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

MoonshotAI: Kimi K3: # Foundry Hall, Which Backed 212 Milwaukee Startups, Will Close in October MILWAUKEE — Foundry Hall, a nonprofit startup accelerator that has operated here since 2018, will close on October 31, undone not by a failed program but by a vanished sponsor. Over seven years, the accelerator supported 212 startups, which collectively employ 1,140 people. It offered early-stage companies what most of them lack: workspace, mentors, and introductions to investors. Its end came down to a single line in a single budget. Blank River Bank, its largest corporate sponsor, supplied 38 percent of Foundry Hall's annual funding. This year, the bank declined to renew. The bank attributed the decision to "a change in our community investment priorities," adding that it was unrelated to Foundry Hall's performance. Executive director Camille Ross spent the months that followed trying to replace the money — meeting with foundations, other banks, individual donors. Nothing arrived in time. "We kept waiting for the replacement check that never came," she said. Replacing more than a third of an operating budget is not the same as raising a third of one. Large commitments move on their own schedule, and Foundry Hall's schedule ran out. Foundry Hall has enough money to operate through October but not enough to run another full program. Six employees will lose their jobs, and the twelve startups currently using its workspace must be out by the end of the month. For the employees, the news turned a workplace into a countdown; for the companies, a home into a deadline. For Dev Malik, the loss is not a budget line. He was one of the founders the accelerator took in, and he traces his company's beginning to its rooms. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." Deputy Mayor Susan Cole said the city is speaking with building owners and local universities about space for the twelve displaced companies. The aim, she said, is to keep them rooted in Milwaukee even as their home closes. The staff's final weeks will go to the practical work of closing: helping founders relocate, connecting employees with new employers, and shutting down a program that, in another year, would have been recruiting its next cohort. The numbers the accelerator leaves behind are the kind it once taught founders to cite in pitch meetings: 212 companies, 1,140 jobs, seven years. But precision was never the same as permanence. The lights in the building stay on through October. After that, the founders who remain will do what founders are trained to do: find another room, and keep pitching.

Anthropic: Claude Fable 5.1: # Foundry Hall to Close After Losing Its Largest Sponsor ## The Milwaukee accelerator helped launch 212 companies. Its last day is October 31. The lights at Foundry Hall will go off for the last time on October 31, ending a seven-year run in which the nonprofit accelerator helped 212 Milwaukee startups get off the ground. The closure follows a decision by Blank River Bank not to renew the sponsorship that had anchored the organization's finances since it opened in 2018. The bank's contribution covered 38 percent of Foundry Hall's annual budget, a share large enough that no other funder came close to replacing it. Executive director Camille Ross said the board spent months trying to close the gap before concluding it couldn't be done in time. "We kept waiting for the replacement check that never came," she said. The organization has enough cash to operate through the end of October, but not enough to run another full program cohort. Rather than start something it could not finish, the board chose to wind down. Six staff members will lose their jobs. Twelve startups currently working out of Foundry Hall's space will have to find new homes. In a statement, a Blank River Bank spokesperson said the decision reflected a shift in the bank's community investment priorities and was not a judgment on Foundry Hall's work. The bank did not say what those new priorities are or whether any of the money would remain in Milwaukee. By the numbers, the record Foundry Hall leaves behind is substantial. The companies it supported now employ a combined 1,140 people. Many started as little more than an idea and a founder with no office. Dev Malik was one of them. He remembers pitching from the back table of a coffee shop before he joined the program. His first investor meeting happened inside Foundry Hall. "Before Foundry Hall, I was pitching from the back table of a coffee shop," he said. For founders like him, the building offered something harder to quantify than a desk: a place where showing up to work meant being taken seriously. City officials say they are trying to prevent the twelve remaining companies from losing momentum. Deputy mayor Susan Cole said the city is in conversations with building owners and local universities about workspace for the displaced startups, though nothing has been finalized. What the city has not announced is any plan to replace the accelerator itself. Foundry Hall's programming — the mentorship, the investor introductions, the cohort structure that turned isolated founders into a network — ends with the building. Ross has not said what comes next for her or her staff. The remaining weeks will be spent helping current companies relocate, closing out grant obligations, and documenting what the organization learned. The building's lease runs through October. After that, the address on hundreds of early pitch decks and incorporation documents will belong to someone else.

MoonshotAI: Kimi K3: 6

Anthropic: Claude Fable 5.1: **6**

MoonshotAI: Kimi K3: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin startup's software maps drainage systems and pairs them with rainfall forecasts, pointing local governments to the drains, pumps and repairs that need attention before a storm hits.* Raincheck, a company that makes flood-risk software for local governments, has raised $8 million in seed funding, the company announced today. The Austin-based company maps municipal drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. The round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has now raised $10.5 million in total. CEO Ana Flores said the software's value lies in helping cities prioritize: > "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 overview of its 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 only; emergency managers retain control over crew deployments and public warnings. Raincheck's competitors include flood-planning vendors such as HydroMap, as well as municipal engineering contractors. The company says the new funding will help it hire eight engineers and expand into Florida.

Anthropic: Claude Fable 5.1: # Raincheck Raises $8 Million to Help Cities Clear Drains Before Storms Hit *The Austin startup's flood-risk software tells local governments which drains, pumps and repairs to prioritize when rain is on the way.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding to help cities decide where to send maintenance crews before a storm. The company maps municipal drainage systems and combines those maps with rainfall forecasts. Cities use the resulting software to direct crews to the drains, pumps and repair sites most likely to cause trouble when rain arrives. The round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has now raised $10.5 million in total. CEO Ana Flores said the challenge for cities is less about knowing where water collects than about triage: > Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives. Raincheck has published [an explanation of how its forecasts are produced](https://example.com/raincheck-methodology). The company stresses that those forecasts are recommendations; emergency managers retain control over crew deployments and public warnings. Raincheck currently works with seven cities in Texas and Louisiana. It declined to identify them, saying some contracts have not yet been publicly announced. The company competes with other flood-planning vendors, including HydroMap, as well as with municipal engineering contractors. Raincheck says the new funding will allow it to hire eight engineers and expand into Florida.

MoonshotAI: Kimi K3: **C** Option C is the strongest lede because it's accurate, specific, and complete: it states what happened (product pulled), when (Tuesday), and why (workspace titles from 4,860 accounts exposed to a public endpoint for 11 hours)—all without overstating the facts. The others fail: - **A** falsely claims private messages were leaked (only titles were exposed). - **B** buries the real story in PR language and omits the exposure entirely. - **D** falsely claims message contents were visible to "hackers" (contents weren't sent, and there's no evidence of malicious access).

Anthropic: Claude Fable 5.1: C

MoonshotAI: Kimi K3: # LedgerLark Wants AI Agents to Handle Post-Trade Drudgery — With Humans Signing Off *The startup is betting that small investment firms will pay for software that finds trade errors and proposes fixes, but stops short of touching the money.* Post-trade operations rarely attracts attention until something goes wrong. LedgerLark, a startup founded by Mina Batra and Owen Carr, builds AI agents for that unglamorous corner of the market: helping small investment firms find and resolve trade errors. The company's pitch rests on a deliberate limit — its software proposes corrections, but a person has to approve them, and it never executes trades or moves money. ## The insight Batra spent seven years in operations at a regional brokerage. She began exploring the idea for LedgerLark after watching a six-person team spend three days resolving a single trade mismatch that traced back to a copied account number. The episode wasn't unusual so much as representative: much of post-trade work is pattern-following detective work, done by experienced staff whose time is expensive. Carr, her co-founder, built reconciliation systems at a large bank. Between them, the pair have seen the problem from both sides — the operations teams that inherit errors and the engineers asked to automate them away. ## The product LedgerLark's agents compare three sources of record: trade confirmations, custodian records, and a firm's internal ledgers. When the records disagree, the agents draft a proposed correction and route it to a human for approval. The decision to stop there is notable. By declining to execute trades or move money, LedgerLark sidesteps some of the thorniest regulatory and risk questions in financial automation. It also gives compliance teams a simpler story: the software recommends, people decide. Whether that boundary holds as customers push for more automation is an open question. ## Early traction Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. The company says one customer reduced its unresolved daily exceptions from 70 to 18. That figure comes from LedgerLark and hasn't been independently verified, and a single customer result says little about the median experience. Still, five signed contracts at those price points suggest the pain is real enough to budget for. The company has raised a $3.5 million seed round led by Motive Row — enough to fund further product work and, presumably, the slow business of selling to financial firms. ## What's uncertain LedgerLark is not entering an empty field. Established reconciliation vendors have sold into financial institutions for years, and many firms already lean on internal software or tools supplied by their banks. LedgerLark's case depends on its agents being meaningfully better than existing rule-based matching — and on small firms having both the budget and the appetite to switch. Sales may be slow. Financial institutions often require lengthy security and procurement reviews, which can stretch the gap between a promising pilot and a signed contract. Five customers is evidence of a start, not of a repeatable sales motion. There are also questions specific to AI in this setting. If an agent proposes a wrong correction, will a reviewer catch it — especially once approving agent suggestions becomes routine? The human-in-the-loop design reduces risk, but it doesn't eliminate it, and its effectiveness depends on reviewers staying skeptical as volumes grow. For now, LedgerLark is a small company with a focused product and a plausible foothold. The more telling signals will come later: whether early customers renew, whether references survive procurement reviews at larger firms, and whether the approval-first model survives customers' inevitable requests to automate one step further.

Anthropic: Claude Fable 5.1: # Startup Spotlight: LedgerLark Wants to Clean Up the Back Office, One Mismatch at a Time Most of what happens after a trade is executed is invisible to the people who placed it. Confirmations arrive from counterparties, custodians report what they hold, and the firm's own ledger records what it believes it owns. When those three sources disagree, someone has to figure out why. LedgerLark, a young company building AI agents for post-trade operations at small investment firms, is betting that this work is both more common and more tractable than it looks from the outside. ## The problem, as the founders saw it Co-founder Mina Batra spent seven years in operations at a regional brokerage. The moment she describes as the origin of the company is unglamorous: a six-person team spent three days resolving a single trade mismatch, which turned out to be caused by a copied account number. Nobody had made a strategic error. Someone had pasted the wrong digits, and the downstream systems had faithfully propagated them. Her co-founder, Owen Carr, came at the same problem from a different direction. He built reconciliation systems at a large bank, where the scale is greater but the tooling, he says, is also more mature. Smaller firms, by their account, often rely on spreadsheets, email threads, and institutional memory. ## What the product actually does LedgerLark's agents ingest three sources—trade confirmations, custodian records, and the firm's internal ledger—and compare them. When they find a discrepancy, they attempt to identify the likely cause and propose a correction. A human then approves or rejects it. Two boundaries are worth noting. The product does not execute trades, and it does not move money. That design choice narrows the scope of what the software can do, but it also narrows the scope of what can go wrong, which is likely to matter when the company sits in front of a compliance officer. ## Early traction Five firms are currently paying customers, with annual contracts ranging from $36,000 to $110,000. The company says one customer reduced its unresolved daily exceptions from 70 to 18. That figure comes from LedgerLark and reflects a single account; it is encouraging, but it is not yet evidence of a pattern across customers with different workflows and data quality. The company has raised a $3.5 million seed round led by Motive Row. That is enough to build and sell for a while, though probably not enough to wait out long procurement cycles at many institutions simultaneously. ## Where the open questions are The competitive landscape is not empty. Established reconciliation vendors have served this market for years, and larger banks have internal software that does much of the same work. LedgerLark's argument is that these tools are either too expensive or too heavy for firms with a handful of operations staff. Whether incumbents respond by moving downmarket—or by adding similar agent-style features to existing products—is an open question. Sales friction is another. Financial institutions, even small ones, may require lengthy security and procurement reviews before connecting a third-party system to custodian and ledger data. Five customers suggests the company has cleared that process a few times; it does not tell us how long each one took or how repeatable it is. There is also a quieter question about the technology itself. Proposing a correction is only useful if the proposal is right often enough that reviewers trust it and wrong rarely enough that they keep checking. Where that threshold sits, and whether the agents can reach it across varied data formats, is something only more customers and more time will reveal. ## Why it's worth watching LedgerLark is not trying to reinvent how trades happen. It is trying to make a tedious, error-prone step less tedious and less error-prone, while keeping a person in the loop. That is a modest ambition in a sector that often prefers grander ones—and it may be exactly why it is worth paying attention to.

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: # 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: # Software Bug at KiteBridge Triggers $2.8 Million in Duplicate Reimbursements *A routine update to the expense-management platform paid out corporate expenses multiple times over, affecting 91 companies.* A faulty software update at expense-management company KiteBridge unleashed more than 13,000 duplicate reimbursement payments totaling $2.8 million on Tuesday, forcing the firm to shut down its automated payment system and leaving 91 corporate customers working to reclaim their money. The incident began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its platform. The first customer complaint arrived at 9:12 a.m., less than half an hour later. By 10:03 a.m., the company had suspended automated payments entirely — but not before the software had generated 13,400 duplicate reimbursements across its customer base. ## How a timeout turned into $2.8 million The failure traces back to a timeout at a payment processor, the outside service that actually moves money between accounts. When a payment request times out, the sender cannot confirm whether the transaction went through, so the standard response is to try again. KiteBridge's fallback system — the backup logic built to handle exactly this situation — misread those retries as brand-new reimbursement requests, paying some expenses two or more times. Employees were not directly harmed. Because corporate reimbursements are funded by employers, no workers had money withdrawn from their own accounts; the duplicate payments came out of company funds, leaving finance teams at 91 businesses to reconcile the damage. The glitch is especially awkward given KiteBridge's own marketing. An April changelog for version 4.0 touted "exactly-once payment protection" — industry shorthand for guaranteeing that a transaction is processed a single time, even when networks falter and systems retry. Tuesday's breakdown occurred under precisely the conditions that feature was designed to prevent, raising questions about whether the protection failed on its own or was broken by the 4.2 update. ## Disclosure lagged behind the fix KiteBridge's public communication trailed its internal response by hours. The company's incident page initially described the problem only as delayed reimbursements, a characterization that significantly understated what had gone wrong. The page did not acknowledge the duplicate payments until six hours after automated payments were suspended — meaning some customers likely spotted the erroneous transactions in their own accounts before hearing an explanation from KiteBridge. ## Recovery underway By Wednesday evening, the company said it had recovered $2.3 million of the $2.8 million, with the remainder still pending. Recovering duplicate payments typically requires reversing transactions through banks and payment processors, a process that can take days depending on where the funds have settled. CEO Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report explaining what went wrong. That report is likely to face close scrutiny from customers who were promised — in the company's own words — that their payments would be processed exactly once.

Anthropic: Claude Fable 5.1: # Software Update Sends $2.8 Million in Duplicate Reimbursements to Employees at 91 Companies A routine software update at KiteBridge, a maker of expense-management software, triggered 13,400 duplicate reimbursement payments totaling $2.8 million on Tuesday morning. The money left corporate accounts for roughly 80 minutes before the company shut down its automated payment system. By Wednesday evening, KiteBridge said it had recovered $2.3 million. The remaining roughly $500,000 was still pending. Ninety-one corporate customers were affected. Because employers, not employees, fund reimbursements through the platform, no worker saw money pulled from a personal account; companies absorbed the duplicate payouts while KiteBridge works to claw them back. Chief Executive Lena Ortiz apologized and said the company will publish a technical report on the failure. ## How it happened KiteBridge deployed version 4.2 of its software at 8:45 a.m. Tuesday. Expense tools like KiteBridge's sit between employers and their staff: an employee submits a receipt, the employer approves it, and the software instructs a third-party payment processor to move the money. Shortly after the deployment, the connection to that payment processor began timing out, meaning the processor was not responding quickly enough to confirm whether payments had gone through. That alone is a common hiccup, and payment systems are built to handle it by trying again. The problem was in how version 4.2 handled those retries. When the processor failed to respond, KiteBridge's fallback system — the backup logic meant to keep payments moving when the primary path stalls — did not recognize the retried requests as repeats of payments already in progress. It treated each one as a brand-new reimbursement request and sent the money again. The first customer complaint arrived at 9:12 a.m., 27 minutes after the release. KiteBridge suspended automated payments at 10:03 a.m. In between, the system issued 13,400 duplicates. ## A protection that was supposed to exist The failure is notable because KiteBridge had told customers it had solved exactly this problem. An April changelog for version 4.0 said the release introduced "exactly-once payment protection," an industry term for safeguards that ensure a single approved expense produces one payment, no matter how many times the instruction is retried. Such protection typically works by tagging every payment with a unique identifier so that if the same instruction arrives a second time, the system recognizes and discards it. Tuesday's incident suggests that safeguard either did not cover the fallback path or was undermined by changes in version 4.2. KiteBridge has not yet said which. ## Disclosure lagged behind the damage Customers also faced a communication gap. KiteBridge's public incident page initially described the problem as delayed reimbursements — for affected finance teams, the opposite of reality. Companies were not waiting on money that had failed to go out; they had sent too much. The page did not acknowledge the duplicate payments until about 4 p.m. Tuesday, six hours after automated payments were halted. For most of the business day, corporate customers watching unexpected outflows had no official confirmation of what had gone wrong or how much was involved. ## What comes next Ortiz's apology and the promised technical report are the company's first substantive public account of the incident. Customers will be looking for answers on three fronts: why the fallback system bypassed the "exactly-once" safeguard, why the incident page misdescribed the problem for six hours, and when the outstanding $500,000 will be returned. For the 91 companies involved, the immediate financial exposure appears limited and shrinking. The longer-term question is one of trust in a product whose core promise is that it moves money exactly as instructed.

MoonshotAI: Kimi K3: # Arcwell Launches Relay, an AI Assistant for Customer Support Teams Arcwell has launched Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to the agents handling them — a product the company is casting as a significant step for its own business and for an industry where AI continues to transform how support work gets done. Relay does more than draft replies. The platform also watches for customers whose messages suggest they may be preparing to cancel, and it produces summaries of support interactions. Together, the features amount to an argument that the same system reading a conversation can both speed it up and flag what is at stake in it. Arcwell tested Relay with 12 companies over six months and reports that agents using it closed tickets 14 percent faster. The company did not release customer-retention results, leaving open the question of whether the cancellation warnings translate into saved accounts. The speed figure is also Arcwell's own measurement from a relatively small pilot, which makes it a promising data point rather than proof. Relay is priced at $65 per agent per month. Chief Executive Priya Shah framed the platform as a fix for a familiar inefficiency. "Support agents spend too much time searching for information that their companies already have." Delivering on that promise requires deep access: to function, Relay must connect to a company's support transcripts and internal help documents, material most businesses guard closely. Arcwell says the data it handles is encrypted and is not used to train shared models, a commitment likely to matter to buyers weighing the platform against their privacy policies. Customers can also choose a 30-day or 90-day retention period, giving them some control over how long their information stays in Arcwell's systems. Whether or not the pilot numbers hold at scale, Relay opens a new chapter for Arcwell. The company is betting that support teams will trade a measure of data access for faster queues and earlier warnings about churn — and that demonstrated gains will matter more to buyers than broad promises about AI reshaping customer service.

Anthropic: Claude Fable 5.1: # Arcwell's Relay Reads Support Chats and Drafts Replies for Agents Arcwell has launched Relay, a platform that uses artificial intelligence to scan customer support conversations and propose answers to the agents handling them. Beyond suggesting replies, Relay flags customers who appear likely to cancel and generates summaries of support interactions, giving agents a running record without requiring them to write one themselves. The pitch rests on a familiar frustration. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Relay is designed to surface that information in the moment, drawing on a company's own transcripts and help documents rather than sending agents off to hunt for it. Arcwell tested the platform with 12 companies over six months. According to the company, agents using Relay closed tickets 14 percent faster than they otherwise would have. That figure comes from Arcwell itself, and the company has not published results on the other half of its pitch: whether identifying at-risk customers actually helped keep them. Retention numbers from the trial were not released. Relay is priced at $65 per agent per month. The platform's usefulness depends on access to sensitive material. To work, Relay needs a company's support transcripts and its internal help documentation, the same records that contain customer complaints, account details and whatever else people disclose when they need help. Arcwell says that data is encrypted and is not used to train models shared across its customer base, a distinction that matters to companies wary of their support history informing a competitor's tooling. Customers can also choose how long Relay holds their data, with 30-day and 90-day retention options available. Those assurances address the most obvious objections, though companies evaluating the product will likely want to see the details for themselves before handing over years of support logs. For Arcwell, Relay is a bet that the next gains in customer service will come not from replacing agents but from putting better information in front of them faster. The early ticket-closure figure gives the company a number to point to. Whether the cancellation predictions prove as valuable remains, for now, an open question.

MoonshotAI: Kimi K3: # HelioDock Launches Night Shift, Software for Managing Warehouse Robot Fleets From Afar *The autonomous-robot maker says a 14-month beta with nine logistics companies cut unplanned stoppages by 17 percent — a figure that has not been independently verified — and announced $24 million in new funding alongside the launch.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely rather than keeping staff on-site around the clock. The company also announced a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. The product takes aim at one of the least automated parts of warehouse automation: what happens when a robot stops working. Autonomous machines can move goods all night, but when one stalls, a person typically has to be physically present to set it right. For operators running several facilities, that means paying staff to be on-site, or on call, at hours when almost no one else is working. Night Shift is HelioDock's attempt to change that equation, and it moves the company beyond its core hardware business into subscription software aimed squarely at the labor costs of automation's remaining human requirements. Mara Velez, HelioDock's chief executive, frames the problem less as an engineering challenge than as a tax on the people who run the buildings. "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 software gives operators remote visibility into their fleets and, when the network connection holds, remote control of the machines themselves — enough, in theory, to resolve routine problems without sending anyone across town. In practice, that means one employee could watch machines in multiple facilities from a single location. Night Shift arrives after 14 months in a private beta involving nine logistics companies and 310 robots, an average of roughly three dozen machines per deployment. According to HelioDock, customers in the program experienced 17 percent fewer unplanned robot stoppages, the metric at the center of the product's pitch. The figure has not been independently verified, so the results are best treated as an early signal rather than proof of how the software performs at scale. Fewer stoppages, if the numbers hold, would mean fewer of the late-night interventions Velez describes. One beta participant offered a qualified endorsement. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages — an experience that suggests the product's value rises and falls with the quality of a facility's network connection. The limitation is structural. When connectivity fails, robots revert to their locally configured behavior, and remote control becomes unavailable; the machines simply continue on their own settings until the connection returns. For warehouses at the edges of reliable cellular coverage, the product's central feature — oversight from a distance — may be intermittently out of reach. HelioDock also enters a market with established options. Competing fleet-management products are already available from Relay Robotics and YardOS, giving warehouse operators alternatives to weigh as they decide whether remote oversight is worth the added cost — and whether HelioDock's version offers anything rivals do not. Night Shift will be generally available on September 15. HelioDock set pricing at $90 per robot per month, which works out to $9,000 a month for a 100-robot deployment and scales directly with fleet size. At that rate, the 310 robots that participated in the beta would carry a combined list price of roughly $27,900 a month.

Anthropic: Claude Fable 5.1: # HelioDock Launches Remote Monitoring Software for Warehouse Robots, Raises $24 Million *Night Shift, which spent 14 months in private beta, lets operators supervise robot fleets from off-site — as long as the network holds up.* HelioDock, a maker of autonomous warehouse robots, on Thursday released Night Shift, software that allows operators to monitor and manage robot fleets remotely. The company also disclosed a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. The product targets a persistent problem in automated warehouses: a single robot stalled by an unexpected obstacle can hold up a shift until a person arrives to sort it out. HelioDock says customers in its private beta experienced 17% fewer unplanned robot stoppages, a figure that has not been independently verified. Night Shift will cost $90 per robot per month and becomes generally available September 15, entering a market where Relay Robotics and YardOS already sell fleet-management tools. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," CEO Mara Velez said. ## Fourteen months, nine customers, 310 robots Night Shift was tested privately for 14 months with nine logistics companies operating a combined 310 robots. The 17% reduction in stoppages comes from that beta group, according to HelioDock. Because the figure is company-supplied and has not been checked by an outside party, it is best read as HelioDock's own assessment of the trial rather than an established benchmark. The beta did produce at least one candid account from a participant. Aaron Pike of Fulcrum Logistics, one of the nine companies in the trial, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. ## What happens when the connection drops Pike's experience reflects how Night Shift is designed. When connectivity fails, robots return to their locally configured behavior and remote control becomes unavailable. In practice, a robot that would have stopped and waited for a human before Night Shift will still stop and wait if the connection drops; the difference is that no one can step in from off-site until service returns. That makes the software's central promise conditional. In a warehouse with reliable coverage, an operator can resolve a stuck robot without leaving home. In a facility served by patchy rural LTE, the two-in-the-morning drive Velez described remains a possibility whenever the network goes down — which may be precisely when a robot is most likely to need help and least likely to get it remotely. ## Pricing and competition At $90 per robot per month, a facility running 50 robots would pay $4,500 a month, or $54,000 a year. Whether that outlay pays for itself depends on how often a given operation experiences stoppages and what each one costs, variables that differ from one warehouse to the next. HelioDock is not the first to offer this kind of oversight. Relay Robotics and YardOS both sell competing fleet-management products, so Night Shift arrives as one option among several rather than a new category. Its distinguishing claims, for now, are the beta results HelioDock has reported and the per-robot pricing it has set. ## The funding round The $24 million Series B, led by North Coast Ventures, brings HelioDock's total raised to $39 million, meaning the company had collected $15 million before this round. The timing pairs a software launch with fresh capital for a company that has until now been known for the robots themselves rather than the tools used to run them. General availability on September 15 will be the first test of whether the beta results hold up across a broader set of customers, including those whose warehouses sit outside the range of a dependable signal.

MoonshotAI: Kimi K3: # Cloudnote Debugging Page Exposed Customer Prompt Excerpts for Four Hours *The page displayed up to 200 characters from prompts across 1,842 user sessions. Cloudnote says no one besides the researcher who reported the flaw appears to have accessed it — but its investigation is not finished.* Cloudnote, a note-taking service with 3.2 million registered users, left a debugging page containing excerpts of customer prompts accessible without authentication for four hours on Monday, the company said. The page was removed 19 minutes after an outside researcher reported it. Logs show the page contained excerpts from 1,842 user sessions, each including up to 200 characters of text customers had entered into prompts. Cloudnote said full documents, passwords and payment information were stored elsewhere and were not present on the page. The exposure was brief, but the material involved is not necessarily trivial. Prompts submitted to services like Cloudnote can contain sensitive personal or commercial information, and even a 200-character excerpt is long enough to capture a name, an address or a confidential plan. Cloudnote said it has found no evidence that anyone other than the researcher accessed the page during the four-hour window. That conclusion is based on the company's own log review, however, and the investigation is not complete. The company has not said how many individual users the 1,842 sessions represent, or whether affected users will be notified. It also remains unclear how the page came to be publicly reachable, or why live customer data was present on it.

Anthropic: Claude Fable 5.1: # Cloudnote Says Debug Page Exposed Prompt Excerpts From 1,842 User Sessions *The page was reachable without a login for four hours Monday. The company says it has found no sign that anyone but the researcher who reported it looked, but its investigation is not finished.* Cloudnote, which has 3.2 million registered users, said a debugging page that displayed excerpts of customer prompts was accessible on the open internet without authentication for about four hours on Monday. The page contained snippets from 1,842 user sessions, according to server logs the company reviewed. Each snippet could include up to 200 characters of a prompt — enough to reveal a fragment of what a user typed, but not an entire document. Cloudnote said full documents, passwords and payment information were stored on separate systems and did not appear on the page. An outside security researcher notified the company of the exposure, and Cloudnote removed the page 19 minutes later. The company said it has found no evidence that anyone other than the researcher viewed the page while it was accessible. That finding is preliminary. Cloudnote said its investigation is still underway, and it has not said how the page came to be exposed, whether it can identify the users whose sessions were affected or what, if anything, it will tell them. The number of affected sessions is small relative to Cloudnote's user base. But prompts can contain whatever a person chose to type — names, personal circumstances, confidential work — which is why even short excerpts can be sensitive.

MoonshotAI: Kimi K3: # The Best Model Wasn't the Most Expensive One A new benchmark of five language models on real customer-support work produced a result that should interest anyone paying for AI by the token: the highest-scoring model was effectively tied with a rival that answered 18% faster and cost roughly 60% less. ## How the test worked Researchers assembled 200 anonymized customer-support tickets and asked each of the five models to draft one proposed answer per ticket. Three experienced support managers then graded every response blindly — they didn't know which model had written which answer. The grades combined three dimensions: factual correctness, usefulness, and tone. That design can't capture everything about a production deployment, but blind human grading on real tickets is a reasonable proxy for the core job of drafting a reply. ## The headline: a tie at the top Model North posted the best average score, 84.2. Model Vale scored 83.8. On paper that's a win for North; in practice it isn't. The researchers report that the difference was not statistically significant, meaning a gap that small could easily be noise rather than a real quality advantage. What is real is the difference in speed and price. Vale returned answers 18% faster, and its estimated cost is $0.74 per 1,000 tickets against $1.90 for North — a saving of about 61% for quality that measured as equivalent. | 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 | The more meaningful separation in the table isn't between North and Vale — it's between that leading pair and the rest of the field. Model Ember trails the leaders by about five points, and Model Harbor sits more than 13 points behind North. The benchmark didn't report significance tests for those lower comparisons, so the precise ordering of Ember, Kite and Harbor should be read cautiously. But the overall shape is clear: two models lead on quality, and below them you trade meaningful quality for steeply falling costs. ## Strengths differ by ticket type The category breakdown adds a wrinkle. North performed best on billing disputes, while Vale performed best on technical troubleshooting. With only 200 tickets in total, category-level results rest on small subsamples, so these are hints rather than verdicts. Still, they're worth weighing against your ticket mix: a company whose queue is dominated by billing questions has a stronger case for North's premium than one buried in technical tickets — pending confirmation with its own data. ## One weakness every model shared The most uniform result in the study was failure. All five models performed poorly on tickets that depended on an image attachment. Because the weakness spans the whole field, switching vendors won't fix it. Teams deploying any of these models need a workflow answer — routing image-dependent tickets to human agents, adding vision-capable tooling, or both — and should retest after any change. ## What this benchmark can't tell you The limits here matter as much as the scores. The tickets came from a single software company and were all in English, so the results may not transfer to other industries, customer bases or writing styles. The published numbers are averages, so they can't show how often a model produced an answer bad enough to need a full rewrite. And the test measured only answer-drafting: it did not cover agentic actions, tool use or multilingual support — capabilities like looking up accounts, issuing refunds or replying in a customer's language that increasingly define real deployments. ## The takeaway For a support organization that looks like the one tested, Vale is the rational default: statistically indistinguishable quality, meaningfully faster, at well under half the cost. North earns its premium only if follow-up testing confirms that its edge on billing disputes is real and relevant at your volumes. And no one should walk away thinking any of these models handles the full job — until image-dependent tickets are solved, a human fallback isn't optional.

Anthropic: Claude Fable 5.1: # The Top Two Models Tied on Quality. One Costs Less Than Half as Much. In a new benchmark of five language models on customer-support work, the headline is not which model scored highest. It is that the two highest scorers were statistically indistinguishable — and one of them costs about 40 percent of what the other does. Model North posted the top average score, 84.2, and Model Vale came in at 83.8. The researchers found that gap too small to be meaningful; it falls within the noise you would expect from 200 tickets graded by three people. Yet North's estimated cost was $1.90 per 1,000 tickets, while Vale's was $0.74. Vale also returned answers 18 percent faster. For a support operation choosing between these two, that combination — equivalent quality, lower price, faster responses — is the clearest result in the study. ## How the test worked The researchers drew 200 anonymized customer-support tickets from a single software company. Each of the five models produced one proposed answer per ticket, with no retries or human editing. Three experienced support managers then graded every answer without knowing which model wrote it. The grading rubric combined three dimensions: factual correctness, usefulness to the customer, and tone. Those were rolled into a single score per answer, then averaged across all 200 tickets for each model. That design has real strengths. Blind grading by people who actually manage support teams is closer to the real standard a deployed model would be held to than an automated metric would be. And 200 tickets is enough to separate models that differ substantially. It is not, however, enough to separate models that differ by less than half a point. That is why the North–Vale result should be read as a tie rather than a narrow win. ## Where the gap is real Below the top pair, the differences get large enough to matter. Model Ember scored 79.1 at $0.31 per 1,000 tickets. Model Kite scored 76.4 at $0.22. Model Harbor scored 71.0 at $0.08. Each step down the list saves money, and the cheapest option is nearly a tenth of Vale's cost. Whether that trade is worth it depends on what a lost point of quality costs in practice — more escalations, more follow-up messages, more customers who leave unhappy. The benchmark does not measure those downstream effects, so it cannot say where the right cutoff is. What it does establish is that the drop from the top tier to the bottom is substantial, roughly 13 points, and that is not noise. ## The two leaders are not interchangeable everywhere Even though North and Vale tied overall, they did not tie on every kind of ticket. North performed best on billing disputes. Vale performed best on technical troubleshooting. These category-level results come from smaller slices of the 200 tickets, so they carry more uncertainty than the overall averages. Still, they suggest that "which model is best" may depend on a team's ticket mix. A company drowning in refund arguments and a company drowning in configuration errors might reasonably reach different conclusions from the same data. ## Where every model struggled One finding cut across all five models: when a ticket depended on an image attachment — a screenshot of an error, a photo of a damaged product — performance dropped sharply for everyone. This is a useful caution for anyone planning a deployment. Image-dependent tickets may be a minority of volume, but they are often the ones where a wrong or vague answer is most frustrating to the customer. Routing those tickets to humans, or at least flagging them for review, looks like a prerequisite rather than a nice-to-have. ## What the benchmark does not tell you The study's scope is narrow in ways that matter. All tickets were in English and came from one software company. A different product, industry, or customer base could shift the rankings. A multilingual support queue is entirely untested. The models only wrote replies. The researchers did not test agentic actions — issuing a refund, resetting a password, looking up an order — or any form of tool use. Many real support deployments depend on exactly those capabilities, and a model that writes well may not act well. Cost figures are estimates and will vary with prompt length, pricing changes, and volume discounts. ## The practical takeaway If you are weighing North against Vale for text-based support replies, this benchmark gives you little reason to pay the premium for North unless billing disputes dominate your queue. If you are considering a cheaper model, expect a measurable quality drop and decide deliberately whether it is acceptable. And whichever model you choose, plan for the fact that none of them handle image-dependent tickets well yet.

MoonshotAI: Kimi K3: # QueueZero Unveils Proxy, an AI Meeting Assistant That Turns Talk Into Tasks QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to attend meetings, produce notes and convert conversation into follow-up work assigned to software agents. Proxy arrives as software makers race to build "agents" that do more than answer questions — systems that can take actions across email, documents and internal tools on a user's behalf. The category has drawn heavy investment, along with growing scrutiny over how much autonomy such systems should have. The idea has moved quickly from demos to paid products, even as early adopters report mixed results with systems that act without being asked. During a meeting, Proxy produces running notes, identifies commitments and compiles a list of follow-ups. Each item can then be handed to a software agent — a program that carries out work in other applications — rather than waiting for a human to pick it up. The launch demonstration offered an early, unscripted look at how literal that pipeline can be. Midway through the event, 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. Within moments, it had created two tasks, both assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. It was, by most measures, the fastest an action item from a product demo has ever been closed. The episode also underscored the central design question facing agentic software: where to draw the line between initiative and oversight. QueueZero said Proxy requires user confirmation before any external message is sent. Internal tasks, by contrast, can be created automatically, and administrators can disable automatic task creation entirely. The demo incident stayed inside that internal system; Proxy created tasks but sent nothing. The distinction matters because Proxy's pitch depends on initiative. An assistant that asks permission before every action saves users little time; one that never asks invites obvious problems. QueueZero's answer is to treat internal task creation as low-stakes and reversible, while gating anything that leaves the organization behind human approval. Wednesday's mishap gave the company an unexpected opportunity to explain that model in concrete terms. Proxy costs $30 per user per month. It enters public beta next week.

Anthropic: Claude Fable 5.1: # QueueZero's New AI Assistant Assigns Its First Tasks — to the Company's CEO QueueZero launched Proxy, an AI meeting assistant, on Wednesday, and within minutes of the demonstration the product had assigned work to the person running the show. Proxy is designed to attend meetings on behalf of a user, produce notes and route follow-up work to software agents that can draft documents, update records or schedule further meetings. The company pitched it as a way to reduce the amount of time employees spend sitting in on calls they do not need to lead. The launch demonstration was running live, with Proxy listening to the room, when a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." The remark was directed at QueueZero chief executive Evan Wu, who was presenting. Proxy interpreted the comment as an instruction. It created two tasks and assigned both to Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task on stage and deleted the first. The pricing sheet, at least, went out faster than it would have under most human workflows. The episode drew attention to a design decision that QueueZero had not emphasized in its launch materials. According to the company, Proxy requires confirmation from a user before it sends any external message, such as an email to a client or a note to another organization. Internal tasks, however, can be created automatically without a confirmation step, which is how the two assignments to Wu appeared without anyone approving them. QueueZero said administrators can disable automatic task creation for their organizations, in which case Proxy will propose tasks for review rather than generating them on its own. The company did not say whether that setting will be enabled by default when the product reaches customers. The incident illustrates a broader challenge for AI assistants that operate on conversational input. Meetings are full of offhand remarks, sarcasm and suggestions that are not meant to be acted upon, and a system built to extract action items has to distinguish between a genuine assignment and a comment made in passing. In this case, Proxy correctly identified the speaker's intent — the customer did want the pricing sheet — but also turned a rhetorical aside into a formal deliverable. Proxy will cost $30 per user per month. It enters public beta next week, and QueueZero said feedback from early users will inform how the task-creation controls are refined before general availability. The company declined to say whether Wu's first task has been reopened.