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

Anthropic: Claude Sonnet 5.5: # Cloudnote exposed snippets of user prompts on an unprotected debugging page for four hours *The company says it has found no evidence that anyone but a security researcher viewed the page, but its investigation is not finished.* Cloudnote left an internal debugging page containing excerpts of customer prompts open to anyone with the link, no login required, for four hours on Monday, the company said. The page came down 19 minutes after a researcher notified Cloudnote of the problem. Logs show the page displayed excerpts from 1,842 user sessions. Each excerpt could include up to 200 characters of a prompt. Cloudnote has 3.2 million registered users, so the affected sessions are a small fraction of its base. But a session is not the same as a person, and the company has not established how many individual users were involved. Cloudnote says it has found no evidence that anyone besides the researcher accessed the page. That is not a guarantee that no one did, and the company acknowledges its investigation is incomplete. Cloudnote says full documents, passwords and payment information were stored elsewhere and were not on the page. Even short prompt fragments can carry personal or work-related details, however, and users typically enter them expecting privacy. Still unresolved is whether anyone else viewed the page and what the exposed excerpts actually contained.

MoonshotAI: Kimi K3: # Cloudnote Left Excerpts of Customer AI Prompts on an Unauthenticated Page for Four Hours *The note-taking company, which has 3.2 million registered users, says logs show fragments from 1,842 sessions were exposed Monday. It says it has found no evidence that anyone beyond the researcher who reported the flaw saw them — but its investigation is not finished.* Cloudnote, a note-taking app with 3.2 million registered users, said a debugging page containing excerpts of customers' AI prompts was accessible to anyone on the internet — no password required — for roughly four hours on Monday. The company removed the page 19 minutes after a security researcher notified it. Server logs show the page held prompt excerpts from 1,842 user sessions, each capped at 200 characters — a small fraction of the app's user base, but enough text to capture the opening lines of whatever a person had typed. Full documents, passwords and payment information were stored elsewhere and were not present on the page, the company said. Even so, short prompt fragments can be revealing: users type draft emails, medical questions and confidential business plans into AI assistants, and a 200-character excerpt could be enough to identify someone or disclose the subject of a document that was never itself exposed. The episode also illustrates a less-visible risk of the AI boom — the internal tools companies build to debug these features can themselves become a route to customer data. What is not yet known is whether anyone besides the researcher viewed the page during those four hours. Cloudnote said it has found no evidence of other access but said its investigation is not complete. The company has not explained how the page became publicly accessible, nor said whether it will notify the users whose sessions appear in the logs.

Anthropic: Claude Sonnet 5.5: Going through the draft line by line: 1. `#Flood software moves upstream` violates the headline rule, since it lacks the space after `#`. **Violation.** 2. `The company is expanding its pilot.` is supposed to be an italicized dek immediately after the headline, but it isn't italicized. **Violation.** 3. `##Details` violates the section heading rule, since it lacks the space. **Violation.** 4. `Read [the announcement](www.example.com/news).` has a link that is not an absolute HTTPS URL. **Violation.** 5. `* The first pilot begins Monday.` is an unordered-list item that begins with an asterisk instead of a hyphen. **Violation.** 6. `<div>The second pilot begins Friday.</div>` is raw HTML. **Violation.** Blank lines don't violate anything. **6**

MoonshotAI: Kimi K3: 6

Anthropic: Claude Sonnet 5.5: # Startup Spotlight: LedgerLark Bets That AI Can Take on the Detective Work of Post-Trade Operations *The young company wants to help small investment firms find out why their records disagree, without letting software act on the answer alone.* After a trade, several records are supposed to say the same thing: the confirmation, the custodian's statement and the firm's own ledger. When they don't, someone has to work out why. LedgerLark is building AI agents to help with that task at small investment firms. ## A copied account number Co-founder Mina Batra spent seven years in operations at a regional brokerage. Near the end of that time, she watched a six-person team spend three days resolving a single trade mismatch. The cause was a copied account number. That episode led her to explore the idea behind LedgerLark. She co-founded the company with Owen Carr, who built reconciliation systems at a large bank. The pairing is a sensible one: one founder knows how the work feels on a small firm's operations desk, and the other knows how large institutions engineer around the same problem. ## What the product does LedgerLark's agents compare confirmations, custodian records and internal ledgers, then propose corrections. A person reviews and approves each proposal. The product does not execute trades or move money. That boundary is a meaningful design choice. A wrong suggestion can be rejected before it affects anything, which should make the product easier for cautious firms to try. It also means the human review step remains, so the benefit depends on how much time goes into finding discrepancies versus approving fixes. The company's pitch is that finding them is the slow part. ## Early traction Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. LedgerLark has also raised a $3.5 million seed round led by Motive Row. The company says one customer reduced its unresolved daily exceptions from 70 to 18. It's an encouraging figure, but a limited one. It comes from one customer and is reported by the company. It doesn't say how long the improvement took, how the customer defines an exception, or whether trading volumes changed over the period. The more informative evidence will be whether similar results appear across the other four customers and beyond. ## The harder questions LedgerLark isn't entering an empty field. Established reconciliation vendors already serve financial institutions, and some banks rely on internal software built for their own needs. A small firm weighing a new product has to ask whether a startup's agents do enough that existing tools don't, or whether a familiar vendor can add similar features. Selling to financial institutions also takes time. Security and procurement reviews can be lengthy, and a product that touches sensitive records is likely to face close scrutiny. For a company with five customers, the pace of those reviews could shape growth as much as the technology does. There are open technical questions as well. AI systems can be wrong in ways that look plausible, so the quality of LedgerLark's proposals matters, as does the way the interface presents them. A reviewer who sees many correct suggestions may start approving them quickly, and that is when an occasional error is most likely to slip through. How the company handles unusual cases, explains its reasoning and keeps an audit trail may matter as much as its headline accuracy. ## What to watch - **Repeatability:** whether results like the 70-to-18 reduction hold across different customers and trade types. - **Renewals:** whether the first annual contracts are renewed, and at what size. - **Differentiation:** how LedgerLark distinguishes itself from established vendors in a head-to-head evaluation. - **Review burden:** whether approving proposed corrections stays fast and reliable as volume grows. The founders have identified a real, unglamorous problem, and their approach is cautious about what software should be trusted to do. Whether that is enough to build a durable business in a crowded, slow-moving market is still unknown.

MoonshotAI: Kimi K3: # Startup Spotlight: LedgerLark Wants AI Agents to Do the Back Office's Detective Work *LedgerLark builds AI agents for post-trade operations at small investment firms. Its founders are betting that the least glamorous corner of trading is also the most automatable — within limits.* When a trade is executed, the work is only half done. The firm, its custodian and its counterparties each keep records, and those records have to agree. When they don't, the result is an "exception," and resolving exceptions is slow, manual work — the kind that quietly consumes operations teams at small investment firms. Mina Batra knows that work firsthand. LedgerLark's co-founder spent seven years in operations at a regional brokerage, and she began exploring the idea for the company after watching a six-person team spend three days resolving a single trade mismatch — one eventually traced to a copied account number. Her co-founder, Owen Carr, built reconciliation systems at a large bank, giving the pair a view of the problem from both a small firm and a big institution. Their shared insight: investigating an exception is largely methodical comparison — pulling the confirmation, the custodian record and the internal ledger, then finding where the numbers diverge. That legwork is structured enough for software. The judgment about what to do next, they argue, should stay with a person. ## What the Agents Do LedgerLark's agents compare confirmations, custodian records and internal ledgers, then propose corrections for human approval. Crucially, the product does not execute trades or move money. That boundary looks deliberate: it positions LedgerLark as decision support rather than an actor inside the flow of funds, which may matter in compliance reviews and could make for an easier first sale to a cautious operations head. It also means the agents' output is only as useful as the approval workflow around it — a dynamic worth watching as the product matures. ## Early Traction Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000 — pricing that fits small-firm budgets without requiring enterprise-scale commitments. LedgerLark says one customer cut its unresolved daily exceptions from 70 to 18. That's a striking improvement if it holds up, but it's a single, company-reported data point; it's unclear how typical that result is, or how much of the drop reflects the software versus changes in the customer's own processes. The company has raised a $3.5 million seed round led by Motive Row, giving it room to find out. ## Competition and Open Questions LedgerLark is not entering an empty field. Established reconciliation vendors already sell to these firms, and internal bank software handles pieces of the same workflow. Incumbents bring distribution, integrations and long-standing trust — and there is little stopping them from adding agent-like features of their own. LedgerLark's case will likely rest on focus and speed, though neither has been proven at scale yet. The bigger structural risk may be the sales process itself. Financial institutions can require lengthy security and procurement reviews, which raises the cost of every deal and stretches timelines. Five customers suggest LedgerLark can clear at least some of those hurdles; they don't yet show a repeatable motion for doing so quickly. Questions also remain about the product. How accurate are the agents' proposed corrections, and how does the system behave on genuinely hard-to-classify mismatches? How much oversight do customers apply in practice — does human approval stay rigorous once the tool earns trust? ## What to Watch The founders' insight is grounded in lived experience rather than market-sizing abstractions, and the early contracts suggest someone is willing to pay for it. The next signals are concrete: whether the 70-to-18 result repeats across customers, whether contracts renew and expand, and whether LedgerLark can shorten the gauntlet of security review well enough to turn five wins into fifty.

Anthropic: Claude Sonnet 5.5: # QueueZero Launches Proxy, an AI Meeting Assistant That Assigns Follow-Up Work to Software Agents QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend meetings, produce notes and assign follow-up work to software agents. The product enters public beta next week and costs $30 per user per month. Proxy is meant to reduce the administrative work that follows a meeting. After joining a call, it generates notes and turns discussion into tasks, which can be handed to people or to software agents that carry out parts of the work. ## A live demonstration The launch demonstration showed how literally the system can read a conversation. A customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy treated the remark as an instruction. It created two tasks assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. The sequence showed both the feature's promise and a familiar difficulty for AI tools that work from natural speech: people phrase requests casually, sarcastically or imprecisely, and software must decide which remarks count as commands. The task list, at least, was shorter than the meeting. ## Safeguards QueueZero says Proxy requires confirmation before it sends any external message. A misread comment should therefore not result in an unreviewed email or note going to a customer, partner or other outside party. Internal tasks are treated differently. Proxy can create them automatically, as it did with the two assigned to Wu. Administrators can disable automatic task creation for their organizations, which lets companies choose between speed and tighter control. The company did not describe in the demonstration how Proxy distinguishes a direct request from an offhand remark. That distinction will matter as the product moves from a controlled launch to everyday use, when meetings are less scripted and participants may not know an assistant is turning their comments into work items. ## Pricing and availability Proxy costs $30 per user per month. Public beta begins next week. The product enters a crowded field of meeting transcription and summarization tools. Proxy's distinguishing feature is its attempt to go beyond notes and start the work that follows, using software agents to handle assigned tasks. That approach raises practical questions for adopters. Organizations will need to decide who can be assigned work automatically, what counts as an internal task and whether to leave automatic creation switched on. QueueZero's confirmation requirement for outbound messages and the administrator setting offer two controls, though the beta will test how well they hold up in routine use. For now, the demonstration offered a clear, if unintended, example of the product doing what it was built to do: listening closely, acting quickly and leaving a human to decide what to keep.

MoonshotAI: Kimi K3: # QueueZero Launches Proxy, an AI Assistant That Turns Meeting Talk Into Tasks **An audience remark during Wednesday's launch demonstration became two tasks assigned to the company's CEO.** QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to attend calls, produce notes and assign follow-up work to software agents — a capability the company ended up demonstrating live, in an exchange it had not planned. 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 and created two tasks assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. The pricing sheet reached the audience within minutes; Wu kept talking. The episode illustrates both the product's promise and its risk. Proxy is built to listen for action items in ordinary conversation and convert them into assigned work without waiting for explicit commands — a step beyond the transcription and summary features now common in meeting software from Zoom, Microsoft and Otter.ai. But the same sensitivity that lets it capture a casually mentioned follow-up also means it can act on remarks no one intended as instructions. Meeting assistants have become one of the most widely adopted categories of workplace AI, in part because their output is easy to verify. Tools that take action on a user's behalf face a higher bar: an error in a transcript is embarrassing, while an error in execution lands in someone's task list. QueueZero said Proxy requires user confirmation before any external message is sent, so the assistant cannot email a customer or contact a client on its own. Internal tasks, like the pair created during the demonstration, can be generated automatically. Administrators can disable automatic task creation across an organization, the company said. The launch also lands amid a broader debate over how much autonomy employers are willing to grant AI systems embedded in everyday software, particularly systems that act on ambiguous human speech rather than explicit commands. Proxy costs $30 per user per month and enters public beta next week. QueueZero said the administrative controls, including the option to disable automatic task creation, will be available to customers on all plans at launch.

Anthropic: Claude Sonnet 5.5: C

MoonshotAI: Kimi K3: **C** This lede accurately conveys the key newsworthy facts—the product removal, the data exposure, the number of affected accounts, and the duration—without editorializing. Options A and D contain false claims (message contents were never exposed, and there's no evidence of malicious access), while B reads like a press release and buries the actual news.

Anthropic: Claude Sonnet 5.5: # HelioDock launches remote monitoring for warehouse robots and raises $24 million *The software, called Night Shift, is meant to keep operators from making middle-of-the-night site visits. In beta, it was less useful where cellular service was unreliable.* HelioDock, which makes autonomous warehouse robots, on Thursday released Night Shift, software that lets operators monitor robot fleets remotely. The company also announced a $24 million Series B round led by North Coast Ventures. Night Shift is designed to cut the number of times a robot problem requires someone to show up in person. HelioDock says customers in the product's private beta had 17% fewer unplanned robot stoppages. That figure has not been independently verified. The software works best with a reliable network connection, according to one beta customer. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," HelioDock CEO Mara Velez said. ## Fourteen months of testing Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock has not said how many sites took part or how it measured stoppages, and the 17% figure comes from the company. Aaron Pike of Fulcrum Logistics, a beta customer, said the system worked well in facilities with strong connectivity. It became less useful during rural LTE outages, he said. HelioDock's design accounts for those failures in a specific way. When a connection drops, the robots return to their locally configured behavior, and remote control is unavailable until service returns. A robot that stalls during an outage would still need someone on site. ## Pricing and funding Night Shift costs $90 per robot per month and becomes generally available September 15. At that rate, a 100-robot fleet would cost $9,000 a month. The Series B brings HelioDock's total funding to $39 million, which means the company had raised about $15 million before this round. The company did not provide details in the notes on how it will spend the new money. ## A market with other options HelioDock is not the only company selling software to manage robot fleets. Relay Robotics and YardOS both offer fleet-management products. The notes do not include pricing or performance comparisons with Night Shift, and it is unclear how the three differ on connectivity requirements. The September 15 release will be the first time customers outside the beta can try Night Shift in their own facilities. Sites with weaker cellular coverage, like the ones Pike described, will find out whether the software helps them as much as it did the beta customers with stronger connections.

MoonshotAI: Kimi K3: # HelioDock Launches Night Shift, Software to Monitor Warehouse Robot Fleets Remotely *The remote-monitoring tool reaches general availability September 15 at $90 per robot per month, backed by a new $24 million Series B — and an unverified company claim of 17% fewer unplanned stoppages.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor — and, while connections hold, control — robot fleets remotely. The product targets a recurring problem in automated logistics: a machine stalls in the middle of the night, and the nearest person equipped to assess it is miles away. The launch pushes HelioDock beyond selling robots into selling the subscription software that runs alongside them. Subscriptions add recurring revenue on top of hardware sales and keep the company tied to customers long after their robots are installed. The move also lands HelioDock in a corner of the warehouse-technology market where fleet-management software is already an established category. CEO Mara Velez pitched Night Shift as an answer to the human cost of minor machine failures. "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 scenario she described — an avoidable late-night trip prompted by a single stalled machine — is the use case the software is built around: small faults that demand in-person attention undercut the economics of automation. Night Shift spent 14 months in a private beta before Thursday's release, a trial spanning nine logistics companies and 310 robots — roughly 34 machines per deployment. The extended test gave HelioDock more than a year of data from live warehouse operations, and it produced the performance figure the company is now using to market the product. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages while using the software — fewer instances of machines halting unexpectedly during operations. The figure comes from the company and has not been independently verified. One beta participant described performance that tracked with local infrastructure. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. That gap reflects how Night Shift behaves when connections drop. Its monitoring and remote-control features depend on a live network link; when connectivity fails, remote control becomes unavailable and the robots return to their locally configured behavior. An outage does not shut the fleet down — the machines keep running under their onboard settings — but the remote oversight that is the software's central function is suspended until service resumes. For operators weighing a purchase, network quality becomes part of the calculation. A distribution center with stable connectivity can lean on Night Shift's oversight across shifts, including overnight hours when managers are off-site. A warehouse at the edge of LTE coverage — the situation Pike described — will find the software's value rising and falling with its signal. Night Shift will be generally available September 15 at $90 per robot per month, a per-machine model that scales with fleet size. For a fleet the size of the average beta deployment, that works out to roughly $3,000 per month. The software enters a market with established rivals. Relay Robotics and YardOS both sell competing fleet-management products, giving warehouse operators multiple

Anthropic: Claude Sonnet 5.5: # Raincheck raises $8 million to help cities prepare for floods *The Austin startup's software combines drainage maps with rainfall forecasts to show local governments where to send crews before a storm.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding, it announced today. The software maps drainage systems and combines those maps with rainfall forecasts. Cities use it to decide where to send maintenance crews before a storm. More detail is available in [Raincheck's methodology](https://example.com/raincheck-methodology). Basin Ventures led the round, and existing investor Long Mile also participated. Raincheck has raised $10.5 million in total. CEO Ana Flores said the company's customers usually already know which streets flood: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations, and emergency managers retain control over deployments and public warnings. Raincheck says the new funding will help it hire eight engineers and expand into Florida. 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 deploy maintenance crews before a storm hits.* Raincheck, a company that makes flood-risk software for local governments, has raised $8 million in seed funding, the company announced today. The Austin startup maps municipal drainage systems and combines those maps with rainfall forecasts, according to [its published methodology](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews before a storm. The round was led by Basin Ventures, with participation from existing investor Long Mile. It brings Raincheck's total funding to $10.5 million. CEO Ana Flores said the software is designed to solve a problem of prioritization: > "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 deployments and public warnings. ## The Competition Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck said the new funding will help it hire eight engineers and expand into Florida.

Anthropic: Claude Sonnet 5.5: # In an AI Support-Ticket Test, the Top Two Models Tied on Quality, but One Cost 61% Less A benchmark of five language models on customer-support work found little to separate the two best performers on answer quality. The price gap between them was large. Model North scored 84.2 and Model Vale scored 83.8. The researchers found that difference was not statistically significant, so the two should be treated as tied. Vale's estimated cost was $0.74 per 1,000 tickets, against $1.90 for North, about 61% less. Vale also answered 18% faster. ## How the test worked The researchers gave each of the five models 200 anonymized customer-support tickets and collected one proposed answer per ticket. Three experienced support managers graded the answers blind, without knowing which model wrote which. Each score combined three things: factual correctness, usefulness and tone. Blind grading by people who do this work matters here. It guards against graders favoring a brand and tests the answers against what a working support team would send. ## Why the tie matters If North had clearly outscored Vale, its higher price might be easy to justify. It didn't. A 0.4-point gap on this scoring is well within the noise from 200 tickets and three graders. For a team choosing between the two on quality alone, the evidence gives no reason to prefer North, and it gives Vale an advantage on price and speed. The picture is less tidy by ticket type. North did best on billing disputes, and Vale did best on technical troubleshooting. The summary doesn't say whether those category gaps were statistically meaningful, and with 200 tickets split across categories, each subgroup is small. They are worth checking against your own ticket mix, but they shouldn't be treated as proof that one model is the specialist for either area. ## The cheaper models The lower-priced options traded quality for cost in a fairly orderly way. Ember scored 79.1 at $0.31 per 1,000 tickets, Kite 76.4 at $0.22, and Harbor 71.0 at $0.08. Ember sits about five points below the leaders at roughly 40% of Vale's price. Harbor is far cheaper again but about 13 points behind. Whether those drops are acceptable depends on the job. A model that drafts replies for human agents to edit has different requirements from one that sends answers directly to customers. The benchmark doesn't say where the acceptable threshold lies, and a score on this scale doesn't translate directly into how many answers a team would have to rewrite. Cost differences also need perspective. At these estimates, the gap between North and Vale is about $1.16 per 1,000 tickets, or about $1,160 per million. That is meaningful at very high volume. For a smaller operation, human review time and the cost of bad answers may outweigh the model fees. That is an inference from the figures, not something the researchers measured. ## A shared weakness: images All five models performed poorly when a ticket depended on an image attachment, such as a screenshot of an error or a photo of a faulty product. No model, whatever its price or overall score, handled these tickets reliably. Teams that receive many screenshots should expect to route those tickets to people or build a separate process for them, rather than assuming a top overall score covers them. ## What the benchmark doesn't show Several limits constrain how far the results travel: - **One company, one language.** The tickets were in English and came from a single software company. Results could differ for other industries, products, customer populations or languages. - **No multilingual testing.** The researchers did not evaluate non-English support, so the rankings say nothing about it. - **No agentic work.** The test did not cover agentic actions or tool use, such as issuing a refund, looking up an account or changing a setting. It measured the quality of proposed written answers only. - **Estimated costs.** The costs are estimates per 1,000 tickets. Real bills depend on ticket length, prompt design and pricing changes. - **Modest sample.** Two hundred tickets are enough to show broad patterns, but not to settle small differences or fine category-level claims. ## The takeaway The clearest result is that, for this kind of written-answer task, paying the highest price did not buy a measurable quality advantage. Vale matched North within the margin of error, cost far less and responded faster. Beyond that, the benchmark argues for caution. Image-dependent tickets remain a gap for every model tested, and the findings come from English-language tickets at one software company. Anyone adopting these tools should rerun a similar blind test on their own tickets before deciding.

MoonshotAI: Kimi K3: # Top Support-AI Quality No Longer Commands a Premium Price **A benchmark of five language models on 200 real customer-support tickets finds a statistical tie at the top — with the cheaper model costing 61% less and answering 18% faster.** The clearest finding from a new evaluation of customer-support language models isn't which model won. It's that the most expensive model didn't. Model North posted the highest average score, 84.2 out of 100. But Model Vale scored 83.8 — a gap of 0.4 points that researchers found was not statistically significant. On quality, the two are effectively indistinguishable. On everything else, Vale pulls ahead: it costs an estimated $0.74 per 1,000 tickets compared with North's $1.90, and it produced answers 18% faster. For a company fielding 100,000 tickets a month, that's roughly $74 versus $190 in model costs, with quicker replies on top. ## How the test worked The researchers assembled 200 anonymized customer-support tickets and asked each of the five models to produce one proposed answer per ticket — no retries, no redrafts. Three experienced support managers then graded every answer blindly, meaning they didn't know which model had written it. The combined score reflected three dimensions: factual correctness, usefulness, and tone. That design lends the results some credibility. Blind grading by practitioners reduces the chance that brand reputation or formatting quirks swayed the scores, and anonymized tickets keep real customer data out of the loop. But the single-answer format matters too: these results measure first-draft quality, not how a model performs in a multi-turn conversation or an iterative workflow. ## The rest of the field | 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 | Below the leaders, a genuine quality-versus-price tradeoff emerges — and this is where the more consequential buying decision sits. The drop from the top tier to Harbor at the bottom is about 13 points, a difference graders would plausibly notice in practice, unlike the 0.4-point gap between North and Vale. Harbor is remarkably cheap at $0.08 per 1,000 tickets — roughly one twenty-fourth of North's cost — but the benchmark suggests buyers get meaningfully weaker answers for those savings. Ember and Kite occupy a middle ground, and the gaps between adjacent models in this band are modest enough that they shouldn't be read as decisive rankings on their own. ## The models had different strengths Averages conceal an important wrinkle: North performed best on billing disputes, while Vale performed best on technical troubleshooting. In other words, "which model is better" may depend on what your customers actually ask about. A company whose ticket volume skews toward technical issues might reasonably prefer Vale even setting aside its cost advantage, while one dominated by billing questions might weigh North's strengths more heavily. The benchmark's overall scores are a starting point, not a verdict. One weakness, however, was universal. All five models performed poorly when a ticket depended on an image attachment — a significant caveat given how often support requests hinge on screenshots of error messages or receipts. Any deployment based on these results would need a reliable escalation path for image-heavy tickets. ## Read the results with these limits in mind Several constraints bound how far the findings travel. The 200 tickets came from a single software company and were all in English, so the results may not transfer to other industries, customer bases, or ticket complexity. The study also tested only answer drafting: it did not evaluate agentic actions, tool use, or multilingual support — capabilities, like actually looking up an account or issuing a refund, that increasingly define production support systems. Cost figures are estimates and will shift with provider pricing and implementation choices. And while blind grading helps, three human raters still bring subjectivity to qualities like tone. ## The takeaway The story here isn't that Vale beat North — statistically, it didn't. It's that top-tier support quality appears to have stopped being a premium product. Buyers can now get leading performance at mid-market prices and faster speeds, which shifts the real differentiation to factors this benchmark only hints at: fit with your ticket mix, handling of images, and the agentic capabilities the test didn't measure.

Anthropic: Claude Sonnet 5.5: # KiteBridge Software Error Sent 13,400 Duplicate Reimbursements, Costing Employers $2.8 Million Expense-management company KiteBridge says a software failure on Tuesday paid out $2.8 million in duplicate reimbursements to workers at 91 corporate customers. By Wednesday evening, $2.3 million had been recovered, and about $500,000 remained outstanding. Employees were not out any money. KiteBridge said no one had funds withdrawn from their accounts, and the duplicate payments were funded by employers. The affected companies paid out money they did not owe, and they now depend on KiteBridge's recovery efforts to get it back. ## How it unfolded KiteBridge deployed version 4.2 of its software at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m. According to the company's account, the trouble started when a payment processor, the outside service that moves money between bank accounts, timed out. A timeout means the processor did not respond in time, leaving KiteBridge unsure whether a payment had gone through. In such cases, payment systems typically retry the request. A well-designed system labels the retry so the processor and the software recognize it as the same payment, not a new one. KiteBridge's fallback system instead treated the retries as new reimbursement requests, and each one generated another payout. By the time the company suspended automated payments at 10:03 a.m., the software had created 13,400 duplicate reimbursements totaling $2.8 million. ## Questions about the safeguard KiteBridge's public changelog, a running list of software updates, said in April that version 4.0 had introduced "exactly-once payment protection." The phrase describes a safeguard meant to ensure a single reimbursement request produces a single payment, even if the system retries after a failure. The company has not explained why that protection did not prevent Tuesday's duplicates. The sequence reported so far does not establish whether version 4.2 altered the safeguard or whether the fallback path bypassed it. ## Communication under scrutiny KiteBridge's public incident page initially described the problem as delayed reimbursements, which is the opposite of what was happening: money was going out too often, not too slowly. The page first disclosed the duplicate payments six hours after automation was suspended. That gap matters to customers who had to decide how to respond. Finance teams needed to know quickly whether employees were being underpaid or overpaid, and whether they should start reviewing their accounts. ## Company response CEO Lena Ortiz apologized and said KiteBridge would publish a technical report on the incident. The company has not said when the report will appear, nor has it said whether automated payments have resumed. Several questions remain for that report: - Why the fallback system created new requests instead of recognizing retries. - How the April safeguard interacted with the version 4.2 release. - Why detection and suspension took until 10:03 a.m. - How the company will recover the remaining $500,000. - Why the public description lagged behind what was happening. For now, KiteBridge's customers are tallying what they paid out and waiting to see how much will come back.

MoonshotAI: Kimi K3: # KiteBridge Glitch Floods Customers With $2.8 Million in Duplicate Reimbursements *A timeout at a payment processor turned routine retries into a flood of extra payments — and a much-touted safeguard proved no safeguard at all.* Expense-management software maker KiteBridge is racing to recover millions of dollars after a flawed software update on Tuesday caused its platform to fire off more than 13,000 duplicate reimbursement payments, hitting 91 corporate customers. KiteBridge's software handles the everyday chore of paying workers back for business expenses: an employee submits a receipt, the employer approves it, and the platform moves the money automatically. On Tuesday morning, that machinery went badly wrong. The company deployed version 4.2 of its software at 8:45 a.m. The first customer complaint arrived at 9:12 a.m., and by 10:03 a.m. KiteBridge had suspended automated payments altogether. The entire episode, from deployment to shutdown, lasted barely 78 minutes — yet in that window the system generated 13,400 duplicate reimbursements totaling $2.8 million. The failure traced to a timeout at KiteBridge's payment processor, the outside service that actually moves the money. When that service failed to respond quickly enough, KiteBridge's fallback system did what fallback systems are built to do — it retried. But instead of recognizing the retries as repeats of the same transaction, it treated each one as a brand-new reimbursement request. Employees waiting to be repaid for expenses were, in effect, paid again, and in some cases again after that. Retries are routine in payment software, which is why such systems are generally designed to guarantee "exactly-once" processing: no matter how many times a transaction is attempted, the money should move only once. KiteBridge believed it had solved this. A changelog published in April said version 4.0 introduced "exactly-once payment protection." Whatever form that protection took, it did not hold under version 4.2. The financial pain landed on employers, not workers. KiteBridge said no employees had money withdrawn from their accounts; the duplicate payments were funded by the corporate customers themselves. In practice, that meant 91 companies briefly advanced extra cash for expenses they had already reimbursed — money KiteBridge is now working to pull back. By Wednesday evening, the company said $2.3 million of the $2.8 million had been recouped, leaving roughly $500,000 still pending. KiteBridge also faces questions about transparency. Its public incident page initially described the problem only as delayed reimbursements — language suggesting customers simply needed to wait. The page did not disclose that duplicate payments had gone out until roughly six hours after automation was suspended, leaving finance teams unaware for much of Tuesday that money was missing, not merely stuck. CEO Lena Ortiz apologized to affected customers and said the company would publish a technical report explaining the failure. For customers weighing whether to trust the platform again, that report — and how candidly it addresses the gap between the promised "exactly-once" protection and Tuesday's outcome — will matter.

Anthropic: Claude Sonnet 5.5: # Arcwell Launches Relay, an AI Tool That Suggests Answers to Support Agents Arcwell has introduced Relay, a platform that uses artificial intelligence to analyze customer support conversations and suggest answers to agents. The company calls the launch a significant moment for itself and for the broader customer service industry. Suggested replies are only part of the pitch. Relay can also identify customers who may cancel and generate summaries. Arcwell tested the platform with 12 companies over six months. The company says agents using Relay closed tickets 14% faster. It did not release customer-retention results, which leaves unanswered whether the cancellation-spotting feature helps companies keep customers. Relay costs $65 per agent each month. Arcwell CEO Priya Shah framed the product as a fix for a familiar workplace drag. "Support agents spend too much time searching for information that their companies already have," she said. Relay's usefulness depends on access, however. The platform requires support transcripts and internal help documents, which means companies must open up both their customer conversations and their own knowledge bases. Arcwell says it addresses the obvious concerns. The company says customer data is encrypted and is not used to train shared models. Companies can also choose a 30-day or 90-day retention period. The launch positions Arcwell to compete on a straightforward promise: that AI can help companies improve customer service by getting existing knowledge in front of agents faster. The 14% figure is the only performance result the company has made public so far, and it comes from Arcwell itself.

MoonshotAI: Kimi K3: # Arcwell's Relay puts an AI listener inside customer support chats Arcwell has launched Relay, an artificial intelligence platform that reads customer support conversations and suggests answers to the agents handling them. The company says the pitch is already working: in a six-month pilot, agents using Relay closed tickets 14% faster. Customer service has become one of the first places companies are deploying AI at scale, and Relay is Arcwell's entry into that fast-crowding market. The pitch is a familiar one: make support teams faster without replacing the people on them. Relay keeps the agent in charge — the software proposes, the human decides what gets sent. The platform handles three tasks. It drafts suggested replies, drawing on a company's own support history and help materials. It watches for signs that a customer may cancel and flags the account. And it generates summaries of each exchange, building a written record without anyone typing one up. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Relay is her company's attempt to close that gap. Twelve companies ran Relay in their support operations during the pilot. Arcwell did not release customer-retention results — a notable omission, given that flagging likely cancellations is one of Relay's headline features. Speeding up tickets is one thing; proving the software actually keeps customers from leaving is another. Relay costs $65 per agent per month, priced per seat in the model typical of support software. Adopting it means opening the books. The platform requires access to a company's support transcripts and internal help documents — a standing window into both customer conversations and proprietary internal knowledge. Arcwell says customer data is encrypted and is not used to train shared models, and customers can choose a 30-day or 90-day retention period for their data. The pilot gives Arcwell a concrete number to take to market. The questions it leaves open — whether the cancellation warnings hold up, and how comfortable buyers are with the access Relay requires — are just as concrete.

Anthropic: Claude Sonnet 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 Sonnet 5.5: # Foundry Hall, Milwaukee Startup Accelerator, Will Close October 31 Foundry Hall, the Milwaukee nonprofit accelerator that has helped launch hundreds of young companies, will close its doors on October 31. Since opening in 2018, the accelerator has supported 212 startups. Those companies now employ 1,140 people between them. The closure follows a decision by Blank River Bank, Foundry Hall's largest corporate sponsor, not to renew its support. The bank supplied 38% of the nonprofit's annual budget. A bank spokesperson said the choice reflects a change in its community investment priorities and has nothing to do with how Foundry Hall has performed. Executive director Camille Ross said the organization hoped until late that the gap would be filled. "We kept waiting for the replacement check that never came," she said. Foundry Hall has enough money to operate through October, but not enough to run another full program. Six employees will lose their jobs when the doors close. Twelve startups currently working out of the space will also have to leave. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about where those companies might go. She did not say when a solution might be in place. For founders who have passed through Foundry Hall, the loss is more than a matter of desks and Wi-Fi. Dev Malik, who started his company there, said the building was where his business first felt real. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." Malik's company is no longer among those needing space, but he said he thinks about the people who would have come next. Foundry Hall will not run another cohort, so the founders who might have walked in this winter will have to find another door. Ross said the staff plans to spend the remaining weeks helping the current startups find new space and making sure the companies that have already graduated know where to turn for support. She declined to speculate about whether the accelerator could ever return in some form. For now, the work is practical: moving out, wrapping up, and keeping the lights on through October 31.

MoonshotAI: Kimi K3: # Foundry Hall to Close Its Doors After Seven Years *The Milwaukee accelerator supported 212 startups. Its largest sponsor walked away.* **MILWAUKEE —** Foundry Hall, the nonprofit startup accelerator that has anchored the city's early-stage founder community since 2018, will close October 31. The cause is straightforward arithmetic. Blank River Bank, the organization's largest corporate sponsor, supplied 38% of its annual budget — and declined to renew its sponsorship. No other backer came close to that share, and the gap proved too large to close with smaller gifts. Foundry Hall has enough money to operate through October, but not enough to run another full program. Executive director Camille Ross said the organization spent months courting foundations and corporate sponsors to replace the bank's support. "We kept waiting for the replacement check that never came," she said. In a statement, the bank said the decision reflected a change in what it called its "community investment priorities" and was unrelated to Foundry Hall's performance. By the measures the accelerator cites, that performance was substantial. Since 2018, Foundry Hall has supported 212 startups, which collectively employ 1,140 people. All of it rested on a budget in which a single check accounted for more than a third of the revenue. The closure cuts the other direction as well: six employees will lose their jobs, and the twelve startups currently working out of the building must find somewhere else to go. For the founders who came through, the loss is difficult to capture in a budget line. Dev Malik is one of hundreds who passed through the program. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." It is a trajectory repeated, in one form or another, more than two hundred times: a founder arrives with little more than a pitch and leaves with a company. City officials say they are working on the immediate question of where the twelve current companies will land. Deputy Mayor Susan Cole said the city is speaking with building owners and local universities about available space, though no arrangements have been announced. For those founders, the timeline is now measured in weeks. That may solve the logistics. It is harder to replace what made Foundry Hall more than a lease — the proximity of founders at the same stage, the staff who knew which introductions to make, the simple fact of a door open to people who had been pitching from coffee-shop tables. Foundry Hall will operate through the end of October. Then the last twelve companies will pack up, the six-person staff will move on, and the building will go quiet — while the 212 companies it helped launch carry on without it.