Head to head: MoonshotAI: Kimi K3 vs Qwen: Qwen3.8 2.4T A95B

MoonshotAI: Kimi K3: # Arcwell launches Relay, an AI assistant for customer support agents Arcwell has launched Relay, an artificial intelligence platform that reads customer support conversations as they unfold and suggests answers to the agents handling them. The launch marks a significant step for Arcwell as it pushes into AI-powered customer service, a corner of the software industry that has been quick to adopt the technology. The product takes aim at a problem Arcwell's chief executive, Priya Shah, describes as endemic to support teams. "Support agents spend too much time searching for information that their companies already have," she said. Relay tackles that problem on three fronts. The software drafts suggested replies in real time, flags customers whose messages indicate they may be preparing to cancel — giving agents a chance to intervene — and generates summaries of each conversation. Arcwell put Relay through a six-month pilot with 12 companies and reports that agents using the platform closed tickets 14% faster, a figure that will serve as the product's headline selling point. The number comes from Arcwell's own testing, and the company did not release results on customer retention, leaving open the question of whether the cancellation warnings actually keep customers from leaving. Relay costs $65 per agent, per month. Using it means handing over a great deal of institutional data. The platform requires access to a company's support transcripts and internal help documents to function. Arcwell says customer data is encrypted and is never used to train shared models, so one client's conversations will not shape the suggestions another receives. Companies can also set their own retention window, choosing to store data for either 30 or 90 days. For Arcwell, Relay is an effort to establish itself in a customer-service industry that is rapidly adopting AI. For the companies weighing a subscription, the calculation is simpler: whether faster ticket times — and the chance to catch unhappy customers before they walk — are worth the price of admission and the data access Relay demands.

Qwen: Qwen3.8 2.4T A95B: # Arcwell launches Relay to suggest support answers and flag cancellation risk Arcwell announced Relay, an artificial intelligence platform that reviews customer support conversations and suggests answers to agents. The launch extends Arcwell’s product line into AI-assisted customer service and adds another option for support teams experimenting with automation. Relay is designed to do more than prompt agents with possible replies. The platform can identify customers who may cancel and produce summaries of support interactions, Arcwell said. Those features place it in a growing category of software that turns support transcripts into operational signals, rather than leaving them as records of past conversations. The company said Relay was tested by 12 companies over six months. During that period, agents using the platform closed tickets 14% faster, according to Arcwell. The company did not release customer-retention results, leaving unanswered how well the cancellation-risk feature performs outside its own measurements. Relay is priced at $65 per agent each month. CEO Priya Shah said the product addresses a familiar inefficiency in support organizations. “Support agents spend too much time searching for information that their companies already have,” Shah said. To provide those suggestions, Relay requires access to support transcripts and internal help documents. That access raises data-governance questions, particularly for companies handling sensitive customer communications. Arcwell said customer data is encrypted and is not used to train shared models. Companies can choose either a 30-day or 90-day retention period, the company said. The platform’s rollout follows broader industry efforts to apply AI to support metrics, including resolution speed, agent workload and customer retention. Its results will likely be scrutinized by support leaders weighing AI spending against privacy requirements and proof of measurable return. Relay enters a customer-service software market where vendors are adding AI features aimed at reducing manual work and predicting customer behavior. For Arcwell, the test will be whether the platform can deliver faster ticket handling and useful churn signals consistently enough to justify its per-agent price.

MoonshotAI: Kimi K3: # QueueZero Launches Proxy, an AI Assistant That Attends Meetings and Assigns the Follow-Up Work QueueZero on Wednesday launched Proxy, an artificial intelligence assistant that attends meetings, produces notes and assigns follow-up work to software agents — a pitch the company wound up demonstrating, unintentionally, on its own chief executive. Proxy joins a meeting on a user's behalf, generates notes and summaries, and converts discussion into tasks that can be routed to automated agents for completion. QueueZero is pitching it as a step beyond transcription: software that does not just record a meeting but acts on it. During the launch 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. The moment drew laughter in the room, but it also neatly framed the central question facing Proxy and products like it: an assistant that acts on what it hears will sometimes act on what it was never meant to hear. The remark was the kind of conversational aside that fills most meetings. QueueZero said Proxy is built with that boundary in mind. The assistant requires confirmation before any external message is sent, the company said, meaning a task such as emailing a customer cannot be executed without a person's approval. Internal tasks, like the ones created during the demonstration, can be generated automatically. Administrators can disable automatic task creation entirely. That division — automatic internally, confirmed externally — reflects the compromise enterprise buyers have come to expect from AI features: useful automation inside the organization, with a human checkpoint before anything leaves it. Proxy enters a crowded market. Microsoft, Zoom and Google have all folded AI summaries into their conferencing products, while startups such as Otter.ai and Fireflies.ai have built businesses on transcription and notes. Vendors are now racing to add task execution, and QueueZero is betting that delegating work to agents, rather than merely documenting it, will distinguish a paid product from features competitors increasingly bundle for free. Proxy costs $30 per user per month and enters public beta next week. QueueZero did not say how many organizations took part in earlier testing. Whether Proxy can consistently tell an instruction from an aside is among the questions the beta period is likely to answer.

Qwen: Qwen3.8 2.4T A95B: # QueueZero Debuts Proxy, an AI Meeting Assistant That Turns Remarks Into Tasks QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant designed to attend meetings, take notes and assign follow-up work to software agents. The product enters public beta next week at $30 per user per month. The company demonstrated the tool before an audience of customers and partners. During the session, a person in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the comment as an instruction. It created two tasks assigned to QueueZero chief executive Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. The moment gave the company an unplanned example of how Proxy parses ordinary workplace language. QueueZero said the system is built to distinguish between internal work items and actions that require human approval. According to the company, Proxy requires confirmation before any external message is sent. Internal tasks can be created automatically, though administrators can disable that automatic task creation. That distinction may matter as businesses race to deploy AI agents into calendars, project management systems and messaging platforms. Meeting assistants are among the most competitive categories in enterprise software, with vendors promising to reduce time spent on note-taking, status updates and administrative follow-through. The challenge is preventing an agent from acting on casual comments, jokes or side remarks. QueueZero is betting that controls and audit settings can make autonomous agents safer. By requiring confirmation for outbound communication, the company is trying to avoid the kind of embarrassing or disruptive automated message that can quickly spread through a workplace. At the same time, allowing automatic internal task creation is meant to keep projects moving without forcing employees to manually enter every action item. Pricing places Proxy in the midrange of AI productivity tools, where companies are increasingly asking employees to justify subscriptions through measurable time savings. For QueueZero, the public beta will test whether customers trust the assistant enough to give it access to meetings, workflows and internal assignments. The launch demonstration suggested that the technology can be remarkably literal. Whether that becomes a selling point or a warning will depend on how well QueueZero’s safeguards work in real offices.

MoonshotAI: Kimi K3: # LedgerLark Wants AI Agents to Take On the Back Office's Slowest Days *Startup Spotlight* A few years ago, Mina Batra watched a six-person operations team spend three days resolving a single trade mismatch. The cause was mundane: a copied account number. The trade itself had gone through fine. Everything after it had not. Batra had spent seven years in operations at a regional brokerage, and that episode pushed her to start exploring whether software could take on more of the work. She founded LedgerLark with Owen Carr, who had built reconciliation systems at a large bank. Their premise is modest-sounding: much of post-trade operations — checking that confirmations, custodian records and internal ledgers agree after a trade — follows recurring patterns that software can learn, while the judgment calls still belong to people. ## What the agents do LedgerLark's AI agents compare trade confirmations, custodian records and a firm's internal ledgers, flag discrepancies and propose corrections. A human reviews and approves each fix before anything changes. The product does not execute trades or move money. That boundary appears deliberate, and probably prudent. It keeps the company clear of the most heavily regulated parts of the trade lifecycle, and it draws a legible line for a buyer's compliance team around what the software can touch. Whether customers find that line reassuring or limiting is an open question — some buyers may eventually want automation that goes further, whatever regulators prefer. The focus on small investment firms also follows the founders' logic. Large banks can afford to build reconciliation tooling in-house; Carr's former employer did. Smaller firms generally cannot, which is where LedgerLark sees room. ## Early customers, early numbers Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The wide spread suggests pricing that scales with firm size or scope of work. At the top of the range, the revenue is meaningful for a young company; at five customers, the sample is too small to say how repeatable the pitch is. The company says one customer reduced its unresolved daily exceptions from 70 to 18 after adopting the software. That figure comes from LedgerLark itself, reflects a single customer and has not been independently verified. Results will presumably vary with the state of a firm's existing records and processes. LedgerLark has raised a $3.5 million seed round led by Motive Row. ## The hard parts ahead LedgerLark is not entering an empty field. Established reconciliation vendors already sell to investment firms, and internal bank software covers some of the same ground at larger institutions. The startup's case is that its agents do more than match records — they propose fixes. Incumbents can presumably build similar features, so any durable advantage may come down to focus, deployment speed or accuracy over time. Accuracy will matter in both directions. Missed discrepancies leave risk in place, while bad proposals waste the reviewer hours the product is meant to save. How LedgerLark measures and shares its error rates may shape whether cautious firms say yes. Procurement poses a separate risk. Financial institutions frequently require lengthy security and procurement reviews before new software touches their books. That stretches timelines for any vendor and can strain a small company working through its first contracts. For now, LedgerLark's wager is that the back office's slowest days are a software problem worth solving carefully rather than loudly. Five signed contracts indicate some firms agree. The next few security reviews will say more.

Qwen: Qwen3.8 2.4T A95B: # Startup Spotlight: LedgerLark Tackles the Messy Work After a Trade LedgerLark is building artificial-intelligence agents for a part of finance that rarely attracts glamour: post-trade operations. The startup’s software compares trade confirmations, custodian records and internal ledgers, then proposes corrections for human approval. It does not execute trades or move money, a boundary that may matter to cautious financial firms. The insight comes from co-founder Mina Batra’s time in operations at a regional brokerage, where she spent seven years dealing with the manual follow-up that can trail a transaction. One episode stood out: a six-person team spent three days resolving a trade mismatch caused by a copied account number. The problem was not exotic. It was the kind of small error that becomes expensive because it requires people to hunt across systems, emails, spreadsheets and counterparties. Co-founder Owen Carr brings a different vantage point. He built reconciliation systems at a large bank, where institutions often solve these problems with bespoke internal software. Together, the founders are aiming at small investment firms that may lack the engineering resources of a large bank but still face similar operational risks. LedgerLark’s pitch is that AI agents can accelerate the investigation stage. Instead of an operations analyst opening every record manually, the agents review the relevant documents, flag discrepancies and draft suggested fixes. A person then decides whether to accept or adjust the proposal. That structure keeps the product in the role of assistant rather than autonomous operator, which may be easier for compliance-minded customers to evaluate. There is early traction, though the numbers are modest. LedgerLark says five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. The company also says one customer reduced unresolved daily exceptions from 70 to 18. As with any vendor-provided metric, it is worth asking how the baseline was measured, over what period, and whether the result would repeat across different workflows. The seed funding is another signal of interest. LedgerLark has raised a $3.5 million round led by Motive Row. That is meaningful capital for a young company, but it also raises the usual question: can the team turn a handful of early users into a repeatable sales motion in a sector known for careful buying cycles? The competitive landscape is not empty. Established reconciliation vendors already serve financial institutions, and some firms rely on internal software built specifically for their operations. Those systems may be old or awkward, but they are known quantities. LedgerLark will need to show that its agents are not merely a new interface, but a genuine reduction in risk, time and manual effort. Security and procurement may be just as important as the technology. Financial institutions often require lengthy reviews before connecting new software to sensitive data. For a startup selling AI agents that read confirmations and ledger entries, trust will likely depend on audit trails, access controls, explainable recommendations and clear limits on what the system can do. Still, the problem looks real enough. Post-trade errors are rarely dramatic, but they accumulate. They consume staff hours, delay reporting and create operational exposure. If LedgerLark can consistently narrow the search space for exceptions, even without full automation, it may find a practical wedge. The company remains early, with a small customer base and unproven scalability. But its focus is specific, and its framing is appropriately restrained. Rather than promising to replace operations teams, LedgerLark is trying to make the messy work after a trade less slow, less manual and less error-prone. In financial operations, that may be exactly the kind of improvement buyers want to test.

MoonshotAI: Kimi K3: # Milwaukee's Foundry Hall to Close After Lead Sponsor Pulls Funding MILWAUKEE — Foundry Hall, the nonprofit accelerator that has backed Milwaukee startups since 2018, will close October 31 after its largest corporate sponsor declined to renew its funding. Blank River Bank supplied 38 percent of the accelerator's annual budget. A bank spokesperson said the decision reflects a change in the bank's community investment priorities and is unrelated to Foundry Hall's performance. For a small nonprofit, the distinction made little difference. Executive director Camille Ross said the organization spent months approaching foundations, other banks, and individual donors, but no single commitment came close to replacing what was lost. "We kept waiting for the replacement check that never came," she said. Foundry Hall has enough money to operate through October, Ross said, but not enough to run another full program. The current cohort will be its last. The record it leaves behind is substantial: 212 startups supported over seven years, companies that now collectively employ 1,140 people. The arithmetic of the closure is smaller and heavier. Six employees will lose their jobs. Twelve startups currently working in the building must find somewhere else to go by the end of the month. For now, the shared workspace still hums — keyboards, phone calls, the daily small talk of young companies — but the month has a fixed end. For Dev Malik, one of the founders who passed through its doors, the loss is hard to reduce to numbers. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." The accelerator offered more than desks, he said — mentors who returned phone calls, introductions that would have taken years to make alone, and the simple credibility of an address. Stories like his, founders say, are the building's common currency: first meetings, first hires, first checks. The city is attempting to soften the landing. Deputy Mayor Susan Cole said officials are speaking with building owners and local universities about space for the twelve displaced companies, with the hope of keeping them in Milwaukee. No arrangement has been finalized, and the founders do not yet know where they will work come November. Ross said the staff's final weeks will look much like the years that came before: helping founders solve problems. This time, the problem is where to go next. The goal, she said, is to leave the building in order and its companies in good hands. For the founders who once pitched from coffee-shop tables, Foundry Hall was proof that Milwaukee had room for them. What proves it now is a question the city has yet to answer.

Qwen: Qwen3.8 2.4T A95B: # The Last Shift at Foundry Hall Foundry Hall will close on October 31, ending a run that began in 2018 with a simple promise: give Milwaukee entrepreneurs a place to be taken seriously. The nonprofit startup accelerator has supported 212 companies, which together employ 1,140 people. Those numbers will remain in reports and grant summaries, but the daily life of the building — whiteboards left full of names, founders rehearsing pitches near the coffee station, the nervous quiet before demo day — will disappear. The closure follows the loss of the hall’s largest corporate sponsor, Blank River Bank, which supplied 38% of its annual budget. The bank declined to renew its sponsorship, explaining that the decision reflected a shift in community investment priorities and was not tied to Foundry Hall’s performance. Executive director Camille Ross said the organization kept waiting for “the replacement check that never came.” That phrase captures the slow nature of the ending. Foundry Hall did not collapse overnight. It had enough money to operate through October, but not enough to fund another full program. Six employees will lose their jobs. Twelve startups currently using the workspace must leave, carrying files, prototypes, and routines into an uncertain next chapter. The shutdown lands differently because the hall was meant to be a bridge, not a landlord. Its programs connected founders to mentors, investors, and one another. The loss of the bank’s support left the organization searching for a new anchor while still paying for work already underway. For some, the building was more than square footage. Founder Dev Malik remembered how Foundry Hall gave him a first real setting for investor conversations, after he had once pitched from the back table of a coffee shop. His recollection points to what accelerators often provide beyond funding: credibility, address, and a room where an idea can be heard. City officials are trying to soften the landing. Deputy mayor Susan Cole said Milwaukee is speaking with building owners and local universities about possible space for the current companies. That effort may help preserve some of the community Foundry Hall gathered around itself. Yet a workspace is hard to replace when it has become a habit of mutual belief. As the October days shorten, the hall’s final cohort will finish meetings, clear desks, and take down signs. The startups may continue, and the employees may find new roles, but Foundry Hall itself will close with the quiet dignity of a place that did what it was built to do. It made room. Now the room is gone.

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.

Qwen: Qwen3.8 2.4T A95B: # 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: 6

Qwen: Qwen3.8 2.4T A95B: 6

MoonshotAI: Kimi K3: # Cloudnote debugging page exposed prompt excerpts from 1,842 user sessions for four hours *The note-taking company says it has found no evidence that anyone besides the researcher who reported the flaw viewed the data — but its investigation is still underway, and it has not said whether affected users will be told.* Cloudnote, a note-taking service with 3.2 million registered users, has confirmed that a debugging page containing excerpts of customer prompts was accessible without any login for four hours on Monday. The company removed the page 19 minutes after an outside security researcher alerted it to the problem. According to Cloudnote's review of its server logs, the page displayed prompt excerpts drawn from 1,842 user sessions. Each excerpt was limited to as many as 200 characters — a fragment of any one document, though potentially enough to capture names, instructions or other sensitive text a user had typed. Cloudnote said full documents, passwords and payment information were stored elsewhere and were never present on the page. The company said it has so far found no evidence that anyone other than the researcher accessed the page during the four-hour window. That assurance, however, rests on an internal review that Cloudnote acknowledges is still underway, and the company has not said when it expects the investigation to conclude or whether it will notify the users whose sessions appear in the logs. The exposure amounts to a small fraction of Cloudnote's user base, but it underscores how easily internal tools built for engineers can become public-facing liabilities. Among the questions still unanswered: how the debugging page was published in the first place, whether the logs capture every visit, and why the page required no authentication at all.

Qwen: Qwen3.8 2.4T A95B: # Cloudnote Debugging Page Exposed Prompt Excerpts From 1,842 Sessions *Four-hour exposure affected 1,842 sessions among 3.2 million registered users; no evidence of wider access has been found, but the investigation remains open.* Cloudnote said that a debugging page containing excerpts from customer prompts was publicly accessible without authentication for four hours on Monday, exposing limited data tied to 1,842 user sessions. The company did not say how many individual users were affected, and it has not determined whether any of the exposed sessions included sensitive information. According to Cloudnote, the page displayed no more than 200 characters from each prompt and did not contain full documents, passwords or payment information, which it said are stored elsewhere. The excerpts could include fragments of questions, instructions or text submitted to the service. The company removed the page 19 minutes after a security researcher notified it of the issue. Cloudnote, which has 3.2 million registered users, said logs show no evidence that anyone besides the researcher accessed the page. But the company has not completed its investigation, leaving open questions about whether the excerpts were captured or viewed by any other party before the page was taken down. The incident highlights how internal debugging tools, even when briefly exposed, can create privacy risks for services that process user-submitted text.

MoonshotAI: Kimi K3: # Second Place, First Choice: Benchmark Finds a Cheaper Model Matching the Leader A new benchmark of five language models on real customer-support work delivers a result that should interest anyone buying AI for a help desk: the top two models finished in a statistical dead heat — but one costs more than twice as much as the other, and answers more slowly. ## How the test worked Researchers assembled 200 anonymized customer-support tickets and had each model produce one proposed answer per ticket. Three experienced support managers graded every response blindly, meaning they didn't know which model wrote which answer. Scores combined three qualities a support leader actually cares about: factual correctness, usefulness, and tone. That design is worth dwelling on, because it makes the results more meaningful than a synthetic quiz. These were real tickets, graded by people who do this work for a living, with branding stripped away. ## A tie at the top — until you look at the invoice Model North earned the highest average score at 84.2. Model Vale scored 83.8. The researchers report that this difference is not statistically significant, and readers should take that seriously: on quality, these two models tied. The resemblance ends there. North's estimated cost is $1.90 per 1,000 tickets; Vale's is $0.74 — roughly 40% of the price. Vale also produced answers 18% faster. For a team handling tens of thousands of tickets a month, choosing the statistical runner-up means equivalent answer quality, quicker responses, and a bill that's less than half the leader's. Below the top two, the field thins out: | Model | Avg. score | Cost per 1,000 tickets | |---|---|---| | North | 84.2 | $1.90 | | Vale | 83.8 | $0.74 | | Ember | 79.1 | $0.31 | | Kite | 76.4 | $0.22 | | Harbor | 71.0 | $0.08 | The pattern here matters more than any single row. From Vale down to Ember, the savings shrink to about 43 cents per 1,000 tickets while the average score drops nearly five points. At the bottom, Harbor's rock-bottom price comes with a 71.0 average — cheap answers that are wrong or unhelpful don't stay cheap for long. One caveat: the study only flagged the North–Vale gap as statistically insignificant, so treat the mid-pack differences as directional rather than proven. ## Category strengths cut both ways The benchmark did find texture beneath the averages. North performed best on billing disputes; Vale performed best on technical troubleshooting. That's an argument for matching models to ticket mix rather than crowning a universal winner — but it cuts against paying a premium by default, since the expensive model's edge lives in one category while its cheaper rival leads in another. The study doesn't report how large those category edges were, so treat them as hypotheses to test on your own tickets, not settled facts. ## The finding every buyer should note The most uniform result in the study had nothing to do with rankings: all five models performed poorly when a ticket depended on an image attachment. For support queues full of screenshots, error dialogs, receipts, or photos of damaged products, none of these models can be trusted to close tickets unsupervised. Any deployment plan needs a routing rule that sends image-dependent tickets to humans — or to a different class of system entirely. ## Read the limits before the leaderboard This was one software company, in English, with 200 tickets. Results may not transfer to retail, telecom, or healthcare queues, or to multilingual support — which the researchers did not test at all. Nor did they test agentic behavior: looking up orders, issuing refunds, or using tools. As support AI shifts from drafting replies to taking actions, that omission looms large. And the composite score blends correctness, usefulness, and tone into one number, hiding trade-offs a buyer might care about — a model can be accurate and still rub customers the wrong way. ## Bottom line For this workload, the premium model bought nothing measurable. The sensible default is the cheaper model at the top of the pack, with category performance, image-heavy tickets, and your own pilot data guiding the final call.

Qwen: Qwen3.8 2.4T A95B: # Vale Matches North at a Fraction of the Cost The clearest result from this benchmark is not that one model swept the field. It is that Model Vale essentially matched Model North on overall answer quality while costing far less and responding faster. Researchers evaluated five language models on 200 anonymized customer-support tickets from one software company. Each model generated one proposed reply per ticket. Three experienced support managers graded the answers blindly, using a combined score for factual correctness, usefulness, and tone. Model North posted the highest average score at 84.2, while Model Vale followed closely at 83.8. But the difference was not statistically significant, meaning the benchmark does not provide reliable evidence that North is truly better overall. That near-tie becomes important when cost and speed enter the picture. North’s estimated cost was $1.90 per 1,000 tickets, while Vale’s was $0.74 per 1,000 tickets. Vale also answered 18% faster than North. For a support organization handling thousands or millions of tickets, that combination can change operating economics quickly. If two models deliver statistically indistinguishable quality, the cheaper and faster one deserves strong consideration. The results do not show that Vale is universally superior. North performed best on billing disputes, where careful wording and financial accuracy may matter greatly. Vale performed best on technical troubleshooting, suggesting it may be especially useful for product issues, diagnostics, and step-by-step guidance. A company whose ticket mix is dominated by billing might still favor North, while one facing heavy troubleshooting may find Vale more attractive. The main point is that overall scores do not justify treating North as the only acceptable top-tier model. The other models form a second tier in this test. Ember scored 79.1, Kite scored 76.4, and Harbor scored 71.0. Their costs were lower, with Harbor estimated at just $0.08 per 1,000 tickets. These models may still be useful for drafts, low-risk tickets, or high-volume triage, especially where human review remains mandatory. However, the benchmark suggests caution before assuming the cheapest model is sufficient for direct customer-facing answers. The gap between Harbor and the leaders is much larger than the gap between North and Vale. Readers should also keep the benchmark’s limits in mind. The test used only English-language tickets from a single software company, so results may not transfer to other industries, products, languages, or customer expectations. The evaluation relied on human judgment, which is valuable but can be shaped by the scoring rubric and reviewers’ experience. Because each model produced one final answer per ticket, the results do not reflect retrieval, escalation, tool calls, or agentic actions. The researchers also did not test multilingual support. One especially clear weakness involved image attachments. All five models performed poorly when a ticket depended on an image. That limitation matters for support teams whose customers send screenshots, error messages, invoices, or configuration photos. A model that looks strong in text-only evaluation may still fail when a problem cannot be understood without visual evidence. Teams should test multimodal handling separately rather than assume these scores cover that need. The methodology is otherwise sensible for an early comparison. Anonymized tickets reduce brand bias, blind grading prevents managers from favoring a known model, and three experienced reviewers add practical judgment. Combining correctness, usefulness, and tone also mirrors what customers actually experience: a reply can be factually accurate but still unhelpful if it is vague, abrupt, or poorly structured. Still, 200 tickets is a modest sample. It can reveal broad patterns, but it cannot prove that one model will dominate every support queue. Small score differences should be treated cautiously, especially when statistical testing shows that the North-Vale gap could be due to chance. The more defensible conclusion is about relative value: Vale delivered comparable overall quality at a much lower estimated cost and faster speed. For teams choosing a support assistant, the practical takeaway is to pilot Vale alongside North rather than defaulting to the highest raw score. Measure performance on your own ticket types, languages, attachments, and escalation paths. If Vale continues to match North where it matters, the cost and speed advantages could make it the more sensible production choice.

MoonshotAI: Kimi K3: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin startup's software maps drainage systems and combines them with rainfall forecasts so local governments can deploy maintenance crews before storms hit.* Raincheck, a startup that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The Austin company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. The seed round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has raised $10.5 million to date. CEO Ana Flores said the company's customers often struggle more with prioritization than prediction: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." More detail on the company's approach is available in [Raincheck's published methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its 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 says the new funding will help it hire eight engineers and expand into Florida.

Qwen: Qwen3.8 2.4T A95B: # Raincheck Raises $8 Million to Help Cities Get Ahead of Flood Risk *The Austin company maps drainage systems and pairs those maps with rainfall forecasts, giving local governments a way to prioritize repairs before storms hit.* Raincheck, a maker of flood-risk software for local governments, said today it has raised $8 million in a seed round. The Austin company maps drainage systems and combines those maps with rainfall forecasts, helping cities decide where to send maintenance crews before a storm. The round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has raised $10.5 million in total. Raincheck 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.” The company outlines its approach in [Raincheck’s methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations, and emergency managers retain control over deployments and public warnings. 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.

MoonshotAI: Kimi K3: # HelioDock Launches Night Shift, Software for Running Warehouse Robot Fleets From Afar *The remote-monitoring tool debuts alongside a $24 million Series B, carrying beta results the company is eager to cite — and a hard dependence on networks that don't always cooperate.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor and control their robot fleets remotely, and announced a $24 million Series B funding round led by North Coast Ventures. The twin announcements amount to a bet that the durable money in warehouse automation sits not just in the robots themselves but in the software that supervises them. Hardware sells once; monitoring software bills monthly. A successful Night Shift would give HelioDock revenue that recurs long after its robots leave the loading dock. The new round brings the company's total funding to $39 million. Night Shift is pitched at one of the more stubborn inconveniences of running an automated warehouse: robots that demand human attention at moments when no human is nearby. Automation has redistributed the human side of warehouse work rather than eliminated it, and a single stalled machine can still force an off-duty manager to choose between halted operations and an unplanned trip to the facility — often overnight, when failures arrive on their own schedule. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," said Mara Velez, HelioDock's chief executive. The software is meant to make that drive unnecessary. Oversight of a fleet moves from the warehouse floor to whatever screen the operator has at hand, and remote control allows staff to intervene without being physically present — in effect, supervision without a standing on-site presence. The product does not arrive cold. Night Shift spent 14 months in a private beta involving nine logistics companies and a combined 310 robots, a testing window that stretched well past a year and now serves as the foundation of HelioDock's sales pitch. According to the company, beta customers experienced 17 percent fewer unplanned robot stoppages. HelioDock supplied that figure, and it has not been independently verified; for now it is a marketing claim, albeit one drawn from more than a year of live deployments. If it holds, it would represent a meaningful cut in downtime for operators running dozens or hundreds of machines. Night Shift becomes generally available September 15, priced at $90 per robot per month. The per-robot structure means costs scale with deployment size — larger fleets produce larger monthly bills and larger recurring revenue for HelioDock. The beta also exposed the product's central limitation: Night Shift is only as dependable as the network beneath it. Aaron Pike of Fulcrum Logistics, one of the participating companies, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. When a connection fails, the robots return to their locally configured behavior and remote control becomes unavailable. The machines keep running under their own settings, but the visibility and intervention that justify the subscription drop out with the signal. None of that makes the product unusable at rural sites — Pike said it performed well where coverage was strong — but it does make network quality a purchasing consideration, and an outage removes exactly the capability Night Shift is sold on. Night Shift will not have the category to itself. Relay Robotics and YardOS both market fleet-management software, giving prospective customers immediate alternatives, so HelioDock's early traction will hinge on whether its beta results and pricing persuade operators to choose it over products already in the field.

Qwen: Qwen3.8 2.4T A95B: # HelioDock Launches Night Shift Remote Monitoring for Warehouse Robots *The software arrives after a 14-month private beta and alongside a $24 million Series B led by North Coast Ventures.* HelioDock on Thursday launched Night Shift, software that lets operators monitor autonomous warehouse robot fleets remotely, the company said. The product is intended to keep fleets under human supervision while reducing the need for on-site visits during overnight work periods. The release gives HelioDock a software product tied to the operation of its autonomous warehouse robots. Night Shift spent 14 months in a private beta that involved nine logistics companies and 310 robots. The company also announced a $24 million Series B led by North Coast Ventures, bringing HelioDock’s total funding to $39 million. HelioDock said beta customers experienced 17% fewer unplanned robot stoppages. The company provided the figure, and it has not been independently verified. The metric points to the kind of operational interruption that can halt a robot mid-task and require a person to intervene. In announcing the product, HelioDock emphasized the operational strain that can occur when a robot needs attention outside normal staffing hours. CEO Mara Velez said the software was meant to remove the expectation that a supervisor must physically travel to a site whenever a machine cannot proceed. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” Velez said. Night Shift is intended to let operators see fleet status and respond to problems without being in the building. The 14-month beta allowed HelioDock to test the software with nine logistics companies before commercial release. With 310 robots involved, the test gave the company a large fleet on which to observe performance and constraints. The participation of multiple logistics companies meant the software was evaluated across more than one customer operation before its scheduled release. One constraint described by a beta customer involved network connectivity. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. When connectivity fails, HelioDock’s robots return to locally configured behavior. Remote control becomes unavailable, leaving the robots to operate according to settings

MoonshotAI: Kimi K3: **C** Option C is the strongest lede because it accurately conveys the essential facts—who, what, when, and the scope of the exposure—without distortion. Option A falsely claims private messages were leaked (contents were not sent). Option B reads like PR spin and buries the actual news. Option D fabricates both the exposure of message contents and hacker involvement, neither of which is supported by the facts.

Qwen: Qwen3.8 2.4T A95B: C

MoonshotAI: Kimi K3: # KiteBridge Software Update Triggers $2.8 Million in Duplicate Reimbursements A failed connection between KiteBridge's expense-management platform and its payment processor turned a routine Tuesday morning software update into a costly cascade, generating 13,400 duplicate reimbursements worth $2.8 million before the company cut off automated payments. KiteBridge sells expense-management software, which companies use to approve and pay out reimbursements that employees file for work-related spending. The incident touched 91 corporate customers — the employers whose accounts fund those payouts. KiteBridge said no employees had money withdrawn from their own accounts; the duplicate payments were funded entirely by employer balances. By Wednesday evening, the company said it had recovered $2.3 million, with the remaining roughly $500,000 still pending. The trouble began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its software. Within half an hour, the first customer complaint had arrived. By 10:03 a.m., the company had suspended automated payments entirely — roughly 80 minutes after the update went live, and about 50 minutes after customers started reporting problems. ## A Retry That Paid Twice According to KiteBridge, the root cause was a timeout at its payment processor, the outside service that actually moves the money. When the processor failed to confirm a transaction in time, KiteBridge's fallback system retried it — but instead of recognizing each retry as a second attempt at the same reimbursement, it logged every attempt as a brand-new request and paid it out again. Payment systems are typically built to prevent precisely this. The design principle, known as "exactly-once" processing, ensures a repeated instruction is matched to the original rather than executed twice — much as a careful payroll office would recognize a resubmitted form instead of paying it a second time. That makes the failure especially awkward for KiteBridge. In an April changelog, the company announced that version 4.0 had introduced "exactly-once payment protection." Whatever safeguard shipped two versions ago, it did not prevent the duplication under version 4.2. ## A Slow Disclosure The company's public communication has also drawn scrutiny. Its incident page initially described the problem as delayed reimbursements — language suggesting customers would simply wait longer for money, not that extra money had flowed out. The page did not disclose the duplicate payments until about six hours after payments were suspended, meaning customers spent much of Tuesday afternoon without knowing the true scope of the incident. CEO Lena Ortiz apologized for the incident and said the company would publish a technical report on what went wrong. For the affected customers, the immediate question is the fate of the still-outstanding $500,000. For KiteBridge, the larger question is whether the promised report can explain how a protection the company marketed as a feature failed so completely — and why customers learned the full story hours after the money had already moved.

Qwen: Qwen3.8 2.4T A95B: ## A $2.8 Million Software Error: How KiteBridge’s Expense Update Created Duplicate Payments KiteBridge’s expense-management platform turned a routine Tuesday software update into a costly payment failure, generating 13,400 duplicate reimbursements worth $2.8 million and forcing the company to halt automated payouts for 91 corporate customers. The trouble began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its software. Soon after, the update encountered a payment-processor timeout. In payment systems, a timeout can leave a transaction in an uncertain state: the sender may not know whether a payment succeeded, failed, or is still waiting to be processed. KiteBridge’s fallback system was designed to recover from that uncertainty by retrying payments. Instead, the system treated those retries as separate, new reimbursement requests. That distinction matters because expense software must ensure that one approved claim produces one payment. When the fallback lost track of which requests had already been submitted, it began creating duplicates. The first customer complaint arrived at 9:12 a.m., less than half an hour after the release went live. Over the next stretch of the morning, the error multiplied. By the time KiteBridge suspended automated payments at 10:03 a.m., the company had already issued thousands of repeated reimbursements. Public communication lagged behind the internal response. KiteBridge’s incident page initially described the problem as delayed reimbursements, a characterization that did not mention the extra money moving out of employer-funded accounts. Six hours after automated payments were suspended, the company updated the page to disclose that duplicate payments had occurred. The incident highlights how fragile automated financial workflows can be when a system cannot determine whether a transaction has completed. Expense-management platforms typically connect approval records, corporate funding sources, and payment networks. If a retry mechanism cannot recognize a previous attempt, a single approved expense can quickly become multiple payment instructions. The company said no employees had money withdrawn from their accounts. Instead, employers funded the duplicate reimbursements, leaving corporate finance teams to identify and recover the excess payments. By Wednesday evening, KiteBridge said it had recovered $2.3 million. The remaining balance was still pending, a recovery process likely to require reconciliation across customers, payment processors, and internal ledgers. The failure also raised questions about earlier assurances. In April, KiteBridge’s changelog for version 4.0 said the product had introduced “exactly-once payment protection,” a technical safeguard meant to prevent the same reimbursement from being paid more than once. Version 4.2’s failure suggests that either the protection did not cover the fallback path, was disabled, or could not handle the specific timeout-and-retry condition that occurred. CEO Lena Ortiz apologized and said KiteBridge would publish a technical report. For customers, the immediate task is reconciliation. For KiteBridge, the report will need to explain why a feature designed to prevent duplicate payments failed during one of the most common failure modes in online transactions: a payment request whose outcome is not immediately known.