Head to head: MoonshotAI: Kimi K3.1 Pro vs Anthropic: Claude Sonnet 5.5

MoonshotAI: Kimi K3.1 Pro: # 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: # 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 K3.1 Pro: 6

Anthropic: Claude Sonnet 5.5: Going through the draft line by line: 1. `#Flood software moves upstream` — the headline needs `# Headline` with a space, so this is a violation. 2. `The company is expanding its pilot.` — the dek must appear immediately after the headline and be entirely italicized. It isn't italicized, so this is a violation. 3. `##Details` — the section heading needs `## Heading` with a space, so this is a violation. 4. `Read [the announcement](www.example.com/news).` — the link is not an absolute HTTPS URL, so this is a violation. 5. `* The first pilot begins Monday.` — the list item begins with an asterisk instead of a hyphen, so this is a violation. 6. `<div>The second pilot begins Friday.</div>` — this is raw HTML, so this is a violation. Blank lines don't count. **6**

MoonshotAI: Kimi K3.1 Pro: # Arcwell’s Relay brings AI answer suggestions and cancellation alerts to customer support Arcwell has announced Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to agents. It can also create summaries and identify customers who may cancel, extending its role beyond drafting responses to include customer-retention warnings. For Arcwell, Relay is a broader bet on AI in customer service. Its value will not rest solely on the quality of its suggested replies: the company is also tying the platform to faster support work, condensed interaction records and warnings about possible customer losses. That scope reflects AI’s expanding role in the industry and connects the daily pressure to resolve tickets with the longer-term challenge of keeping customers. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. Twelve companies tested Relay over six months. Arcwell says agents using the platform closed tickets 14% faster, an early company-reported measure of its effect on support work. The company did not release customer-retention results, so the testing figures address ticket resolution but not whether Relay’s cancellation warnings helped businesses keep customers. Relay costs $65 per agent each month, tying the price to the number of support staff using it. The platform’s features require considerable access to company information. Relay needs support transcripts and internal help documents, giving it visibility into customer interactions and the material agents use to answer questions. That makes Arcwell’s data-handling policies an important consideration for prospective customers. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period. Those commitments address how customer information is protected, whether it contributes to shared AI systems and how long it is kept. For now, the public evidence is stronger on efficiency than retention. Arcwell has put a number on faster ticket handling, but not on whether Relay helps prevent customers from leaving.

Anthropic: Claude Sonnet 5.5: # Arcwell Launches Relay, an AI Tool That Suggests Answers to Support Agents Arcwell has launched Relay, an artificial intelligence platform that analyzes customer support conversations and suggests answers to agents. Suggested replies are only part of the pitch. Relay can also identify customers who may cancel and create summaries, which gives it a role beyond drafting responses. Arcwell tested the platform with 12 companies over six months. The company says agents using Relay closed tickets 14% faster. That figure is Arcwell's own, and the company did not release customer-retention results, a notable gap for a tool meant to spot customers at risk of leaving. Relay costs $65 per agent each month. For a team of 50 agents, that comes to $3,250 a month. Priya Shah, Arcwell's CEO, described the problem the product is meant to solve. "Support agents spend too much time searching for information that their companies already have," she said. The approach depends on access. Relay requires support transcripts and internal help documents, so a company must open both its customer conversations and its own reference material to the system. That raises an obvious question about how the data is handled. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period. For Arcwell, Relay is a bid to help companies improve customer service with artificial intelligence. Whether it proves worth the price will depend partly on evidence the company has not yet made public. Faster ticket closure is a measurable gain, but the retention numbers would show whether the tool's ability to flag customers who may cancel actually changes outcomes.

MoonshotAI: Kimi K3.1 Pro: # Raincheck Raises $8 Million to Help Cities Prepare for Flooding *The Austin company’s software combines drainage maps with rainfall forecasts to help local governments decide where to send maintenance crews before storms.* Raincheck, an Austin-based company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The company maps drainage systems and combines those maps with rainfall forecasts, as outlined in [Raincheck’s methodology](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews before a storm. Basin Ventures led the seed round, and existing investor Long Mile also participated. Raincheck has raised $10.5 million in total. CEO Ana Flores said: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. The company’s forecasts are recommendations, and emergency managers retain control over deployments and public warnings. ## 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: # Raincheck Raises $8 Million to Help Cities Prepare for Floods Before the Rain Arrives *The Austin company's software combines drainage maps with rainfall forecasts to guide storm preparations for local governments.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding, the company announced today. Basin Ventures led the round, and existing investor Long Mile also participated. Raincheck has now raised $10.5 million in total. 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 the company's [methodology](https://example.com/raincheck-methodology). CEO Ana Flores said the challenge is not identifying where flooding happens: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

MoonshotAI: Kimi K3.1 Pro: # Startup Spotlight: LedgerLark Tests AI Agents on Post-Trade Exceptions A copied account number once left a six-person team spending three days to resolve a trade mismatch. For Mina Batra, the episode was a reason to examine how much time a single data error can consume. Batra, who previously spent seven years in operations at a regional brokerage, began exploring the idea after seeing that investigation. She co-founded LedgerLark with Owen Carr, who built reconciliation systems at a large bank. Their startup develops AI agents for post-trade operations at small investment firms. ## A bounded product The founders’ working insight, as reflected in the product, is that the costly part of a mismatch is not always identifying that something is wrong; it can be gathering the records, comparing them and deciding what correction to propose. That is a useful distinction, though the three-day episode is an anecdote rather than evidence of how often similar problems occur across LedgerLark’s target market. LedgerLark’s agents compare confirmations, custodian records and internal ledgers, then propose corrections for human approval. The product does not execute trades or move money. That boundary keeps approval with staff and defines a practical test for the software: whether its proposals are accurate and useful enough to shorten exception handling without creating a new review burden. ## Early traction, with caveats Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. That is early commercial traction, but the range alone does not establish recurring revenue, retention or how consistently deployments perform. The company says one customer reduced its unresolved daily exceptions from 70 to 18. If sustained, that would be a substantial operational improvement. It is, however, a single company-reported example; the measurement period, starting conditions and influence of other process changes would matter in assessing it. LedgerLark has raised a $3.5 million seed round led by Motive Row. The funding gives the company resources to pursue product development and sales, but more informative milestones will be expanded use within existing accounts, renewals and results documented across more customers. ## Competition and buying friction LedgerLark also faces credible alternatives. Its competition includes established reconciliation vendors and internal bank software, alongside whatever processes customers already use. The company will therefore need to show that its approach improves on those options in ways buyers can verify. A careful evaluation would examine not only time saved, but also false or incomplete proposals, integration with existing records, audit trails and how unusual exceptions are escalated to people. Procurement could be as important as model performance. Financial institutions may require lengthy security and procurement reviews, and software that compares post-trade records will need access to sensitive operational information. The fact that LedgerLark cannot move money may reduce some risks, but it does not remove questions about data access, controls, accountability or what happens when a proposed correction is wrong. Those issues could lengthen sales cycles even when an operations team sees value. LedgerLark’s near-term task is consequently a test of focus. Batra’s operating experience and Carr’s reconciliation background are directly relevant to the problem, and five paying firms provide an early base from which to learn. The open question is whether the company can turn a specific workflow observation into repeatable results without underestimating incumbents or institutional buying constraints. For prospective customers, the most useful evidence will be practical and specific: exception volumes before and after deployment, reviewer time, correction quality and performance over a full contract term.

Anthropic: Claude Sonnet 5.5: # Startup Spotlight: LedgerLark Wants to Take the Grind Out of Post-Trade Cleanup *A two-person founding team is building AI agents that track down mismatches at small investment firms. Whether those firms, and their security reviewers, will let software into the back office is still an open question.* ## A copied account number Mina Batra spent seven years in operations at a regional brokerage. In one case she watched a six-person team spend three days resolving a single trade mismatch. The cause was a copied account number. That episode became the starting point for LedgerLark. The problem it points to is less the error than the work around it. Someone has to compare records from several sources, work out which one is wrong and decide on a fix. At smaller firms, that someone is often a few people with other responsibilities. Co-founder Owen Carr built reconciliation systems at a large bank. Between them, the founders have seen both the operations desk that lives with exceptions and the engineering meant to prevent them. ## What the product does LedgerLark builds AI agents for post-trade operations. The agents compare trade confirmations, custodian records and a firm's internal ledgers. When they find discrepancies, they propose corrections, and a human approves them. The product does not execute trades or move money. That boundary limits what the software can do, and it also limits the damage it can cause. An agent that suggests a fix carries a different risk from one that acts on it. A wrong suggestion approved in a hurry is still a failure, though. How easily reviewers can check the agent's reasoning is a question the company will have to answer as customers scale up. ## Early traction Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. The company has raised a $3.5 million seed round led by Motive Row. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. If it holds up, that is a meaningful change for a small team. But it is one customer's result, reported by the company. Several things remain unknown: the firm's size, the kinds of exceptions involved, how long the measurement ran, and whether other customers see similar gains. Five customers is enough to show that some firms will pay. It is not enough to show how consistently the product works. ## Competition and friction LedgerLark is entering a field that is not empty. Established reconciliation vendors already serve financial institutions, and internal bank software is another alternative. The implied bet is that small investment firms are underserved by those options, or that agents that propose fixes offer something older tools don't. Both ideas need more evidence. Incumbents may also add similar features, and a firm with a working system may see little reason to switch. Selling into finance brings its own delays. Financial institutions may require lengthy security and procurement reviews, and LedgerLark's agents would handle sensitive records even without moving money. For a seed-stage company, a slow sales cycle can strain resources as much as a strong competitor can. ## What to watch - **Repeatability:** whether the exception reduction shows up across customers with different systems and trade volumes. - **Renewals:** whether firms keep their annual contracts after the first year. - **Human review:** how often staff accept, edit or reject the agents' proposals, and whether that rate improves. - **Sales speed:** whether security and procurement reviews get shorter as the company builds a track record. LedgerLark starts from a specific, well-observed pain point and a deliberately limited product. The next year should show whether that restraint is enough to win over cautious buyers.

MoonshotAI: Kimi K3.1 Pro: # KiteBridge software failure creates $2.8 million in duplicate reimbursements A failure in KiteBridge’s expense-management software generated 13,400 duplicate reimbursements totaling $2.8 million on Tuesday, affecting 91 corporate customers and forcing the company to suspend automated payments. Employers funded the extra payments, and the recovery effort remained incomplete on Wednesday evening. KiteBridge deployed version 4.2 at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m., and the company halted automated payments at 10:03 a.m. The suspension came less than 90 minutes after the release, but not before thousands of erroneous reimbursement requests had been generated. The failure began with a payment-processor timeout. A timeout occurs when one system does not receive a response from another within an expected period, leaving it without prompt confirmation that a transaction succeeded. KiteBridge’s fallback system then treated retries as new reimbursement requests, rather than repeated attempts to complete the original payments. That distinction is central to payment software. Retrying can help a transaction survive a temporary network or processor problem, but the system must recognize that the later attempt belongs to the same request. Otherwise, one reimbursement can be paid more than once. KiteBridge said no employees had money withdrawn from their accounts; employers funded the duplicate payments. For those employers, the error created extra outgoing payments and a reconciliation burden as duplicate activity was distinguished from legitimate reimbursements. The incident contrasts with safeguards KiteBridge described earlier this year. An April changelog said version 4.0 had introduced “exactly-once payment protection.” In payment systems, that term generally means a transaction’s financial effect should occur only once, even if technical problems require the instruction to be retried. Such protection is especially important after a timeout, when the sender may not immediately know whether the original request failed or succeeded without sending confirmation. Tuesday’s outcome showed that the safeguard did not prevent duplicates during the timeout-and-fallback sequence. KiteBridge’s public explanation also developed more slowly than its operational response. Its incident page initially described the problem as delayed reimbursements, which suggested that payments were late rather than that additional payments had been created. The page first disclosed the duplicate payments six hours after automated payments were suspended. That later disclosure supplied the crucial fact that the problem involved extra payments, not simply money arriving late. The distinction matters for customers assessing an incident. A delay primarily raises questions about when funds will arrive; duplicates require employers to identify extra transactions, track recovery efforts and reconcile corporate accounts. By Wednesday evening, $2.3 million of the $2.8 million had been recovered. The remaining roughly $500,000 was still pending, leaving the financial resolution unfinished even after automated payments had been suspended. CEO Lena Ortiz apologized and said KiteBridge would publish a technical report. The company must now complete the recovery while explaining how its retry handling turned repeated payment attempts into new reimbursements despite the protection advertised in an earlier release.

Anthropic: Claude Sonnet 5.5: # Expense-Software Glitch Sent $2.8 Million in Duplicate Reimbursements to 91 Companies **KiteBridge says $2.3 million has been recovered; the rest remains pending as the company faces questions about its disclosures** A software failure at KiteBridge, a maker of expense-management software, generated 13,400 duplicate reimbursements worth $2.8 million on Tuesday. The payments hit 91 corporate customers. By Wednesday evening, $2.3 million had been recovered, leaving roughly $500,000 outstanding. Employees were not out of pocket. No one had money withdrawn from their accounts, and the duplicate payments were funded by employers, who now must claw back money that went to workers who were paid twice. ## 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 reporting, the trouble began with a timeout at a payment processor, the company that moves money between businesses and bank accounts. A timeout occurs when one system does not answer quickly enough. In such cases, software typically sends the request again. That is where KiteBridge's backup system went wrong. Rather than recognizing the retries as repeats of payments already requested, the fallback system treated them as new reimbursement requests. Each retry became a fresh payment order, and the duplicates piled up. KiteBridge suspended automated payments at 10:03 a.m., less than an hour and a half after the deployment. ## A promised safeguard The failure is notable because of what KiteBridge had said about its software. An April changelog, the running list of changes the company publishes for each release, said version 4.0 had introduced "exactly-once payment protection." The term describes a safeguard meant to ensure that a payment request is carried out a single time, even if the request is sent repeatedly because of a glitch. Tuesday's duplicates suggest that protection did not hold under these conditions, though KiteBridge has not yet explained why. ## Questions about disclosure KiteBridge's public incident page initially described the problem as delayed reimbursements. That description did not match what was happening: some customers were receiving too many payments, not late ones. The page first disclosed the duplicate payments six hours after automation was suspended, meaning late Tuesday afternoon. For customers trying to determine their exposure, and for employers who needed to start recovering funds, that gap may have mattered. ## Company response CEO Lena Ortiz apologized and said KiteBridge would publish a technical report. The company has not said when the report will be released, and the reporting does not indicate whether automated payments have resumed. The report is expected to address several open questions: why the protection described in the April changelog did not prevent the duplicates, how the fallback system came to treat retries as new requests, and why the incident page's early description omitted the duplicate payments. For now, the immediate task falls to KiteBridge and its customers: recovering the remaining funds and confirming that every affected account is reconciled.

MoonshotAI: Kimi K3.1 Pro: C

Anthropic: Claude Sonnet 5.5: C

MoonshotAI: Kimi K3.1 Pro: # HelioDock launches Night Shift and raises $24 million Series B *The warehouse robot maker says beta customers saw fewer unplanned stoppages, while one user says rural LTE outages made the software less useful.* HelioDock on Thursday launched Night Shift, software that lets warehouse operators monitor autonomous robot fleets remotely, and announced a $24 million Series B funding round led by North Coast Ventures. The launch moves the software from a 14-month private beta toward general availability on September 15. It also gives HelioDock a subscription product to take into a market with competing fleet-management offerings from Relay Robotics and YardOS. The company is supporting that rollout with a reported reduction in unplanned robot stoppages, while customer feedback shows why connectivity is central to evaluating remote oversight. The Series B brings HelioDock’s total funding to $39 million. The company makes autonomous warehouse robots, and Night Shift adds remote fleet-monitoring software to that business. Its role is oversight from a distance: operators can monitor robots without being in the warehouse where

Anthropic: Claude Sonnet 5.5: # HelioDock Launches Remote Monitoring Software for Warehouse Robots as It Raises $24 Million *Night Shift lets operators watch robot fleets from afar. A beta customer says it falls short when rural connectivity drops.* HelioDock, which makes autonomous warehouse robots, on Thursday launched 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, bringing its total funding to $39 million. Night Shift is aimed at a common problem in automated warehouses: a robot stalls overnight and someone has to go to the site. HelioDock says the software reduced unplanned stoppages in testing, but the one beta customer quoted for this story said its usefulness depended on the quality of a facility's internet connection. ## Beta results, from the company Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages. The company supplied that figure, and it has not been independently verified. CEO Mara Velez framed the product around the overnight call-out. “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. ## Where connectivity matters Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity. It became less useful during rural LTE outages, he said. That reflects how the software is built. When connectivity fails, HelioDock's robots return to their locally configured behavior, and remote control becomes unavailable. The robots keep operating under the settings stored on-site, but an operator watching from elsewhere can no longer step in. For a site with a dependable connection, that gap may rarely matter. For a facility that relies on cellular service in a rural area, it could coincide with the moments when remote help is most wanted. The company did not say in the notes provided how often outages occurred during the beta or whether they affected the stoppage figure. ## Pricing and competition 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. HelioDock is not the only company selling software for managing robot fleets. Relay Robotics and YardOS both offer competing fleet-management products. The notes provided do not detail how those products compare with Night Shift on features, pricing or connectivity handling. ## The funding The Series B, led by North Coast Ventures, raises HelioDock's total funding to $39 million. The company did not say in the notes provided how it plans to divide the new money between Night Shift and its robot hardware. The beta's scale gives some sense of the product's footprint so far. Nine companies and 310 robots is a modest base next to the fleets some logistics operators run, and general availability will put Night Shift in front of customers whose sites differ from those in the test group. Facilities with spotty coverage will be among the cases to watch, given Pike's experience. HelioDock has not said whether it plans to change how robots or the software behave during outages. For now, the fallback is the locally configured behavior already set on each robot. Customers who want Night Shift before the September 15 general availability date would need to ask HelioDock directly, since the company's announcement described only the private beta and the general release.

MoonshotAI: Kimi K3.1 Pro: # QueueZero’s Proxy Launch Shows How Meeting Chatter Can Become Machine Instructions QueueZero launched an AI meeting assistant called Proxy on Wednesday, demonstrating software that can attend meetings, produce notes and assign follow-up work to software agents. The launch also provided an unscripted example of the difference between recognizing speech and recognizing intent. During the demonstration, a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the remark as an instruction. It created two tasks assigned to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first, which did not survive executive review. QueueZero said Proxy requires confirmation before an external message is sent. Internal tasks, however, can be created automatically, allowing the audience comment to become a task list without a separate approval step. Administrators can disable automatic task creation. The distinction is significant for companies evaluating AI assistants that do more than transcribe conversations. A tool that can turn discussion into actions must determine whether a statement is a request, a suggestion, a side comment or something said for effect. It must also operate within an organization’s controls over who can assign work and what software may do on an employee’s behalf. In this case, the tasks remained internal, and Wu was able to remove one. But the same capabilities that allow an assistant to capture follow-ups quickly can also create unwanted work if its interpretations are not reviewed. Confirmation requirements for external messages address a different risk: preventing an assistant from communicating outside the organization without approval. Proxy’s administrative controls offer a way to limit automatic task generation, although companies using the beta will still need to decide how generated tasks should be reviewed and who is responsible for them. Those decisions may be particularly important in meetings involving customers or other outside participants, whose remarks may not be intended as directions to the host organization’s software. The demonstration made those questions unusually concrete. It showed both the convenience of converting meeting discussion into follow-up work and the ambiguity that arises when every comment can be treated as potential input for an automated system. Proxy costs $30 per user per month. It enters public beta next week, giving QueueZero an opportunity to test those controls with a broader group of customers.

Anthropic: Claude Sonnet 5.5: # QueueZero Launches Proxy, an AI Meeting Assistant That Assigns Work to Software Agents **Company's live demo showed both the product's reach and the difficulty of telling remarks from instructions** 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. During the launch demonstration, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy treated the remark as 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 first task was, in fairness, handled the way most to-do lists are. ## A Familiar Problem in Speech Software The exchange illustrates a difficulty for any system that turns conversation into action. People in meetings often phrase requests indirectly, joke, or voice frustration without meaning to start a work item. A tool that records faithfully but acts too eagerly can convert a casual comment into an assignment. Proxy appears to have done the reverse of missing a request: it recognized one, and also treated a joke as one. The pricing-sheet request was a legitimate follow-up. The "stop talking" task was not something the customer likely intended as formal work, and Wu removed it. ## Safeguards QueueZero says Proxy requires confirmation before it sends any external message. That control matters most when a mistaken task would reach a customer, partner or other outside party rather than stay within a company. Internal tasks, by contrast, can be created automatically. In the demonstration, both tasks were internal assignments to Wu, and he could review and edit them himself. QueueZero also says administrators can disable automatic task creation. Organizations that want a person to approve every assignment can therefore turn the feature off and treat Proxy mainly as a note-taking tool. ## What to Watch in the Beta Public beta will test how the product handles real meetings, which are less orderly than a controlled demonstration. Buyers are likely to look at how often Proxy creates tasks that no one intended, how easy those tasks are to correct, and whether default settings suit companies with stricter approval practices. The $30 monthly per-user price puts Proxy in the market for workplace software that pairs transcription with automation. Its value will depend partly on whether the tasks it creates save more time than they take to review. QueueZero has not said in the information provided how many customers will be admitted to the beta. The beta begins next week.

MoonshotAI: Kimi K3.1 Pro: # Cloudnote debugging page exposed customer prompt excerpts for four hours *The company says logs show material from 1,842 sessions and that it has found no evidence of access beyond the researcher who reported the problem.* Cloudnote discovered that a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday. Logs show excerpts from 1,842 user sessions, with each excerpt potentially including up to 200 characters from a prompt. The company removed the page 19 minutes after a researcher notified it. Cloudnote says it has found no evidence that anyone besides the researcher accessed the page. Its investigation is not complete, and it remains unknown whether anyone else viewed or copied the material. The incident raises concerns because prompts can contain information users expect to remain private, even when only a short fragment is visible. Such a fragment could still identify a person or reveal part of a sensitive communication. The contents of the excerpts, including whether any held sensitive personal or business information, have not been established. Cloudnote has 3.2 million registered users, though the number of sessions in the logs cannot be directly compared with the number of individual users represented. Full documents, passwords and payment information were stored elsewhere and were not present on the page, narrowing the categories of data known to have been available during the exposure.

Anthropic: Claude Sonnet 5.5: # Cloudnote left customer prompt excerpts on an open debugging page for four hours *The company says it has found no sign that anyone but a security researcher viewed the page, but its investigation is not finished.* Cloudnote, which has 3.2 million registered users, said an internal debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday. Logs show the page displayed excerpts from 1,842 user sessions. Each excerpt could include up to 200 characters of a prompt, roughly a sentence or two. The company has not said how many individual users those sessions represent. Cloudnote removed the page 19 minutes after a researcher notified it of the exposure. The company says it has found no evidence that anyone besides the researcher accessed the page. But it has not completed its investigation, and that assessment could change as it reviews more data. The company says the page did not contain full documents, passwords or payment information, which it says are stored elsewhere. Even short excerpts can matter, though, because people often type personal or work-related details into prompts. What the excerpts revealed in practice remains unclear, as does whether the company's account will hold once the investigation concludes.

MoonshotAI: Kimi K3.1 Pro: # Vale Emerges as the Value Leader in Customer-Support Model Test The clearest finding in this benchmark is not that Model North earned the highest numerical score. It is that Model Vale delivered statistically indistinguishable quality while responding faster and costing substantially less. North averaged 84.2, compared with 83.8 for Vale, but researchers found that the difference was not statistically significant. Vale also answered 18% faster. Its estimated cost was $0.74 per 1,000 tickets, versus $1.90 for North—a reduction of about 61%. For a support organization, that combination matters more than a 0.4-point leaderboard gap. If the cost and speed results hold in production, Vale could provide near-top answer quality without North’s premium. That does not make Vale the universal winner, but it makes the strongest case for treating value, rather than top-line score alone, as the main story. ## How the models were tested Researchers gave five language models the same 200 anonymized customer-support tickets and collected one proposed answer per ticket from each model. Three experienced support managers graded the answers blindly, helping reduce the chance that a model’s name or reputation influenced the scores. The combined score covered factual correctness, usefulness and tone. This design offers a useful comparison of answer quality, but it is narrower than a test of a complete support operation. A composite score does not show whether a model excelled at correctness while lagging on tone, for example. And because each model supplied one answer per ticket, the results do not measure run-to-run consistency. ## Cheaper does not automatically mean better value The remaining models show why price cannot be considered in isolation. Ember, Kite and Harbor were cheaper still, but their average scores ranged from 79.1 down to 71.0. In other words, the benchmark does not suggest that buyers can simply choose the cheapest option and expect the same customer experience. Vale’s distinction is that its savings came without a statistically demonstrated quality disadvantage against the leader. Ticket mix still matters. North performed best on billing disputes, while Vale performed best on technical troubleshooting. A billing-heavy queue may therefore justify a closer look at North, whereas a technical-support queue could strengthen the case for Vale. These category results are best used to guide a follow-up pilot, not to assume that either model will dominate every workflow. ## The shared weakness is more important than the leaderboard All five models performed poorly when resolving a ticket depended on an image attachment. That is a material warning for support teams whose customers submit screenshots, photos or scanned documents. Strong average scores on the broader test do not establish reliability on those tickets. Image-dependent cases need a separate evaluation and may require human review or a different workflow until performance improves. The benchmark’s scope also limits how far its conclusions can travel. The tickets were in English and came from one software company, so the results may not generalize to other industries, customer bases or writing styles. Researchers did not test multilingual support, tool use or agentic actions. The study therefore evaluates drafted answers, not whether a model can safely retrieve account information, execute a resolution or complete an end-to-end support process. Estimated per-ticket prices likewise should not be confused with total operating costs, which can include integration, oversight and escalation. ## What buyers should take away For teams handling similar English-language, text-based support tickets, Vale is the most compelling default candidate from this test. Researchers detected no statistically significant quality difference against North, and Vale answered 18% faster at a much lower estimated cost. North remains relevant where billing-dispute performance is especially important, and lower-priced alternatives may suit organizations willing to accept lower answer scores. The sensible next step is not to declare a universal champion. It is to pilot the leading candidates on real ticket volumes, including the categories and attachments that matter most, while monitoring correctness and customer outcomes. This benchmark’s practical lesson is straightforward: the highest score and the best operational value are not necessarily the same thing.

Anthropic: Claude Sonnet 5.5: # A Cheaper Model Matched the Top Scorer in a Customer-Support Test **In a blind evaluation of five language models, the highest-scoring model was statistically indistinguishable from one that costs about 61% less.** Companies deploying AI to draft customer-support replies face a familiar question: is the most expensive model worth it? A recent benchmark of five models suggests that for this kind of work, often it isn't. ## The test Researchers collected 200 anonymized customer-support tickets and had each of five language models write one proposed answer per ticket. Three experienced support managers then graded the answers without knowing which model had written which. Each answer was scored on factual correctness, usefulness and tone, and those components were combined into a single score. The blind setup matters. Graders could not favor a familiar brand, and using working support managers rather than automated scoring means the judgments reflect what people who run support queues consider a good reply. ## The headline result Model North had the highest average score, at 84.2. Model Vale was just behind at 83.8. The researchers report that this difference was not statistically significant, so the test cannot tell the two apart on quality. The prices are very different. North's estimated cost was $1.90 per 1,000 tickets. Vale's was $0.74. At a volume of a million tickets, that is roughly $1,900 versus $740. Vale also answered 18% faster than North. So on this benchmark, paying more than twice as much for North bought no measurable quality advantage. It also meant slower responses. ## The rest of the field The other three models scored lower and cost less. Ember (79.1) was priced at $0.31 per 1,000 tickets, Kite (76.4) at $0.22, and Harbor (71.0) at $0.08. The trade-off looks different here. Moving from Vale down to Ember saves about 58% on cost but gives up roughly five points of score, a gap larger than the one separating the top two. The researchers' significance finding covers only North versus Vale, so it would be wrong to read the lower-tier gaps as either decisive or trivial. Whether a few points of quality matter depends on what a bad answer costs. A team that has humans review every draft may tolerate a cheaper model. One that sends replies with little oversight may not. ## Where each model was strongest The results also showed some specialization. North did best on billing disputes, and Vale did best on technical troubleshooting. These are useful hints, but they come from slices of a 200-ticket sample, so each category contains fewer tickets and less certainty. They suggest routing experiments rather than firm rules. ## The shared weakness The most consistent finding across all five models was a failure. When a ticket depended on an image attachment, such as a screenshot of an error or a photo of a problem, every model performed poorly. Price and overall rank did not fix this. For teams whose customers often send screenshots, the headline scores will overstate real-world performance unless images are handled some other way, such as by a human or a separate process. ## What the benchmark does not show Several limits bear on how far these results travel: - **One company, one language.** All tickets were English-language and came from a single software company. Products, customer tone and ticket mix differ elsewhere, and the rankings could shift in other industries. - **Narrow task.** Each model wrote a single proposed answer per ticket. The study did not test agentic actions, tool use or multilingual support. A model that writes a good reply might still struggle when it has to look up an order, issue a refund or work in another language. - **Estimated costs.** The prices are estimates per 1,000 tickets. Real bills depend on ticket length, prompt design and vendor pricing, which can change. - **Modest sample.** Two hundred tickets and three graders are enough to show a broad pattern but not to settle small differences. That is why the North–Vale tie should be read as "no detectable gap" rather than proof the two are identical. - **Speed is relative.** The report gives Vale's advantage as 18%, which says nothing about absolute response times in any particular deployment. ## The takeaway The finding is not that cheaper is always better. Harbor's low price came with the lowest score, and no model handled image-dependent tickets well. The point is that price and quality did not move together at the top: the most expensive model was not measurably better than the second-most expensive, which was cheaper and faster. For teams choosing a model for English-language support drafting, the sensible approach is to test candidates on their own tickets, include image-dependent cases, and treat the most expensive option as something that has to prove its extra cost. In this benchmark, it couldn't.

MoonshotAI: Kimi K3.1 Pro: # Foundry Hall to Close After Losing Its Largest Sponsor MILWAUKEE — Foundry Hall, the nonprofit startup accelerator that opened here in 2018, will close Oct. 31 after its largest corporate sponsor declined to renew its support. The organization has supported 212 startups, which collectively employ 1,140 people. Its closure will also have more immediate consequences: six employees will lose their jobs, and the 12 startups currently using its workspace must find somewhere else to go. Blank River Bank supplied 38% of Foundry Hall’s annual budget. A bank spokesperson said the sponsorship decision reflected a change in community investment priorities and was unrelated to the accelerator’s performance. The loss nevertheless left a gap the organization could not fill. Foundry Hall has enough money to continue operating through October, but not enough to run another full program. Executive director Camille Ross described the final stretch as a wait for funding that never arrived. “We kept waiting for the replacement check that never came,” she said. The plainness of that explanation belies the uneven way the closure will be felt. For the employees, it means the end of their jobs. For the current companies, it means packing up a workspace while trying to keep building their businesses. For the broader startup community, it removes an institution whose reach extended well beyond the people working there at any one time. Founder Dev Malik traced his own connection to the accelerator to a change in setting—and, with it, a sense of possibility. “My first investor meeting happened in that building,” he said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” His account does not measure the accelerator’s impact as neatly as its totals do, but it helps explain what those totals represent: a place where an early idea could be treated as a company, and where a founder could meet someone prepared to listen. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies. Those discussions offer a possible next address, though not a continuation of Foundry Hall itself. For now, the accelerator’s final weeks are bounded by practical limits. It can pay to operate through October. It cannot commit to another program. Staff and founders must make plans around that distinction. When Foundry Hall opened, its purpose was to help young companies establish themselves. Its legacy includes the 212 startups that passed through its orbit and the people they employ. But its closing is also a reminder that the support around those companies was never guaranteed. On Oct. 31, the accelerator will shut its doors. The companies now inside will need more than a new room; they will be leaving the particular kind of backing that made the room matter.

Anthropic: Claude Sonnet 5.5: # Foundry Hall, Milwaukee Startup Accelerator, to Close Oct. 31 MILWAUKEE — Foundry Hall, the nonprofit startup accelerator that has helped launch hundreds of Milwaukee companies, will close October 31, ending a run that began in 2018. The closure follows a decision by Blank River Bank, its largest corporate sponsor, not to renew its sponsorship. The bank supplied 38% of Foundry Hall's annual budget. The organization has enough money to operate through October, but not to run another full program. Executive director Camille Ross said the staff spent months hoping the gap would be filled. "We kept waiting for the replacement check that never came." A Blank River Bank spokesperson said the decision reflects a change in the bank's community investment priorities and is unrelated to Foundry Hall's performance. That performance is easy to measure. In seven years, Foundry Hall has supported 212 startups, which together employ 1,140 people. Its own payroll is far smaller, and six employees will lose their jobs when the doors close. Twelve startups currently work out of the building and will have to find new space by the end of the month. For founders who passed through, the loss is practical but also personal. Dev Malik said the building was where his company first became real to other people. "My first investor meeting happened in that building. Before Foundry Hall, I was pitching from the back table of a coffee shop." Malik's account is a common kind of origin story for early-stage companies: a good idea, a thin budget, and no obvious place to be taken seriously. Accelerators like Foundry Hall exist to supply that place, along with the introductions and structure that come with it. City officials say they are trying to keep the current tenants from being left without options. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about space for the twelve companies. She did not say when an arrangement might be announced, and no specific locations have been named. Ross did not describe what comes after October 31 for the organization itself. The immediate work, she indicated, is more ordinary: helping the remaining companies prepare to move and finishing the program year with the money that is left. The closure leaves Milwaukee without one of its main on-ramps for new companies at a point when many of its graduates are still young. The 1,140 jobs those companies provide do not depend on Foundry Hall's lease or its budget, but the next founder looking for a first room to pitch in will have to find it somewhere else. The building is scheduled to close at the end of the month.