to end it. Executive director Camille Ross said the organization held out for new funding to close the gap. "We kept waiting for the replacement check that never came," she said. Six employees will lose their jobs. Twelve startups that now work out of the building will have to leave. ## A record that wasn't in question Since opening in 2018, Foundry Hall has supported 212 startups. Together, those companies employ 1,140 people. Those figures are the simplest measure of what the accelerator did in its years of operation. The bank does not dispute that record. A spokesperson said the decision had nothing to do with Foundry Hall's performance and instead reflected "a change in our community investment priorities." For the accelerator, the explanation does not change the arithmetic. More than a third of its budget is gone, and what remains runs out with the month. An organization whose performance, by its sponsor's own account, was not the problem is closing anyway. ## Where the companies go The most immediate question is what happens to the 12 current tenants. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about finding space for them. Those conversations are about the companies. The six people who work at Foundry Hall are in a different position: their jobs end when it closes. And some of what the accelerator offered may be harder to replace than desks. Dev Malik, a founder, held his first investor meeting in the building. He remembers what came before it. "Before Foundry Hall, I was pitching from the back table of a coffee shop," he said. Foundry Hall will stay open through the end of October. The companies still inside have until then to find somewhere else to work. After that, there will be no next program and no new group of founders. What remains are the companies it already supported and the 1,140 people who work for them.

MoonshotAI: Kimi K3: # Foundry Hall to Close After Its Biggest Sponsor Walks Away **MILWAUKEE —** Foundry Hall, the nonprofit startup accelerator that has anchored the city's early-stage tech community since 2018, will close on October 31. Its largest corporate sponsor declined to renew its funding, and no replacement came forward. "We kept waiting for the replacement check that never came," said Camille Ross, the organization's executive director. Foundry Hall has enough money to operate through October, she said, but not enough to run another full program. The closure ends a seven-year run. Over that time, Foundry Hall supported 212 startups, which collectively employ 1,140 people. Not all of those companies survived — startups rarely do — but the jobs figure became the accelerator's best argument for itself: evidence that founders could build in Milwaukee without first leaving it. Dev Malik was one of those founders. "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 vulnerability, in the end, was structural. Blank River Bank, the accelerator's largest corporate sponsor, supplied 38 percent of its annual budget. When the bank declined to renew its sponsorship, the gap proved too large to close. The math stopped working. In a written statement, a bank spokesperson said the decision reflected "a change in our community investment priorities" and was unrelated to Foundry Hall's performance. The statement made no mention of what the sponsorship had paid for. The consequences arrive quickly. Six employees will lose their jobs. The twelve startups currently working out of Foundry Hall's space must find somewhere else to be by the end of the month — a particular sting for companies that joined a program, not just a lease. Deputy Mayor Susan Cole said the city is speaking with building owners and local universities about space for the displaced companies. Desks, though, were only part of the offering. The introductions, the mentorship, the simple fact of a room where investors agreed to take meetings — those do not transfer with a set of keys. For the remaining weeks, the staff's work continues: preparing founders for meetings that will now happen somewhere else. What comes after — for the six employees, the twelve companies, and the belief that a mid-sized city can grow its own startups — is unresolved. On November 1, for the first time in seven years, there will be no program to apply to.

Anthropic: Claude Fable 5.2: # In a Support-Ticket Benchmark, the Most Expensive Model Bought No Measurable Edge A new benchmark of five language models on customer-support tickets found that the two top performers were statistically indistinguishable on quality. One of them, Vale, cost about 61% less to run than the other, North, and answered 18% faster. For teams choosing a model to draft support replies, the priciest option showed no advantage the data can support. ## How the test worked Researchers drew 200 anonymized tickets from a single software company and had each model produce one proposed answer per ticket. Three experienced support managers graded the answers without knowing which model wrote them. Each score combined factual correctness, usefulness and tone. The design has two strengths. Blind grading keeps brand reputation out of the judgment, and the graders were practitioners who know what a good support reply looks like, not an automated metric. ## A tie at the top | Model | Average 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 | North nominally leads, but the 0.4-point gap between it and Vale was not statistically significant. That does not prove the two are equal, only that this test cannot separate the difference from noise. Declaring North the winner would read more into the numbers than they contain. What does separate the two is cost and speed. North costs roughly 2.6 times as much per ticket, and Vale responds 18% faster. The cost gap needs perspective, though. All of these prices are small in absolute terms, and at a million tickets the difference between North and Vale is about $1,160. For many teams, the speed advantage may matter more than the savings. The benchmark also hints at different strengths. North performed best on billing disputes and Vale on technical troubleshooting. The results do not say how large those leads were, and category subsets of a 200-ticket sample are necessarily small. They are reasons to test on your own ticket mix, not grounds for a routing strategy. ## What the cheaper models give up Below the top pair, scores fall as prices do. Ember trails the leaders by about five points at less than half Vale's cost. Harbor costs roughly one twenty-fourth of what North does and scores about 13 points lower. The results do not say whether these gaps are statistically significant. The large ones, such as Harbor's distance from the leaders, are more likely to reflect real differences than the few points between Ember and Kite. Whether a cheaper model is good enough depends on how it is used. If a human agent reviews every draft, a lower-scoring model may be acceptable. If answers go to customers with little oversight, lost quality costs far more than the cents saved, since every model here is cheap next to a mishandled ticket. ## Where every model struggled The most practical finding may be one that has nothing to do with rankings: all five models performed poorly when a ticket depended on an image attachment. Screenshots of error messages, invoices and settings pages are common in software support. Paying more did not fix this weakness. Any deployment should plan for it, for example by sending image-dependent tickets to human agents. ## What the benchmark cannot tell you Several limits affect how far these results travel: - **Narrow sample.** The tickets were in English and came from one software company. Results may differ in other industries, products and customer bases. - **Untested capabilities.** Researchers did not evaluate agentic actions, tool use or multilingual support. A model that drafts a good reply may not be the best at looking up an account or issuing a refund. - **Single attempts.** Each model answered each ticket once, so the test does not show how consistent a model is across repeated runs. - **Composite scoring.** A single blended score hides whether a model lost points on facts or on tone, and a wrong answer usually costs more than a curt one. - **Estimated costs.** The per-ticket figures are estimates, and real costs will vary with ticket length and pricing changes. ## The bottom line On this evidence, North's higher price buys no measurable gain in answer quality over Vale. Either model, or a cheaper one, could be the right choice depending on ticket mix, review process and tolerance for error. The benchmark supports a shortlist, and a trial on your own tickets, including ones with screenshots, should make the final call.

MoonshotAI: Kimi K3: # The Top Model in This Support Benchmark Isn't the One Worth Paying For As more companies hand first-draft support replies to language models, the practical question has shifted from "which model is smartest?" to "which model is good enough, at a price that works at scale?" A new benchmark of five models on real support work offers an unusually clean answer — and it isn't the model that finished first. Researchers fed 200 anonymized customer-support tickets to five models and had three experienced support managers grade each proposed reply blindly. Scores combined factual correctness, usefulness and tone. Model North posted the highest average at 84.2. Model Vale scored 83.8 — a gap the researchers found was not statistically significant. That tie is the story. Vale costs an estimated $0.74 per 1,000 tickets, about 40% of North's $1.90, and it produced answers 18% faster. On every dimension a buyer can act on, the runner-up is at least as good: equal measured quality, lower cost, quicker responses. The premium attached to the top score buys nothing the graders could reliably detect. ## Where cheaper starts to cost you Below the top two, the trade-offs become real. The difference between North and Vale is statistical noise; the difference between Vale and third-place Ember (79.1) is nearly five points — more than ten times as large. Stepping down to the budget tier is a genuinely different product decision, not just a thriftier version of the same one. The arithmetic compounds at scale, which is what makes the budget tier tempting. A team handling 100,000 tickets a month would spend roughly $190 with North, $74 with Vale and $8 with Harbor, the cheapest model at $0.08 per 1,000 tickets. But Harbor's 71.0 average implies a substantial share of replies that managers judged weak on correctness, usefulness or tone — savings that would show up in the queue, not just the invoice. | 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 | ## Different strengths, one shared blind spot The averages hide some texture. North performed best on billing disputes; Vale led on technical troubleshooting. "Best model," in other words, depends partly on the mix of tickets a company actually receives — though with the overall gap inside the margin of error, even that texture should be weighed cautiously. More striking is what none of the models could do. All five performed poorly on tickets that depended on an image attachment — a screenshot of an error message, a photo of a damaged shipment. For support teams that isn't an edge case but a large slice of real volume. Whatever model wins the text-only comparison, deployment will need a human fallback or a separate pipeline for image-based tickets. ## Read the limitations before the leaderboard Blind grading by three managers protects the scores from brand bias, and the methodology is sound as far as it goes. But it goes a short distance. The composite score blends three distinct qualities into one number, and the researchers did not publish the weighting or how often the graders agreed — both worth knowing before treating any gap as meaningful. The scope is narrower still. All 200 tickets came from a single software company, in English. A model tuned to one company's products, vocabulary and customers may not transfer cleanly to a retailer, a bank or a hardware maker. And the models were asked only to draft answers — not to take actions, use tools, pull account data or process refunds, and not to work in other languages. Support automation increasingly involves exactly those agentic capabilities; this benchmark measures one slice of the job, which is writing a good reply to a written question. Two hundred tickets is also a modest sample, which limits how confidently close scores can be ranked. ## The takeaway For most teams, Vale is the rational default: quality statistically indistinguishable from the leader, at roughly 40% of the cost and with faster turnaround. The budget models earn their place only where volume is extreme and quality tolerances are loose. And nobody should deploy any of these five without a plan for image-based tickets — the one category where every model, cheap or expensive, fell short.