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

Anthropic: Claude Opus 5.5: # HelioDock Launches Remote Fleet Software for Warehouse Robots, Raises $24 Million *The robot maker says Night Shift reduced unplanned stoppages during a 14-month beta, but one customer says its value drops when the network does* 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 funding round led by North Coast Ventures. The new product addresses a common problem for warehouses that run robots around the clock. When a machine stops unexpectedly, someone has to respond, often in person and often at inconvenient hours. HelioDock is betting that operators will pay a monthly fee per robot to handle more of those problems from a distance. Results from the company's beta point to a benefit, but a customer's experience shows that the software depends on the quality of each facility's network connection. "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. ## What the beta showed Night Shift spent 14 months in a private beta with nine logistics companies operating a combined 310 robots. HelioDock says those customers saw 17% fewer unplanned robot stoppages during that period. The figure has not been independently verified. It comes from HelioDock alone, so prospective buyers have only the company's word on how much the software reduces downtime. ## The connectivity question Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity. He said it became less useful during LTE outages at rural sites. HelioDock's design accounts for those outages, but only partly. When a connection fails, the robots return to the behavior configured locally at the facility, and remote control becomes unavailable. An outage therefore doesn't leave robots without instructions. It does remove the remote intervention that Night Shift is built to provide, at least until the connection comes back. That distinction matters most for operators in places where the network is least reliable. At a rural facility running on LTE, an outage that coincides with a stuck robot could still require the kind of late-night trip Velez says the software is meant to eliminate. In well-connected buildings, Pike's account suggests the product performs as intended. ## Pricing and availability Night Shift will cost $90 per robot per month and become generally available September 15. That works out to $1,080 per robot per year. At that rate, a fleet the size of the entire beta group, 310 robots, would cost about $27,900 a month. Because the price is set per robot, costs rise directly with fleet size. Operators weighing the product will likely compare that monthly expense against the cost of downtime and of sending staff to handle stoppages on site, the trade-off that the 17% claim is meant to address. ## Funding 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 investment. The funding arrives as HelioDock moves from selling robots alone to charging recurring fees for software that runs on top of them. Night Shift is the company's first subscription offering among the details it announced Thursday, and its per-robot pricing ties the company's software revenue to the number of machines its customers deploy. ## The competition HelioDock is not alone in selling remote oversight of robot fleets. Relay Robotics and YardOS both offer competing fleet-management products, so Night Shift will have to win over customers who can choose among several options for keeping watch on their machines after hours.

MoonshotAI: Kimi K3: # HelioDock Launches Night Shift, Software for Managing Warehouse Robot Fleets Remotely *Tested for 14 months across nine logistics companies and 310 robots, the platform reaches general availability September 15 at $90 per robot per month — and arrives alongside a $24 million Series B.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely and intervene when a machine runs into trouble, without anyone having to be in the building. The pitch speaks to one of the least glamorous realities of warehouse automation: robots stall, often at inconvenient hours, and getting them moving again has typically meant sending a person. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," said CEO Mara Velez. HelioDock's answer is to move that intervention off the warehouse floor entirely, shifting oversight from whoever is nearest the robot to an operator who could be anywhere. ## Fourteen Months in Beta The launch follows a 14-month private beta involving nine logistics companies and a combined fleet of 310 robots, giving the software more than a year of use in live logistics operations before its commercial debut. HelioDock says customers in that program experienced 17% fewer unplanned robot stoppages — a company-supplied figure that has not been independently verified. Unplanned stoppages, in which a robot halts mid-operation and waits for help, are precisely the failures Night Shift is built to catch, and HelioDock credits the software with the improvement. If the number holds up, it would translate directly into fewer of the late-night scenarios Velez described. 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. ## The Connectivity Catch Pike's caveat cuts to the product's central dependency. Night Shift's monitoring and remote-control features require a live connection to each machine, and LTE refers to the cellular data networks on which many out-of-the-way facilities depend. When connectivity fails, the robots do not simply shut down: they return to their locally configured behavior and keep working under their own settings. But remote control becomes unavailable until service is restored, leaving an off-site operator with nothing to monitor and no way to step in. For warehouses in areas where cellular coverage is unreliable — the scenario Pike described — that fallback means the software's most distinctive features are the first thing to go when the network does. ## Pricing, Funding and Competition Night Shift becomes generally available on September 15, priced at $90 per robot per month, a subscription structure that scales with the size of an operator's fleet. At that rate, running the software across a deployment the size of the full beta fleet — 310 robots — would cost $27,900 per month. HelioDock paired the launch with a funding announcement: a $24 million Series B led by North Coast Ventures, which brings the company's total funding to $39 million. The new capital lands as HelioDock expands from building robots into managing them — a category where it will not be alone. Relay Robotics and YardOS both offer competing fleet-management products, so warehouse operators weighing Night Shift will have alternatives on the table from the moment it launches, and HelioDock's per-robot pricing and connectivity requirements will be measured against them.

Anthropic: Claude Opus 5.5: # The Top Two Support AIs Tied on Quality. One Costs Less Than Half as Much. A new benchmark of five language models on real customer-support tickets has one clear result. The two highest-scoring models performed about equally well, but one was far cheaper and faster than the other. Model North had the highest average score, 84.2. Model Vale scored 83.8. The researchers found that this 0.4-point gap was **not statistically significant**, so the test does not show that either model is better overall. The practical differences were large, though. Vale's estimated cost was $0.74 per 1,000 tickets, compared with $1.90 for North, which means Vale cost about 39% as much. Vale also answered 18% faster. For teams choosing a model to draft support replies, the test suggests that paying more for North did not buy measurably better answers. ## How the test worked The researchers used 200 anonymized customer-support tickets. Each of the five models wrote one proposed answer per ticket. Three experienced support managers then graded every answer without knowing which model wrote it. The scores combined three qualities: - **Factual correctness:** whether the answer was accurate - **Usefulness:** whether it would help the customer - **Tone:** whether it sounded appropriate for a support exchange Blind grading by practitioners is a strength of this design. The graders were the kind of people who review support replies in real work, and they couldn't favor a model they recognized. Combining the three qualities into a single score has a cost, however. A model that is accurate but blunt could score about the same as one that is warm but occasionally wrong, and those are very different risks in production. ## Differences matter more down the list Below the top two, the gaps grow: | 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 table shows a familiar pattern: cheaper models generally scored lower. Ember and Kite cost a fraction of Vale's price and fell a few points behind it. Harbor was by far the cheapest and trailed North by about 13 points. Whether these lower-cost options are good enough depends on the job. A team that wants rough drafts for human agents to edit might accept a lower score. A team that sends replies straight to customers probably would not. The cost figures also need context. They are estimates, and at these rates the absolute dollar amounts are small. For a team handling a few thousand tickets a month, the gap between North and Vale is a rounding error next to staff salaries. The ratio matters more at very high volumes, or where the system makes several model calls per ticket. ## Each top model had a specialty Overall averages can hide differences by ticket type. North performed best on billing disputes, and Vale performed best on technical troubleshooting. A support operation dominated by one kind of request might reasonably prefer the model that led in that category. These results come with a caution. The 200 tickets were split across categories, so each category contained fewer tickets than the full set. The researchers did not report whether these category differences were statistically significant. They are worth a closer look in a company's own testing, but they do not establish that either model is reliably better at a ticket type. ## A blind spot every model shared The most consistent weakness had nothing to do with price. **All five models performed poorly when a ticket depended on an image attachment**, such as a screenshot of an error. Customers send screenshots often, especially for technical problems. This result suggests that teams should keep humans responsible for image-dependent tickets, or at least flag them, whichever model they choose. ## What the benchmark doesn't show The study's scope limits how far its 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 products, other industries or customers with different writing styles. - **No agentic tasks.** The researchers did not test whether models could take actions, such as issuing refunds or updating accounts, or use tools like knowledge-base search. Many real support deployments depend on these abilities. - **No multilingual support.** Companies serving non-English-speaking customers will need their own evidence. - **A modest sample.** Two hundred tickets with one answer per model is enough to separate the strongest and weakest models. It is not enough to rank close competitors with confidence, which is why the North–Vale gap could not be resolved. ## The takeaway Within its scope, the benchmark gives a fairly clear answer to a common question. You don't have to pay for the top-ranked model to get top-tier support drafts. Vale matched North's quality within the test's margin, cost much less and responded faster. Organizations should still test candidate models on their own tickets before switching. They should pay particular attention to their most common ticket types and to any requests that depend on screenshots, since no model in this test handled those well.

MoonshotAI: Kimi K3: # Tied on Quality, Miles Apart on Price: Five AI Models Face 200 Real Support Tickets When researchers put five language models to work answering real customer-support tickets, the clearest result wasn't which model won — it was how little separated the two leaders, and how much more one of them charges for that tie. Model North earned the highest average score, 84.2. But Model Vale scored 83.8, a gap of just 0.4 points that the researchers found was **not statistically significant**. In practical terms, the two models performed indistinguishably. What is distinguishable is the bill: North's estimated cost is $1.90 per 1,000 tickets, while Vale's is $0.74 — roughly 60% cheaper. Vale also produced its answers 18% faster. For a support operation processing a million tickets a year, that works out to about $1,900 versus $740 for essentially the same quality, delivered more quickly. The "best" model in this test is, by the numbers, a tie — and one of the tied models costs far less. ## How the test worked The researchers assembled 200 anonymized customer-support tickets and had each model produce one proposed answer per ticket. Three experienced support managers then graded every response blindly — they didn't know which model wrote which answer — scoring on a combination of factual correctness, usefulness, and tone. That design matters. Blind grading by working managers reduces the brand bias that can creep into vendor-run demos, and scoring tone alongside accuracy reflects how support actually works: a correct answer delivered rudely is still a bad answer. At the same time, a 200-ticket sample graded by three people is a modest foundation, and the blended score can't reveal whether models differed more on accuracy or on style. ## Where the field actually separated Below the top two, the gaps widen. Model Ember averaged 79.1 at $0.31 per 1,000 tickets; Model Kite scored 76.4 at $0.22; and Model Harbor, the cheapest option at $0.08, trailed at 71.0. The step down from Vale to Ember is more than ten times the North–Vale gap, which suggests the real decision point isn't between the leaders but between the top tier and the budget tier. Harbor's score shows that cost-cutting eventually stops being free: the cheapest model saved about 90% versus Vale while giving up nearly 13 points of quality. One caveat: the researchers reported significance testing only for the top pairing, so the gaps in the middle of the table are suggestive rather than proven. The leaders also showed different strengths. North performed best on billing disputes, while Vale led on technical troubleshooting. For teams whose ticket mix skews one way, that could serve as a tiebreaker — though category-level results from a modest sample shouldn't be treated as settled fact. ## The weakness no model escaped Every model struggled when a ticket depended on an image attachment. That's a practical warning: screenshots of error messages, photos of damaged shipments, and scanned invoices are common in support queues, and on this evidence none of the five can reliably handle them. Teams deploying any of these models should route image-dependent tickets to humans or to vision-capable systems rather than assuming coverage. ## What this benchmark can't tell you The test drew exclusively on English-language tickets from a single software company. Results might not transfer to retail, telecom, or healthcare support, let alone to other languages. And the models were asked only to draft answers: the researchers didn't test agentic actions, tool use, or multilingual support, so this says nothing about whether a model can actually issue a refund, look up an order, or resolve a ticket end to end. Finally, each model got one shot per ticket. Real support is often a back-and-forth conversation, and single-turn performance is at best a proxy for it. ## The takeaway The headline here isn't a winner — it's the compression at the top. The two best models are effectively tied on quality, so paying well over double for one of them is hard to justify on these results. The more useful questions for buyers are simpler: Which model handles your dominant ticket type best? And have you carved out the image-heavy tickets that none of them can handle?

Anthropic: Claude Opus 5.5: # Foundry Hall to Close After Losing Its Largest Sponsor *The Milwaukee startup accelerator, which has supported 212 companies since 2018, will shut its doors October 31.* Foundry Hall, the Milwaukee nonprofit that has helped hundreds of young companies find their footing, 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 operate through the end of October but cannot fund another full program. Six employees will lose their jobs, and the twelve startups now working out of its space will have to leave. Executive director Camille Ross said the organization had hoped another funder would step in to fill the gap before it came to this. "We kept waiting for the replacement check that never came," she said. In a statement, the bank said its decision reflected a shift in its community investment priorities and was not a judgment on Foundry Hall's performance. By the accelerator's own numbers, that performance was substantial. Since opening in 2018, Foundry Hall has supported 212 startups, which together employ 1,140 people. For a nonprofit of its size, those figures represent a meaningful share of the city's early-stage business activity, and they reflect years of mentoring, introductions, and the simple provision of a place to work. That last part is easy to overlook. For many founders, the value of an accelerator lies less in any single workshop than in having an address, a desk, and a room where serious conversations can happen. Dev Malik remembers it that way. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." The twelve companies still in residence now face the more immediate problem of finding somewhere else to go. Deputy Mayor Susan Cole said the city is in talks with building owners and local universities about space for them. She did not describe a timeline or specific locations. The closing raises a familiar question for nonprofits that depend heavily on a single benefactor. When one sponsor provides more than a third of the budget, its departure can leave an organization with no margin to adjust, regardless of how well its programs are working. Foundry Hall's outcomes did not protect it from a decision made for reasons unrelated to those outcomes. For the six staff members, the end comes with weeks of notice rather than months. For the startups, it means packing up in the middle of their work. Ross did not suggest that Foundry Hall might be revived under a new arrangement, and there was no indication that a rescue was in the works. What remains, for now, is a fixed date and a short stretch of time before it arrives. On October 31, the building where Malik once took his first meeting will close. The companies that would have followed him there will have to start somewhere else.

MoonshotAI: Kimi K3: # Foundry Hall to close after seven years, leaving a dozen startups without a home *The Milwaukee accelerator, which supported 212 startups, lost the sponsor that covered more than a third of its budget.* Foundry Hall, the nonprofit startup accelerator that has spent seven years helping Milwaukee founders turn side projects into companies, will close October 31 after its largest corporate sponsor declined to renew its funding. Executive Director Camille Ross said the organization has enough money to operate through October but not enough to run another full program. Blank River Bank had supplied 38 percent of the accelerator's annual budget, and in the months after the bank's decision, no other funder emerged to take its place. "We kept waiting for the replacement check that never came," Ross said. Nonprofit accelerators run on a patchwork of sponsorships, grants and donations, and a lead sponsor covering more than a third of the budget is not the kind of loss that can be replaced on a deadline. The closure ends a run that began in 2018, when Foundry Hall opened its workspace to a first class of founders. Since then, it has supported 212 startups, which collectively employ about 1,140 people — a small but real share of the city's economy, assembled one pitch meeting at a time. For founders, the building itself was part of the offering: a professional address, meeting space and a room where investors took them seriously. Dev Malik, whose company came through the program, is one of 212 founders who could tell some version of that story. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." Blank River Bank said the decision reflected "a change in our community investment priorities" and was unrelated to Foundry Hall's performance. The bank gave no indication it would reconsider. The consequences now are concrete. Six employees will lose their jobs as the organization winds down. Twelve startups currently using the workspace must relocate by the end of the month. Deputy Mayor Susan Cole said the city is speaking with building owners and local universities about space for the displaced companies, though no arrangement has been announced, and time is short. What remains is a wind-down: staff spending their final weeks helping the current companies land somewhere new — the last round of introductions for an organization that built its mission on making them. For Malik, the loss is simpler than budgets and sponsorships. After October, Milwaukee will have one fewer place where a founder can stop pitching from the back table of a coffee shop.

Anthropic: Claude Opus 5.5: # Arcwell's Relay Puts an AI Assistant Beside Customer Support Agents Arcwell is betting that customer support agents waste much of their day looking for answers their employers already have. The company has announced Relay, a platform that uses artificial intelligence to examine customer support conversations and suggest answers to the agents handling them. Relay can also flag customers who may be about to cancel and generate summaries, which puts it in direct competition with a growing number of AI tools aimed at the support desk. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. That pitch comes with one hard number. Twelve companies tested Relay over six months, and Arcwell says agents using it closed tickets 14% faster. The figure is the company's own, and it covers only part of the story. Arcwell did not release customer-retention results, so there is no public evidence yet on whether Relay's ability to spot customers likely to cancel actually helps keep them. The distinction matters to buyers weighing the cost. Relay is priced at $65 per agent each month, or $6,500 a month for a 100-person support team. Faster ticket handling can justify part of that expense, but businesses may want proof that the retention feature works before counting on it to pay for the rest. Relay also asks customers for a significant amount of trust. To function, the platform needs access to support transcripts and internal help documents. Those records often include customer details and institutional knowledge that companies would rather keep in-house. Arcwell says it encrypts customer data and does not use it to train shared models. That second assurance addresses a common worry among businesses adopting AI tools: that their information could end up improving a system other clients also use. Companies can also choose a 30-day or 90-day retention period for their data. Those safeguards will likely shape how quickly Relay gets past the pilot stage. A 14% improvement in ticket speed is a concrete selling point. Whether companies will hand over their support histories in exchange, and whether Relay can show it keeps customers from leaving, will decide how far the product goes.

MoonshotAI: Kimi K3: # Arcwell Launches Relay, an AI Assistant for Customer Support Teams Arcwell has launched Relay, an artificial intelligence platform that analyzes customer support conversations and suggests answers to the agents handling them. The platform is built around three capabilities. It suggests answers agents can use in customer conversations. It flags customers whose messages indicate they may be preparing to cancel, giving companies a window to respond. And it generates summaries of support interactions, so the substance of a long exchange can be reviewed without rereading the full thread. For chief executive Priya Shah, the problem Relay sets out to fix is not a shortage of information but the time agents lose hunting for it. "Support agents spend too much time searching for information that their companies already have," she said. Relay's answer is to bring that information to the agent rather than send the agent looking for it. Twelve companies tested Relay over a six-month pilot, and agents using the platform closed support tickets 14 percent faster, according to Arcwell. That figure comes from a small sample. Notably, the company did not release results on customer retention, leaving open the question of whether Relay's cancellation warnings actually help businesses keep the customers it flags — a conspicuous gap for a product that counts churn detection among its headline features. Relay is priced at $65 per agent per month, so costs scale directly with the size of a support team. None of it works without access. Generating suggestions requires Relay to read a company's support transcripts and internal help documents. Arcwell says customer data is encrypted and is not used to train shared models, meaning one customer's conversations never inform another's answers. Companies can choose to retain their data for either 30 or 90 days. Whether the 14 percent improvement holds up beyond a 12-company pilot — and whether faster tickets translate into customers who stay — are questions Arcwell will have to answer with data it has not yet shared.

Anthropic: Claude Opus 5.5: # KiteBridge Glitch Sends $2.8 Million in Duplicate Reimbursements, Months After Promise of "Exactly-Once" Payments Ninety-one companies paid out roughly $2.8 million they did not owe on Tuesday after a software update at KiteBridge, a maker of expense-management software, issued the same employee reimbursements twice. The error produced 13,400 duplicate payments in a little over an hour. By Wednesday evening, $2.3 million had been recovered. About $500,000 remained pending. The failure is especially awkward for KiteBridge because of a promise it made in April. A changelog for version 4.0 said the release had introduced "exactly-once payment protection," a safeguard meant to ensure that each reimbursement is paid only once, even when systems falter. ## How the duplicates happened KiteBridge's software handles the process by which employees submit work expenses and are paid back by their employers. On Tuesday at 8:45 a.m., the company deployed version 4.2 of its software. The problem began with a timeout at a payment processor, the outside service that actually moves money between bank accounts. A timeout occurs when a system waits for a response that does not arrive in time. Payment software typically responds by retrying the request, on the assumption that the first attempt may not have gone through. Those retries are supposed to be recognized as repeats of the original payment. Instead, KiteBridge's fallback system, the backup logic that takes over when the normal process breaks down, treated each retry as a brand-new reimbursement request. In many cases the original payment had already gone through, so employees were paid a second time. The first customer complaint reached KiteBridge at 9:12 a.m., less than half an hour after the update went live. The company suspended automated payments at 10:03 a.m., about 50 minutes after that complaint and 78 minutes after the deployment. By then, 13,400 duplicate reimbursements had been issued. ## Who bore the cost The duplicate payments were funded by employers, not employees. No employees had money withdrawn from their accounts as a result of the error, so the immediate financial exposure fell on the 91 corporate customers whose accounts paid out the extra reimbursements. Recovery efforts moved quickly. Within roughly a day and a half, more than 80 percent of the money had been returned. The remaining sum was still pending as of Wednesday evening. ## A slow public disclosure KiteBridge's handling of public communication drew attention alongside the technical failure. The company's public incident page, where software firms typically post updates during outages, initially described the problem only as delayed reimbursements. That description reflected the effect of the suspension: with automated payments halted, legitimate reimbursements were held up. It did not mention that thousands of payments had already gone out twice. The incident page first disclosed the duplicate payments six hours after automation was suspended, around 4 p.m. Tuesday. Until then, customers relying on the page would not have learned the full nature of the problem from it. ## Company response KiteBridge chief executive Lena Ortiz apologized for the incident and said the company would publish a technical report. That report may address the central question raised by Tuesday's failure: how a system advertised as preventing duplicate payments came to generate thousands of them. The company's April changelog described exactly-once protection as part of version 4.0. The duplicates occurred after the deployment of version 4.2, through a fallback path that handled retries as new requests.

MoonshotAI: Kimi K3: # KiteBridge Software Bug Triggers $2.8 Million in Duplicate Reimbursements Expense-management software maker KiteBridge is working to recover $2.8 million in duplicate reimbursements after a faulty software update on Tuesday caused its system to pay thousands of expense claims twice, hitting 91 corporate customers. No employees had money withdrawn from their accounts — the duplicate payments were funded by employers, meaning workers briefly saw extra reimbursement deposits land in their bank accounts. But for the companies involved, the error created an unwelcome accounting mess. As of Wednesday evening, KiteBridge said $2.3 million had been recovered, with the remaining roughly $500,000 still pending. The failure began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its software. The first customer complaint arrived less than half an hour later, at 9:12 a.m. The company suspended automated payments at 10:03 a.m., but by then the software had generated 13,400 duplicate reimbursements. The root cause lay in how the system handled a payment-processor timeout. When KiteBridge's software sends a reimbursement to an outside payment processor and doesn't receive a confirmation in time, a fallback system retries the request — a routine safeguard, since network hiccups are common. The risk is that the original request actually succeeded, and a blind retry pays the same claim twice. Well-designed payment systems guard against this with a technique known as idempotency: attaching a unique identifier to each transaction so repeats can be recognized and discarded. KiteBridge's fallback system instead treated each retry as an entirely new reimbursement request, and the duplicates flowed through. The incident is especially awkward for the company because of promises it made earlier this year. An April changelog for version 4.0 advertised the introduction of "exactly-once payment protection" — industry shorthand for precisely the guarantee that a payment will be processed a single time, never more. Whatever protection that feature provided apparently did not extend to the fallback path triggered on Tuesday. KiteBridge's public communication also lagged behind events. Its incident page initially described the problem as delayed reimbursements — a characterization that omitted the far more serious issue. The page did not disclose the duplicate payments until roughly six hours after automation was suspended, putting the first public acknowledgment in the late afternoon. Recovering erroneous payments typically involves reversing transfers where possible and requesting the return of funds where it isn't, a process that can stall depending on bank cooperation and how quickly recipients respond. KiteBridge has not said how long it expects the remaining recovery to take. CEO Lena Ortiz apologized for the incident and said the company would publish a technical report explaining what went wrong. That report is likely to face close scrutiny. Customers will want to know not only why the fallback system malfunctioned, but why a safeguard announced in April failed to prevent the duplicates — and why the company's first public account of the incident understated what had actually happened. For the 91 affected companies, the immediate question is simpler: when the rest of the money comes back.

Anthropic: Claude Opus 5.5: # Startup Spotlight: LedgerLark **A young company is betting that AI agents can take on the investigative grind of post-trade operations, as long as people keep the final say.** Mina Batra once watched a six-person team spend three days resolving a single trade mismatch. The cause turned out to be a copied account number. For anyone outside financial operations, that ratio of effort to error can seem absurd. For people inside it, the story is familiar. After a trade is executed, records have to agree across several places: the trade confirmation, the custodian's books and the firm's own internal ledger. When they don't, someone has to figure out why. That work is often slow and manual, and it rarely gets attention until something goes wrong. Batra, who spent seven years in operations at a regional brokerage, began exploring the problem after that episode. She co-founded LedgerLark with Owen Carr, who previously built reconciliation systems at a large bank. Their combined backgrounds cover both sides of the problem: the daily experience of chasing exceptions and the engineering of systems meant to catch them. ## What the product does LedgerLark builds AI agents aimed at post-trade operations at small investment firms. The agents compare confirmations, custodian records and internal ledgers, flag where they diverge, and propose corrections. Those proposals go to a human for approval. The product does not execute trades or move money. That boundary matters. It limits what can go wrong if an agent misreads a record, and it may make the tool easier to explain to compliance teams wary of automated systems acting on their own. The focus on smaller firms is also notable. Large institutions can build or buy heavyweight reconciliation infrastructure. Smaller shops may face the same kinds of mismatches with fewer people and less tooling to absorb them. ## Early traction LedgerLark says 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 is promising, but it should be read with care. It comes from the company, it describes one customer, and it doesn't say on its own how much of the improvement came from the software versus other changes, or whether the result holds across firms with different setups. Five customers is a meaningful start for a seed-stage company. It is not yet a broad sample. LedgerLark has raised a $3.5 million seed round led by Motive Row. ## The harder questions LedgerLark is entering a field that already has occupants. Established reconciliation vendors have long-standing relationships with financial firms, and many banks run their own internal software. The startup's case rests partly on whether AI agents that investigate and propose fixes offer something meaningfully different from rule-based matching tools. Its case also depends on whether smaller firms are underserved by existing options. Then there is the pace of selling into finance. Financial institutions may require lengthy security and procurement reviews, particularly for software that touches sensitive account and transaction data. Even with the human-approval design, a young company may find that sales cycles stretch longer than its runway would prefer. Accuracy is another open question for any AI tool in this setting. A proposed correction is useful only if reviewers can trust it, or at least quickly verify it. If staff end up double-checking every suggestion from scratch, the time savings shrink. ## What to watch The next chapter for LedgerLark will likely turn on a few things: - Whether results like the reported drop in exceptions repeat across more customers. - How quickly it can clear security and procurement reviews. - Whether it can carve out a clear position next to incumbent vendors and in-house systems. The underlying observation is hard to dismiss: small errors in financial records can consume a surprising amount of skilled human time. Whether LedgerLark is the company that meaningfully reduces that burden is still an open question.

MoonshotAI: Kimi K3: # Startup Spotlight: LedgerLark Bets AI Agents Can Untangle Post-Trade Work Trade mismatches rarely make headlines, but they eat a surprising share of the workday at small investment firms. LedgerLark, founded by Mina Batra and Owen Carr, is building AI agents designed to find those errors and help fix them before they pile up. ## A three-day error Batra spent seven years in operations at a regional brokerage, where she says the job was defined less by judgment than by drudgery: matching records across systems that didn't quite agree. The idea for LedgerLark took shape after she watched a six-person team spend three days resolving a single trade mismatch traced to a copied account number. The error itself was trivial, she says; finding it was not. Carr, her co-founder, built reconciliation systems at a large bank, giving the pair a view of the problem from both ends of the market. The insight the founders draw from experiences like that one: most firms don't struggle to *detect* mismatches—existing systems already surface exceptions by the dozens. The costly step is investigation, tracing an exception to its cause and drafting a fix. That is the work LedgerLark's agents are meant to compress. Reconciliation—the daily task of confirming that a firm's books match what counterparties and custodians report—is mature software territory. LedgerLark's argument is that small firms running lean operations teams are underserved: incumbent tools can be costly to install and rigid in what they flag. ## What the agents do "Agent" covers a lot of ground these days. In LedgerLark's case, it means software that follows an exception across systems—comparing trade confirmations, custodian records and internal ledgers—then assembles a proposed correction for a human to approve. The company is explicit about what the product does not do: it does not execute trades or move money. That boundary matters, both for regulatory reasons and for customer trust, and keeping a person in the approval loop is likely to remain central to how the product is sold. ## Early traction Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The company says one customer cut its unresolved daily exceptions from 70 to 18 after adopting the product. If that result held across a broader base, it would represent a meaningful reduction in operational drag—but the figure comes from LedgerLark itself and reflects a single customer. Outcomes will depend on data quality and how exception-heavy a firm's book is. The company recently closed a $3.5 million seed round led by Motive Row—useful runway, though not much cushion if sales cycles stretch. ## Competition, and the slow path to yes Rivals arrive from two directions. Established reconciliation vendors hold long-standing relationships with financial institutions, and several are adding AI-assisted features of their own. From the other side, larger banks often rely on software built in-house, while some smaller firms may decide their custodian's existing tools are good enough. LedgerLark's bet is that a focused product—aimed at firms too small for enterprise software and too busy for manual work—can hold durable ground between those poles. The harder test may be procurement rather than technology. Financial institutions typically subject new vendors that touch trade data to lengthy security and compliance reviews, which can stretch for months and strain a small team. Batra acknowledges the constraint and says the human-approval design and limits on agent access are partly a response to it. Whether AI agents earn a lasting place in post-trade operations remains an open question. The workflow LedgerLark targets is unglamorous, rules-heavy and expensive to get wrong—which is exactly what makes it a useful test of where agentic software will be trusted, and where firms will insist a person stays in charge.

Anthropic: Claude Opus 5.5: # Raincheck Raises $8 Million for Flood-Risk Software Used by Cities *The Austin company's tools combine drainage maps with rainfall forecasts to help local governments decide where to send 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 a seed round led by Basin Ventures. Existing investor Long Mile also participated, bringing Raincheck's total funding to $10.5 million. 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. Chief Executive Ana Flores said the challenge is less about identifying flood-prone areas than about prioritizing the work that could prevent flooding: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." More detail is available in [Raincheck's description of its methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. The software's forecasts serve as recommendations. Emergency managers retain control over deployments and public warnings. Raincheck competes with other flood-planning vendors, including HydroMap, as well as municipal engineering contractors. The company says the new funding will help it hire eight engineers and expand into Florida.

MoonshotAI: Kimi K3: # Raincheck Raises $8 Million to Help Cities Get Ahead of Flood Risk *The Austin startup maps drainage systems and pairs them with rainfall forecasts so local governments can deploy maintenance crews before the rain arrives.* Raincheck, an Austin-based company that builds flood-risk software for local governments, has raised $8 million in seed funding, the company announced today. Raincheck maps drainage systems and combines those maps with rainfall forecasts, an approach detailed in [the company's published methodology](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews ahead of a storm. Basin Ventures led the seed round, with participation from existing investor Long Mile. Raincheck has now raised $10.5 million in total. CEO Ana Flores said the software is designed to close the gap between knowing where flooding happens and knowing what to fix first: > "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; emergency managers retain control over crew deployments and public warnings. Raincheck competes with other flood-planning vendors, including HydroMap and municipal engineering contractors. The company said the new funding will allow it to hire eight engineers and expand into Florida.

Anthropic: Claude Opus 5.5: **6** 1. `#Flood software moves upstream` has no space after `#`, so it does not follow the `# Headline` form. 2. `The company is expanding its pilot.` is the dek, but it is not italicized. 3. `##Details` has no space after `##`, so it does not follow the `## Heading` form. 4. `Read [the announcement](www.example.com/news).` uses a link that is not an absolute HTTPS URL. 5. `* The first pilot begins Monday.` starts the list item with an asterisk instead of a hyphen. 6. `<div>The second pilot begins Friday.</div>` contains raw HTML.

MoonshotAI: Kimi K3: 6

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

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

Anthropic: Claude Opus 5.5: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 Sessions *The page required no login and was reachable for four hours on 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 Cloudnote debugging page that displayed fragments of customers' prompts was accessible without authentication for four hours on Monday, the company said, before a security researcher alerted it to the problem. According to Cloudnote, logs show the page held excerpts from 1,842 user sessions, each potentially containing up to 200 characters of a prompt. The company removed the page 19 minutes after the researcher's notification. Cloudnote has 3.2 million registered users. It has not said how many individual accounts the affected sessions represent. The exposure appears limited in scope. Full documents, passwords and payment information were stored in separate systems and did not appear on the page, the company said. Short prompt fragments can still matter, however. People often type names, work details or other sensitive material into such tools, and 200 characters is enough to contain it. Key questions remain open. Cloudnote says it has found no evidence that anyone besides the researcher accessed the page, but it has not completed its investigation. It is also not yet clear how a page meant for internal debugging came to be reachable without a login.

MoonshotAI: Kimi K3: # Cloudnote Left Customer Prompt Excerpts Readable Without a Login for Four Hours *Logs show text from 1,842 user sessions was exposed on Monday. The company says it has found no evidence anyone besides the researcher who reported it viewed the page — but its investigation is not finished.* Cloudnote said a debugging page containing excerpts of its customers' prompts was accessible without any login for four hours on Monday, exposing text from 1,842 user sessions before the page was taken down. A researcher discovered the page and alerted the company, which removed it 19 minutes later, according to Cloudnote's account of the incident. The company said it has found no evidence that anyone other than the researcher accessed the page while it was live. The page held short excerpts — up to 200 characters from a single prompt — rather than complete documents. Cloudnote said full documents, saved passwords and payment information are stored elsewhere and were not present on the exposed page. What was exposed is limited in volume but not necessarily in sensitivity. People routinely type names, unpublished drafts, business plans and other private material into prompt fields, and 200 characters is enough to capture a detail its author never meant to share. The 1,842 sessions are a small slice of Cloudnote's 3.2 million registered users, but the company has not said how many individual accounts those sessions represent. Cloudnote said its investigation is still underway. Among the unanswered questions: how the debugging page became publicly reachable in the first place, whether the users whose prompts appeared on it will be notified, and what the company's log review can — and cannot — rule out about who else may have seen it.

Anthropic: Claude Opus 5.5: C

MoonshotAI: Kimi K3: C

Anthropic: Claude Opus 5.5: # QueueZero's AI Meeting Assistant Assigns Its CEO a Task During Launch Demo **Proxy, which attends meetings and hands follow-up work to software agents, enters public beta next week at $30 per user per month.** QueueZero launched Proxy, an AI meeting assistant, on Wednesday. Its first public demonstration offered an unscripted look at how the software turns spoken comments into action. Proxy is designed to attend meetings, produce notes and assign follow-up work to software agents. During the launch demonstration, a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the comment as an instruction. It created two tasks assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. It was, in a narrow sense, the product doing exactly what it was built to do. ## Where Proxy draws the line The episode highlights a core challenge for software that converts conversation into work: distinguishing a directive from an offhand remark. Meetings are full of half-serious suggestions, rhetorical asides and comments aimed at the room rather than at any one person. A human note-taker usually filters those out. An automated system has to infer intent from words alone. QueueZero says Proxy requires confirmation before an external message is sent. Internal tasks, however, can be created automatically, which is what happened on stage. Administrators can disable automatic task creation. That split determines the cost of a misreading. In Wednesday's demonstration, the error stayed inside a task list, where it was visible and easy to remove. The confirmation requirement for external messages means a comparable mistake should not reach a client or partner without a person approving it. ## What buyers may want to consider For organizations evaluating Proxy, the demonstration suggests several practical questions: - **Default settings:** whether automatic task creation is on or off when Proxy is first deployed, and who in the organization controls that setting. - **Review process:** who is responsible for checking tasks the assistant creates, particularly tasks assigned to people who did not ask for them. - **Scope of confirmation:** exactly which actions count as "external" and therefore require approval. Teams that hold large meetings or meetings with outside participants may find the distinction especially relevant. In those settings, more of what gets said comes from people who are not directing the work. ## Pricing and availability Proxy costs $30 per user per month. It enters public beta next week.

MoonshotAI: Kimi K3: # QueueZero's AI Meeting Assistant Takes Audience Feedback Literally at Launch QueueZero on Wednesday launched Proxy, an artificial-intelligence assistant that can attend a meeting, produce notes and assign follow-up work to software agents. The event, staged in front of customers, doubled as an unscripted product test. During the demonstration, a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the comment as an instruction. It created two tasks and assigned them to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. By at least one measure, the demo succeeded: Proxy did exactly what it was told. The exchange drew laughter in the room, but it also highlighted the central design tension in a fast-growing category of workplace software: how much autonomy to grant tools that hear everything and are built to act on what they hear. It is the kind of edge case vendors rarely plan for — ambient remarks and rhetorical asides can sound like directives to a system trained to find action items in conversation. ## Guardrails QueueZero said Proxy operated within its intended limits, and that no external messages were sent during the demonstration. The assistant requires human confirmation before any external message goes out, meaning no pricing sheet reaches a customer without approval. Internal tasks, however, can be created automatically, which is how the two assignments landed in Wu's queue. Administrators can disable automatic task creation entirely. That split reflects a common approach among vendors building so-called agentic AI — systems that move beyond answering questions to taking actions on a user's behalf. The design keeps a person in the loop for anything that leaves the organization while letting software handle routine internal work on its own. ## Pricing and Availability Proxy arrives in a crowded field. AI notetakers have become fixtures of corporate video calls, and vendors are racing to differentiate by taking on more of the work meetings generate. QueueZero's bet is that routing action items to software agents, rather than leaving a to-do list in a document, will make its assistant stickier. Proxy costs $30 per user per month and enters public beta next week. The pricing sheet was sent.