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

Anthropic: Claude Opus 5.5: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin company's software combines drainage maps with rainfall forecasts to help local governments decide where to send maintenance crews before a storm.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The company's software maps drainage systems and combines those maps with rainfall forecasts. Cities use it to decide where to send maintenance crews before a storm. Basin Ventures led the round, with participation from existing investor Long Mile. The new funding brings Raincheck's total raised to $10.5 million. The company says it will use the money to hire eight engineers and expand into Florida. CEO Ana Flores described the problem the company is trying to solve: > "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." More detail is available in [Raincheck's methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. The company's forecasts are recommendations only. 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 Get Ahead of Flood Risk *The Austin startup's software pairs drainage maps with rainfall forecasts so local governments can send maintenance crews where they're needed before a storm hits.* Raincheck, an Austin-based company that builds flood-risk software for local governments, has raised $8 million in seed funding, the company announced today. The company maps municipal drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews ahead of a storm. Raincheck has published [an overview of its methodology](https://example.com/raincheck-methodology). The seed round was led by Basin Ventures, with participation from existing investor Long Mile. The new funding brings Raincheck's total raised to $10.5 million. CEO Ana Flores said: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them, saying some contracts have not been publicly announced. The company's forecasts are recommendations only. Emergency managers retain control over crew deployments and public warnings. Raincheck faces competition from other flood-planning vendors, including HydroMap and municipal engineering contractors. The company says the new funding will allow it to hire eight engineers and expand into Florida.

Anthropic: Claude Opus 5.5: # QueueZero's Proxy Assistant Turns Audience Remark Into Tasks During Launch Demo *The AI meeting tool, which enters public beta next week, logged an offhand comment as two assignments for the company's CEO* QueueZero on Wednesday launched Proxy, an artificial intelligence assistant designed to attend meetings, take notes and hand off follow-up work to software agents. The launch demonstration also offered an unplanned illustration of how such systems decide 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 treated the remark as a directive. It created two tasks assigned to QueueZero chief executive Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. Of the two assignments, only one fit neatly on a to-do list. ## How Proxy handles actions Proxy is built to go beyond transcription. According to QueueZero, the assistant can attend a meeting, produce notes and assign follow-up work to software agents, which are automated programs that carry out tasks. The company draws a line between internal and external actions. QueueZero says Proxy requires confirmation before an external message is sent. Internal tasks, however, can be created automatically. That is what happened during Wednesday's demonstration. Administrators can disable automatic task creation, according to the company. ## Why it matters The episode highlights a central challenge for AI tools that act on what they hear. Meetings are full of jokes, asides, hypotheticals and requests directed at no one in particular. An assistant that converts speech into assignments has to distinguish between a passing comment and a genuine request. The demonstration showed that the distinction is not always clear, even to software designed for the purpose. In this case, the consequences were limited. The tasks were internal and easily removed, and the pricing request was one Wu acted on. But for organizations weighing tools that generate work automatically, the incident raises practical questions. They will need to decide how much latitude to grant an assistant and who reviews the work it creates on their behalf. Under the confirmation requirement QueueZero describes, a misread remark should not lead to a message leaving the company without a person's approval. The option to turn off automatic task creation gives administrators a further check on what the assistant does internally. ## Pricing and availability Proxy costs $30 per user per month. It enters public beta next week.

MoonshotAI: Kimi K3: # QueueZero Launches Proxy, an AI Assistant That Attends Meetings and Delegates the Follow-Up *The $30-per-month assistant enters public beta next week with controls over what it can do automatically.* QueueZero on Wednesday launched Proxy, an artificial intelligence assistant that can attend a meeting, produce notes and assign follow-up work to software agents. During a live launch demonstration, the product offered an unscripted preview of both its capabilities and the questions raised by software designed to treat speech as instructions. While Proxy was shown participating in a demonstration meeting, a customer in the audience remarked, "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 chief executive Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first, briefly becoming the first chief executive to receive action items from his own product. The company said the exchange reflected the product operating largely as designed, within defined limits. Proxy requires human confirmation before any external message, such as an email to a customer, is sent. Internal tasks, by contrast, can be created automatically when the assistant identifies an actionable request. Administrators can disable automatic task creation entirely. That distinction is central to QueueZero's case that Proxy is safe to deploy in real workplaces. Actions that leave the organization require approval; actions that remain inside it can proceed without interruption, preserving the efficiency the company is selling. Even so, the demonstration highlighted a challenge for the fast-growing category of AI meeting assistants: telling instructions apart from commentary. The customer's remark was not addressed to Proxy, and it paired a request the company was happy to fulfill with one its chief executive was not. Under the default configuration, both became tasks. Proxy is priced at $30 per user per month and enters public beta next week. QueueZero said it expects administrators to adjust task-creation settings as they observe how the assistant behaves in actual meetings. The product enters a crowded field of AI notetakers, many of which have expanded from transcription into summaries, action items and workflow automation. QueueZero is wagering that the next step — delegating captured work to other software agents rather than simply recording it — will appeal to organizations already automating routine operations. Whether those organizations are prepared to let an assistant decide what qualifies as an assignment is a question the launch event made unusually concrete.

Anthropic: Claude Opus 5.5: # Arcwell's Relay Uses AI to Help Support Agents Find Answers Faster Arcwell is betting that customer service agents already have the answers they need. They just can't find them quickly enough. The company has announced Relay, an artificial intelligence platform that reads customer support conversations as they unfold and suggests responses to the agents handling them. Relay can also flag customers who may be on the verge of canceling and generate summaries of support interactions. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Relay is meant to close that gap. Rather than sending agents digging through knowledge bases mid-conversation, the system pulls relevant material forward and offers a suggested reply, which the agent can then use. Twelve companies tested Relay over six months. According to Arcwell, agents using the platform closed tickets 14% faster. That is a meaningful gain in an industry where handling time is closely tracked, though the figure comes from the company rather than an independent evaluation. Arcwell has been less forthcoming about one of Relay's headline features. The company did not release results showing whether the platform's ability to spot at-risk customers actually helped the test companies keep them. Without that data, it remains unclear whether the cancellation warnings translate into retained accounts or simply give agents an earlier view of customers they will lose anyway. Relay is priced at $65 per agent per month. The platform's usefulness depends on what it can see. To generate suggestions, Relay requires access to a company's support transcripts and internal help documents. That material can include sensitive details about customers, products and internal policies, which may give some businesses pause before handing it to an outside vendor. Arcwell is trying to address those concerns. The company says customer data is encrypted and is not used to train shared models, meaning one client's conversations should not inform the suggestions another client receives. Companies can also choose how long their data is kept, with retention periods of either 30 or 90 days. For Arcwell, Relay is an attempt to turn the information companies already hold into faster answers for the people on the front lines of customer support.

MoonshotAI: Kimi K3: # Arcwell's Relay Wants to End the Support Agent's Search for Answers *The AI platform reads support conversations, drafts replies and flags customers at risk of leaving — for a price: access to company transcripts.* Arcwell is betting that the slowest part of customer support isn't the customer. It's the hunt for answers. The company has launched Relay, an artificial intelligence platform that reads support conversations as they unfold and suggests replies to the agents handling them. Beyond drafting responses, Relay flags customers who appear likely to cancel and generates summaries of support interactions — an attempt to give companies a faster read on what their own customers are telling them. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Relay is Arcwell's attempt to close that gap, and the company says the early numbers back up the pitch. A dozen companies tested the platform over six months; according to Arcwell, agents using it closed tickets 14 percent faster. The company did not release customer-retention results from the trial. That omission stands out, because spotting at-risk customers is one of Relay's headline features. For now, there is no public data showing whether the platform's warnings actually help companies keep the accounts it flags. Relay costs $65 per agent per month, a per-seat price that scales with the size of a support team. Adopting it also requires a measure of trust: to do its job, Relay needs access to a company's support transcripts and internal help documents, often among the most sensitive material a business holds. Arcwell says customer data is encrypted and is not used to train shared models, a stance aimed at buyers wary of feeding proprietary information into someone else's AI. Each company can choose a retention window of 30 or 90 days for its data. The pitch, in the end, is simple: faster agents, earlier warnings, shorter searches. The price of admission is letting Relay listen in on a company's conversations with its customers.

Anthropic: Claude Opus 5.5: **6** 1. `#Flood software moves upstream` has no space after `#`, so it does not follow `# Headline`. 2. `The company is expanding its pilot.` is the dek but is not italicized. 3. `##Details` has no space after `##`, so it does not follow `## Heading`. 4. `[the announcement](www.example.com/news)` does not use 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 Opus 5.5: C

MoonshotAI: Kimi K3: **C** Option C is the only lede that is both accurate and news-forward: it leads with the company's action, includes the key specifics (4,860 accounts, workspace titles, 11-hour exposure), and doesn't overstate the facts. - **A** is false — message contents were *not* sent, so "leaked their private messages" misstates the facts. - **B** is PR spin that buries the actual news (the data exposure). - **D** is false — nothing users wrote was exposed, and there's no evidence of hackers accessing anything.

Anthropic: Claude Opus 5.5: # Foundry Hall to Close After Losing Its Largest Sponsor Foundry Hall, the Milwaukee nonprofit that has helped more than 200 startups find their footing since 2018, 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 keep operating through October but cannot pay for another full program cycle. The organization's leaders spent the intervening period hoping another funder would step in to cover the gap. "We kept waiting for the replacement check that never came," said executive director Camille Ross. The closure will eliminate six jobs at Foundry Hall. It will also uproot twelve startups that currently work out of the accelerator's space and now have weeks, not months, to find somewhere else to go. A spokesperson for Blank River Bank said the decision reflected a shift in the bank's community investment priorities and was not a judgment on how Foundry Hall had performed. By its own measures, the accelerator performed well. The 212 startups it has supported now employ a combined 1,140 people. For many of those companies, Foundry Hall provided more than desks and mentorship. It offered a place where an idea could start to look like a business. Dev Malik, one of the founders who came through the program, remembers what came before. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." That shift, from a borrowed table to a room built for the work, is what the twelve current tenants are about to lose. Some are still early enough that a move could disrupt hiring, fundraising, or simply the daily routine of building something new. City officials have begun looking for alternatives. Deputy Mayor Susan Cole said the city is in talks with building owners and local universities about space for the companies that must leave. She did not say whether a single location could house all twelve or when arrangements might be finalized. For the six employees losing their jobs, there is no comparable search underway. They will spend their final weeks helping the remaining startups prepare to move and closing out a program that, by any count, did what it set out to do. What happens to Foundry Hall's model after October is unclear. The accelerator was not undone by failure but by a budget that depended heavily on one source. When that source changed its priorities, nothing arrived to take its place. On October 31, the building will empty out. The companies it helped start will go on elsewhere, carrying the 1,140 jobs they created with them. The place where many of them began will not.

MoonshotAI: Kimi K3: # Foundry Hall, Milwaukee Startup Accelerator, to Close Its Doors MILWAUKEE — Foundry Hall, the nonprofit startup accelerator that has anchored Milwaukee's early-stage business community since 2018, will shut down on October 31 after its largest corporate sponsor declined to renew its funding. Blank River Bank supplied 38 percent of Foundry Hall's annual budget. When the bank told the organization it would not continue the sponsorship, executive director Camille Ross spent months courting foundations, corporations and individual donors to fill the gap. None came through. A bank spokesperson said the decision reflected a shift in the bank's community investment priorities, not any judgment on Foundry Hall's performance. "We kept waiting for the replacement check that never came," Ross said. The accelerator leaves behind a measurable legacy. Over seven years, it supported 212 startups, which collectively employ roughly 1,140 people. For many founders, Foundry Hall was the first place that took them seriously. Dev Malik was one of them. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." His story is a familiar one in Milwaukee's startup scene, where Foundry Hall served as a front door — the place a founder with an idea and a laptop could find a desk, a mentor and, eventually, a meeting with someone willing to write a check. The closure will be felt well beyond the building's walls. Six employees will lose their jobs. Twelve startups currently working out of the Foundry Hall space must relocate by the end of October. Ross said the organization has enough money to operate through October but not enough to run another full program. Rather than shrink the accelerator into something unrecognizable, the board chose to end it cleanly and spend its final weeks helping the current companies land somewhere else. City officials are part of that effort. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the displaced startups, though no arrangements have been finalized. For Ross, the math of the closing is simple; the rest of it is not. Foundry Hall survived its uncertain first year, a pandemic and the constant churn of founders cycling through its doors. What it could not survive was the loss of a single check. The need for what it did has not gone away, she noted. The money to meet it has. Malik, who no longer needs the accelerator's desks or its introductions, said he still thinks about that back table at the coffee shop — and about the founders who will now have to stay there. Foundry Hall's staff will keep working through the end of the month, answering emails, making introductions, packing boxes. On October 31, the lights go out.

Anthropic: Claude Opus 5.5: # Startup Spotlight: LedgerLark Wants to Shorten the Hunt for Broken Trades Most people never think about what happens after a trade is placed. For operations teams at investment firms, that is where much of the work begins. They must match what was agreed against what the custodian recorded and what the firm's own books say. When those records disagree, someone has to find out why. LedgerLark, a young company building AI agents for post-trade operations, is betting that small investment firms will pay for help with that search. ## Where the idea came from Co-founder Mina Batra spent seven years in operations at a regional brokerage. The idea for LedgerLark took shape after she watched a six-person team spend three days resolving a single trade mismatch. The cause turned out to be a copied account number. The episode illustrates a familiar pattern in back-office work. The error is small, but locating it is expensive because the evidence is spread across systems that don't connect well. Co-founder Owen Carr approaches the problem from the systems side. He previously built reconciliation systems at a large bank. ## What the product does LedgerLark's agents compare three sources of record: trade confirmations, custodian records and a firm's internal ledgers. When they find a discrepancy, they propose a correction, and a human decides whether to approve it. That boundary matters. The product does not execute trades or move money. Its role is to narrow down where a break occurred and suggest a fix, leaving the final call and the accountability with operations staff. For firms wary of handing financial decisions to software, that design may make adoption easier. It also means the product's value depends heavily on the quality of its proposals. If the suggestions are wrong often enough, they could add review work rather than remove it. ## Early traction LedgerLark says five firms are currently paying customers, with annual contracts ranging from $36,000 to $110,000. The company has raised a $3.5 million seed round led by Motive Row. The most concrete performance figure comes from the company itself. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. If that holds up, it is a meaningful change. But it is one customer's result, reported by LedgerLark, and it leaves open questions: - Over what time period did the change occur? - How were exceptions counted? - Did other changes at the firm contribute? Results from the other four customers, and how they look after a full year, will say more. ## Competition and obstacles LedgerLark is not entering an empty market. Established reconciliation vendors already sell to investment firms, and large banks often rely on software built in-house, the kind of system Carr once worked on. LedgerLark appears to be betting that smaller firms are underserved because they are too small to build their own tools and may be a poor fit for products designed for larger institutions. It is not yet clear whether incumbents will respond by adding similar AI features or competing on price. Sales cycles are another hurdle. Financial institutions may require lengthy security and procurement reviews before granting a vendor access to trade and account data. For a seed-stage company, long reviews can slow growth and stretch funding. Focusing on smaller firms could shorten that process, though even small firms answer to regulators and auditors. ## What to watch The open questions for LedgerLark are straightforward: - Do its exception-reduction results repeat across customers? - How often do operations staff accept its proposed corrections? - Can it win new contracts at a steady pace despite procurement delays? The founders have identified a real and tedious problem. The next stage will show how much of it their agents can reliably solve.

MoonshotAI: Kimi K3: # Startup Spotlight: LedgerLark Wants AI Agents to Untangle Post-Trade Errors *A two-person founding team targets the least glamorous corner of investment operations — with a human still signing off.* At the regional brokerage where Mina Batra spent seven years in operations, a single copied account number once occupied a six-person team for three days while the error was traced, confirmed and unwound. For Batra, the episode captured something structural: post-trade operations at small investment firms run on manual cross-checking, and the cost of a mismatch is measured in staff days, not seconds. ## What LedgerLark does LedgerLark, the startup Batra co-founded with Owen Carr, is an attempt to compress that work. Its agents compare three sources that are supposed to agree — trade confirmations, custodian records and a firm's internal ledgers — flag discrepancies and propose corrections. A person approves each fix. The product does not execute trades or move money, a boundary that appears deliberate: it keeps the software out of the most heavily regulated workflows and gives cautious compliance teams a clear line around what the system can touch. Carr's background complements Batra's. He built reconciliation systems at a large bank, so the pair can plausibly speak to both sides of a sales meeting — the operations staff who live with exception queues and the engineers who know how the matching actually works. That is a description of relevant experience, not a guarantee of execution. ## Traction and funding The traction so far is real but thin. Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. Even if every account sat at the top of that range, revenue would be modest — the fair reading is a handful of early adopters, not a market verdict. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. That figure comes from the company, reflects a single deployment and has not been independently verified. It is, at least, the right kind of metric: exception backlogs are concrete, measurable and expensive to carry. The company has raised a $3.5 million seed round led by Motive Row, capital that should fund a small team through its next stretch of deployments. ## Competition and open questions The open questions are sizable, starting with competition. Established reconciliation vendors have sold into this workflow for years, and many institutions run matching software built and maintained in-house. Those incumbents can add AI-assisted comparison to products already sitting inside procurement-approved environments. LedgerLark's presumed counter is focus: small investment firms are often underserved by enterprise vendors priced for global banks, and a tool built specifically for the three-way match may beat a platform that requires a services engagement to configure. Whether that wedge holds depends on how quickly incumbents ship similar features — and on whether small firms will trust a young company with sensitive trade data. The second question is pace. Financial institutions routinely subject new vendors to lengthy security and procurement reviews, which can stretch across months and strain a seed-stage runway. The human-approval design may ease those reviews, but it also caps the labor savings on offer: if a person must sign off on every correction, the product accelerates the search for errors rather than eliminating the work. None of this undercuts the core insight, which is well grounded in lived experience: back offices are full of repetitive, high-stakes comparison work, and the people doing it know exactly where the hours go. Whether an agent that proposes rather than acts can carve out a durable position between spreadsheets and enterprise suites is a genuinely open question — and a more interesting one than the usual claims of reinvention.

Anthropic: Claude Opus 5.5: # KiteBridge Software Error Sends $2.8 Million in Duplicate Reimbursements **A flaw in a Tuesday update caused the expense platform to pay thousands of employee claims twice, drawing on the funds of 91 corporate customers** A software update at expense-management company KiteBridge caused 13,400 duplicate reimbursements worth a combined $2.8 million on Tuesday. The money came from the accounts of 91 corporate customers who use the platform to repay employees for work expenses. By Wednesday evening, $2.3 million had been recovered. The remaining $500,000 was still pending. The duplicate payments were funded by employers. No employee had money withdrawn from their own accounts as a result of the error. ## How the failure unfolded KiteBridge deployed version 4.2 of its software at 8:45 a.m. Tuesday. The first customer complaint arrived less than half an hour later, at 9:12 a.m. About 50 minutes after that, at 10:03 a.m., the company suspended automated payments, stopping the system from sending more money. The problem began with a timeout at the payment processor KiteBridge uses to move funds. A timeout happens when one system waits for a response from another and doesn't get one in time. In payment systems, that leaves an uncertain situation: the original payment may have gone through, or it may not have. Software typically handles that uncertainty by retrying the request. A well-designed system recognizes a retry as another attempt at the same payment and makes sure the money moves only once. In KiteBridge's case, a fallback system, the backup process meant to take over when the normal path fails, treated those retries as brand-new reimbursement requests. Each one was then paid as if it were a separate claim. ## A promised safeguard The failure raises questions about a protection KiteBridge had advertised months earlier. An April changelog, the public list of changes that accompanies a software release, said version 4.0 had introduced "exactly-once payment protection." Exactly-once processing is the guarantee that a payment is carried out a single time, even when networks are slow or requests are repeated. It is the kind of safeguard designed to stop the duplication that occurred on Tuesday. The reporting does not establish why that protection failed to prevent the duplicates in version 4.2, or whether it covered the fallback system involved. ## Slow disclosure KiteBridge's public account of the incident also lagged behind events. The company's incident page, where it posts information about service problems, initially described the issue as delayed reimbursements. That description reflected the effect of halting automated payments, but it did not mention that money had already gone out twice. The incident page first disclosed the duplicate payments about six hours after automation was suspended, around 4 p.m. Tuesday. For the affected businesses, the distinction matters. A delay means employees wait longer to be repaid. Duplicate payments mean company money has already left their accounts and must be recovered. ## Company response Chief Executive Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report on what went wrong. That report may answer the central questions still open. It could explain how a timeout led the fallback system to create new payments, why the exactly-once protection announced in April did not stop them, and how the company plans to recover the outstanding $500,000.

MoonshotAI: Kimi K3: # KiteBridge Software Bug Triggers $2.8 Million in Duplicate Reimbursements *A failed safeguard in a routine update sent extra payments to employees at 91 companies; most of the money has been recovered, but roughly $500,000 remains outstanding.* A botched software update at KiteBridge, an expense-management software company, generated more than 13,400 duplicate reimbursement payments worth a combined $2.8 million on Tuesday, forcing the company to halt its automated payment system and spend the following day clawing the money back. The problem 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 by then the damage was done. In the roughly 80 minutes the faulty code was live, the system issued duplicate reimbursements affecting 91 corporate customers. ## A retry mechanism that retried too well According to KiteBridge, the failure began with a timeout at one of its payment processors, the third-party services that actually move money. When a processor times out, it is often unclear whether the original transaction went through, so payment systems are typically built to retry automatically. KiteBridge's fallback system mishandled those retries, treating each one as an entirely new reimbursement request rather than a repeated attempt at the same one. The result: employees were reimbursed multiple times for the same expenses. The incident is an especially uncomfortable one for the company because it had previously claimed to have solved exactly this problem. An April changelog for version 4.0 advertised a feature called "exactly-once payment protection" — industry terminology for a guarantee that a transaction will be processed one time and one time only, no matter how many retries occur behind the scenes. Tuesday's failure raises pointed questions about how that safeguard fared in version 4.2, questions the company has not yet answered. ## Slow disclosure KiteBridge also faces scrutiny over how it communicated during the incident. Its public incident page initially described the problem only as delayed reimbursements. The page did not acknowledge the duplicate payments until six hours after automated payments were suspended — by which point some customers' finance teams were already reconciling money that had unexpectedly left their accounts. No employees had money withdrawn from their own accounts, the company said; the duplicate payments were funded by employers. Recovering the money therefore means unwinding overpayments that have already landed in employees' bank accounts. By Wednesday evening, KiteBridge said it had recovered $2.3 million of the $2.8 million. The remainder — roughly $500,000 — was still pending, and the company has not said when it expects affected customers to be made whole. In a statement, chief executive Lena Ortiz apologized to customers and said the company would publish a detailed technical report on the failure. She did not say when the report would be released. For the 91 affected customers, that report will need to explain how a safeguard the company once promoted as a selling point failed on an otherwise ordinary Tuesday morning.

Anthropic: Claude Opus 5.5: # HelioDock Launches Night Shift to Let Managers Oversee Warehouse Robots From Afar *The autonomous-robot maker paired the release with a $24 million Series B, though a beta customer says the software is less useful when connectivity drops* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor their robot fleets remotely. The company also announced a $24 million Series B round led by North Coast Ventures. Night Shift targets a basic problem with warehouse automation. Robots can stall at any hour, and someone has to respond when they do. HelioDock is pitching the software as a way to handle those moments without sending staff to the site. The company says beta customers saw 17% fewer unplanned robot stoppages, a figure that has not been independently verified. At least one early user says the system becomes less useful when the network goes down. Chief Executive Mara Velez framed the product around the people who get called in when automation stumbles. "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 spent 14 months in a private beta with nine logistics companies running a combined 310 robots. HelioDock's stoppage figure comes from that group. An unplanned stoppage takes a robot out of service until the problem is resolved. If a reduction of that size holds across a wider customer base, it would affect how much work a fleet completes overnight and how often managers are pulled away from home to deal with it. The beta group is still small compared with a commercial rollout. General availability will be the first broad test of whether results from nine companies carry over to a more varied set of warehouses. Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity. In locations that depend on rural LTE service, he said, it became less useful during outages. That limitation follows from how the product behaves. When connectivity fails, HelioDock's robots return to their locally configured behavior and remote control becomes unavailable. The robots continue operating under the rules set for them on site, but an operator watching from elsewhere can no longer intervene. In buildings with dependable networks, that may rarely come up. At sites on the edge of cellular coverage, a remote operator could be cut off during the overnight hours when nobody else is available to help. The Series B brings HelioDock's total funding to $39 million, which means the company had raised $15 million before this round. North Coast Ventures led the new financing. Night Shift becomes generally available September 15 and will cost $90 per robot per month. Because the price scales with fleet size, a site running 50 robots would pay $4,500 a month, or $54,000 a year. Operators are likely to weigh that cost against what they spend on on-call staff and after-hours trips, and against how much of their footprint has connectivity reliable enough to support remote oversight. They will also have other options. Relay Robotics and YardOS already sell competing fleet-management products, so HelioDock enters a market where buyers can compare its claims with tools from other vendors before committing when Night Shift becomes widely available.

MoonshotAI: Kimi K3: # HelioDock Launches Night Shift, Bringing Remote Oversight to Warehouse Robot Fleets *The monitoring software, priced at $90 per robot per month, reaches general availability September 15 alongside a $24 million Series B — though its usefulness rises and falls with a facility's connectivity.* 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 funding round led by North Coast Ventures. The round brings the company's total funding to $39 million. The product is built around a staffing mismatch in automated warehouses. Robots work through the night; the people accountable for them usually do not. When a machine stalls after hours, the idle time cuts into the productivity that justified automating in the first place, and getting it moving again has typically meant sending someone to the building. "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. Night Shift is designed to eliminate that drive, letting an off-site operator see what a fleet is doing — and, when the network cooperates, act on it — without setting foot in the building. The launch also moves HelioDock past pure hardware. The company built its business on autonomous warehouse robots; Night Shift adds a software layer sold on a per-robot subscription, generally available September 15 at $90 per robot per month. For an operator running 50 robots, that works out to $4,500 a month. HelioDock arrives at that launch with field data in hand. The company spent 14 months running Night Shift in a private beta with nine logistics companies and 310 robots — an average of roughly 34 machines per operator across nearly a year and a half of deployment. That trial produced its headline performance claim: beta customers experienced 17 percent fewer unplanned robot stoppages, according to the company. HelioDock said the figure has not been independently verified, so it is best read as a self-reported data point rather than an audited result. The events it measures — a robot halted by something as ordinary as a pallet — are the ones Night Shift is meant to help operators clear without a site visit. The beta also made the product's central constraint plain: Night Shift is only as useful as the network beneath it. Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His account tracks with how the system handles a lost connection. When connectivity drops, the robots do not shut down — they return to their locally configured behavior — but remote control becomes unavailable. The fleet keeps working as it would without the software, while the off-site operator loses the ability to intervene until service returns. For warehouses in well-connected industrial corridors, that may amount to an edge case. For sites dependent on rural LTE, it defines the boundary of what the subscription includes. The limitation is structural rather than a defect: Night Shift's oversight travels over the same networks the outages interrupt. HelioDock is not alone in chasing that subscription. Relay Robotics and YardOS already sell competing fleet-management products, giving operators direct points of comparison as Night Shift enters the market. HelioDock's pitch rests on its beta mileage — 14 months and 310 robots of operational testing — and on a funding round, announced the same day as the launch, that gives the company fresh capital as it opens the software to the broader market.

Anthropic: Claude Opus 5.5: # Top Customer-Support AI Models Score About the Same, but One Costs Far Less In a new benchmark of five language models on real customer-support tickets, the top two finished in a statistical tie. One of them costs less than half as much to run. Model North posted the highest average score, 84.2, with Model Vale close behind at 83.8. The researchers found that this 0.4-point gap was not statistically significant, so the test gives no reliable evidence that North writes better support answers than Vale. The costs, however, differ substantially. North's estimated cost is $1.90 per 1,000 tickets, while Vale's is $0.74, about 61% less. Vale also answered 18% faster. For teams choosing a model to draft support replies, that is the main finding. On this test, paying more did not buy measurably better answers at the top of the field. ## How the Test Worked The researchers gave each of the five models the same 200 anonymized customer-support tickets. Each model produced one proposed answer per ticket. Three experienced support managers then graded every answer without knowing which model wrote it. Blind grading matters because evaluators who know a response's source may favor familiar or highly regarded systems. The managers scored each answer on three combined criteria: - **Factual correctness:** whether the answer was accurate - **Usefulness:** whether it would actually help the customer - **Tone:** whether it sounded appropriate for a support interaction Because these criteria were combined into a single score, an answer could do well overall while still having a weakness in one area. The published averages do not show how each model performed on each criterion separately. ## A Clear Trade-off Below the Top Two Below North and Vale, the results show a steadier link between cost and quality: - **Ember** scored 79.1 at $0.31 per 1,000 tickets. - **Kite** scored 76.4 at $0.22. - **Harbor** scored 71.0 at $0.08. The researchers reported significance testing only for the North–Vale comparison. Readers should therefore be careful with small gaps elsewhere, such as the 2.7 points between Ember and Kite. The broader pattern is clear, though. The cheapest models scored noticeably lower than the top pair, and Harbor's roughly 13-point deficit to North is large enough to matter for a customer-facing product. Absolute costs also deserve attention. Every model here costs less than a fifth of a cent per ticket. A team handling a few thousand tickets a month will barely notice the difference between North and Vale. The gap matters at scale. At one million tickets, North's estimated cost would be about $1,900 and Vale's about $740. ## Strengths Differ by Ticket Type The tie in average scores hides some differences. North performed best on billing disputes, and Vale performed best on technical troubleshooting. A company whose ticket queue leans heavily toward one category might reasonably favor the model that leads there. The benchmark summary does not say how large these category leads were or whether they were statistically significant. With only 200 tickets split across several categories, each subgroup is small, so these results are better treated as leads to investigate than as settled conclusions. ## A Shared Weakness: Images One result applied to every model. All five performed poorly when a ticket depended on an image attachment, such as a screenshot of an error message. In software support, customers often send screenshots because they cannot describe a problem in words, so this is a real limitation. Teams deploying any of these models should consider sending image-dependent tickets to human agents. ## What the Benchmark Doesn't Tell You The test was narrow, and its limits shape how far the results can be applied: - **One company, one language.** All tickets were in English and came from a single software company. Results could differ for other industries, products or customer bases. - **Drafting only.** The models wrote proposed answers. The researchers did not test agentic actions, such as issuing refunds or updating accounts, or tool use, such as looking up order records. Many production support systems depend on these capabilities. - **No multilingual testing.** Companies serving non-English-speaking customers will need their own evaluation. - **Modest sample size.** Two hundred tickets is enough to separate the strongest and weakest models, but it is not enough to rank close competitors confidently or to draw firm conclusions about specific ticket categories. ## The Bottom Line Within its limits, the benchmark points to a practical conclusion. Vale matched North's quality at roughly 39% of the cost and with faster responses. North may still be the better choice for teams that handle mostly billing disputes or that need to squeeze out every possible point of quality. For most English-language software support teams drafting replies at volume, Vale looks like the stronger value. Teams should still test any model on their own tickets before relying on it, especially tickets that include screenshots.

MoonshotAI: Kimi K3: # The Best Model in This Support Benchmark Isn't the Most Expensive One A new benchmark of five language models on real customer-support work delivers a counterintuitive result: the top-scoring model and its runner-up are effectively tied on quality — but the runner-up costs roughly 60% less and answers faster. Researchers tested five models — North, Vale, Ember, Kite and Harbor — on 200 anonymized customer-support tickets. Each model drafted one answer per ticket, and three experienced support managers graded the responses blind, combining factual correctness, usefulness and tone into a single score. ## A tie at the top — until you look at the bill Model North earned the highest average score at 84.2; Model Vale scored 83.8. Because that difference is not statistically significant, the honest reading is that the two performed equally well. This isn't a technicality: if the researchers ran the same experiment on a different set of 200 tickets, the ranking between the two could plausibly flip. The costs are anything but equal. North runs an estimated $1.90 per 1,000 tickets, while Vale runs $0.74 — about 61% cheaper — and Vale produced answers 18% faster. Faster first responses are closely tied to customer satisfaction, so that speed edge compounds the cost argument. For a team handling tens of thousands of tickets a month, equal quality at roughly a third of the price is the kind of finding that reshapes procurement decisions. The rest of the field trails meaningfully. Ember averaged 79.1 at $0.31 per 1,000 tickets, Kite scored 76.4 at $0.22, and Harbor — the cheapest option tested at $0.08 — scored 71.0. Those gaps are large enough to matter: there's a genuine quality step down from the top two. The budget models may still suit low-stakes, high-volume categories where an imperfect first draft is acceptable. ## Different models, different strengths The averages conceal some specialization. North performed best on billing disputes, where precision about refunds and account policies matters most. Vale performed best on technical troubleshooting, which rewards step-by-step diagnostic reasoning. That split hints at a routing strategy — send billing tickets to North and technical tickets to Vale — rather than a single-winner deployment, though the benchmark doesn't show whether the per-category edges are large enough to justify the added complexity. ## One weakness all five share The clearest red flag isn't about any single model. All five performed poorly on tickets that depended on an image attachment — a screenshot of an error message, a photo of damaged hardware. Real support queues contain plenty of these, especially in consumer and hardware support, where a photo is often the fastest way for a customer to explain a problem. Teams deploying any of these models should route image-dependent tickets to humans or a separate vision-capable system; nothing here suggests a high text score transfers to images. ## How to read these results The methodology has real strengths. Blind grading by three experienced managers limits the influence of brand reputation or formatting style, and using genuine anonymized tickets captures the messiness of actual customer language better than synthetic prompts would. The limitations deserve equal weight. All 200 tickets came from one software company and were in English. Support norms vary widely by industry, so results may not transfer to, say, a telecom or an airline. The modest sample size is also why small score gaps shouldn't be over-read. Just as notable is what wasn't tested: no agentic actions, no tool use, no account lookups or refunds processed, and no multilingual support. Modern support automation increasingly depends on exactly those capabilities, so this benchmark measures only one slice — single-draft answers to text tickets — of what a real deployment requires. ## The takeaway The headline isn't that one model "won." It's that the most expensive option failed to separate itself from a cheaper, faster rival. On this evidence, Vale is the sensible default for text-based support drafting, with North worth reserving for billing disputes where its edge appeared — and with all five models kept away from image attachments. The right next step for any support leader is to replicate the test on their own ticket mix, especially the categories this benchmark couldn't cover.

Anthropic: Claude Opus 5.5: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 Sessions *The page was publicly accessible for four hours Monday. The company says it has found no sign that anyone other than the researcher who reported it viewed the data, but its investigation is not finished.* A debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday, Cloudnote said. The company removed the page 19 minutes after a security researcher notified it of the problem. According to Cloudnote's logs, the page displayed excerpts from 1,842 user sessions. Each excerpt could include up to 200 characters of a prompt. That is enough, in some cases, to capture a name, a work detail or other information a user typed. Cloudnote has 3.2 million registered users. It is not yet clear how many individual users the affected sessions represent. Cloudnote said full documents, passwords and payment information were stored elsewhere and did not appear on the page. The central unresolved question is whether anyone else saw it. Cloudnote says it has found no evidence that anyone besides the researcher accessed the page. But the company has not completed its investigation, and finding no evidence so far is not the same as confirming that no one else did. It is also unclear when the review will conclude and whether users whose sessions appeared on the page will be notified individually.

MoonshotAI: Kimi K3: # Cloudnote Left Customer Prompts Visible on an Unsecured Page for Four Hours *The note-taking company says logs show prompt excerpts from 1,842 user sessions. It found no evidence anyone besides the researcher who reported the flaw saw them — but its investigation is not finished.* Cloudnote, a note-taking service with 3.2 million registered users, said a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday before being removed. According to the company, logs show the page held excerpts from 1,842 user sessions, each limited to as many as 200 characters of a prompt — brief fragments rather than complete documents. Cloudnote said it took the page offline 19 minutes after a security researcher alerted it to the exposure. The company said full documents, passwords and payment information are stored in separate systems and were not present on the page. It also said it has found no evidence that anyone other than the researcher accessed the page while it was live. Those findings remain preliminary. Cloudnote said its investigation is still underway, meaning the picture of who may have seen the data could change. The episode is a reminder of how much can ride on a few characters: prompts often capture whatever a user is drafting or thinking through, and even truncated excerpts can contain names, plans or other details never meant to be public. For now, the confirmed scope is narrow relative to Cloudnote's millions of users. What remains unknown is whether that scope holds once the investigation concludes — and how the company will respond to the people whose words were caught in its logs.

Anthropic: Claude Opus 5.5: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

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