Head to head: Qwen: Qwen3.8 Max Prime vs grok-4.7

Qwen: Qwen3.8 Max Prime: # 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.

grok-4.7: # 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 Max Prime: # Cloudnote Debugging Page Exposed Prompt Excerpts for Four Hours *The company says only a security researcher appears to have viewed the page, but logs show snippets from 1,842 user sessions.* Cloudnote on Monday left a debugging page accessible without authentication for four hours, exposing short excerpts from customer prompts to anyone who found the URL. The incident was disclosed after a security researcher notified the company. The company confirmed that logs recorded excerpts from 1,842 user sessions. Each excerpt could contain as many as 200 characters from a prompt, a limited but potentially sensitive slice of interactions on a platform with 3.2 million registered users. Because prompts can include personal or work-related information, even truncated records may carry privacy implications. Cloudnote removed the page 19 minutes after the researcher reported it. The company said it has found no evidence that anyone other than the researcher accessed the page while it was live, though it cautioned that its investigation is not complete. It has not explained why the page lacked authentication or whether it was intended for internal use. The exposure does not appear to have reached Cloudnote’s broader data stores. The company said full documents, passwords and payment information are stored elsewhere and were not present on the exposed page. What remains unclear is whether the prompt snippets were indexed, saved or misused before removal, and whether affected users will be notified.

grok-4.7: # Cloudnote debugging page exposed prompt excerpts from 1,842 sessions *A debugging page was open without a login for four hours Monday. Cloudnote says only a researcher appears to have viewed it; the company has not finished its review.* A Cloudnote debugging page containing customer prompt excerpts was accessible without authentication for four hours on Monday. Logs show excerpts from 1,842 user sessions. Those excerpts could include up to 200 characters from a prompt. Set against Cloudnote’s 3.2 million registered users, 1,842 sessions is a limited share of the base. It still matters, because an excerpt can show part of what a user typed into the product. Full documents, passwords and payment information were stored elsewhere and were not present on the page. 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. That statement is the company’s account, not a closed finding. Cloudnote has not completed its investigation, so it is not yet known whether anyone else opened the page during the four hours it was exposed. Confirmed so far are the missing login check, the session count in the logs, the 200-character limit, the 19-minute takedown after notice, and the fact that passwords, payment data and full documents were not on the page. Until that review is done, the clearest picture is a short-lived exposure of prompt fragments, not a confirmed leak of complete customer files.

Qwen: Qwen3.8 Max Prime: # Arcwell Unveils Relay to Suggest Support Answers and Flag At-Risk Customers Arcwell has announced Relay, an artificial-intelligence platform that examines customer support conversations and suggests answers to agents. The release marks Arcwell’s latest effort to apply AI to customer service, with tools intended to speed up agent work and highlight customers who may be preparing to cancel. Relay draws on support transcripts and internal help documents to produce suggested responses during customer interactions. It can also identify customers who may cancel and create summaries of support conversations. Together, those capabilities extend Relay beyond basic answer suggestions and toward the broader workflow of handling and documenting a support case. The platform was tested by 12 companies over six months. Arcwell says agents using Relay closed tickets 14% faster. The company did not release customer-retention results, so it remains unclear whether the cancellation warnings produced measurable reductions in churn. That leaves answer speed as the clearest disclosed result from the testing period. Chief Executive Priya Shah said the product addresses a familiar problem in support work. “Support agents spend too much time searching for information that their companies already have,” Shah said. The comment points to one of Relay’s central tasks: locating relevant information inside existing company materials while an agent handles a customer. Relay costs $65 per agent each month. To work as described, the platform requires access to support transcripts and internal help documents. 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. Those controls give Arcwell a way to address an obvious concern for buyers evaluating AI support software: how much customer material the system can see, whether encrypted data is used to train shared models, and how long the information remains available. With Relay, Arcwell is seeking to help companies improve customer service using artificial intelligence, pairing faster ticket handling with cancellation signals and automatic summaries for support teams.

grok-4.7: # Arcwell Launches Relay, an AI Platform for Support Agents Arcwell has announced Relay, a platform that uses artificial intelligence to examine customer support conversations and suggest answers to agents. In a test with 12 companies over six months, the company says agents using it closed tickets 14% faster. Relay can also identify customers who may cancel and create summaries. Arcwell did not release customer-retention results. The platform costs $65 per agent each month. The speed figure is the performance result the company chose to publish. Identifying customers who may cancel is part of the product, but without retention results there is no public reading on whether those identifications changed who stayed. Summaries sit beside suggested answers as a second cut at the same conversations: not a proposed reply, but a shorter account of what was said. The price, unlike the retention outcome, is stated outright. Chief executive Priya Shah described the bottleneck Relay is meant to relieve. "Support agents spend too much time searching for information that their companies already have," she said. Suggested answers are the direct response to that complaint. The platform examines the conversation and proposes a reply drawn from material the company already holds, rather than leaving the search to the person on the ticket. Identifying customers who may cancel reaches past the wording of a ticket to the relationship behind it. None of those functions runs without the underlying record. Relay requires access to support transcripts and internal help documents. 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. The policy separates two uses of the same history: assisting a customer's own agents, and training models shared with others. How long the history is kept is the buyer's choice, limited to those two windows. For Arcwell, the launch is a move to put artificial intelligence inside routine customer service work. The case it can make today is specific — a six-month trial, a 14% gain in ticket speed, a monthly price, encrypted data, and a ban on shared-model training. The case it cannot yet make is the one its cancellation feature implies. Faster tickets are a reported result. Fewer customers lost are not.

Qwen: Qwen3.8 Max Prime: # HelioDock Launches Night Shift Fleet-Monitoring Software and Raises $24 Million *The company says beta testing reduced unplanned robot stoppages, but rural LTE failures can limit remote control.* HelioDock, a maker of autonomous warehouse robots, launched Night Shift on Thursday, software that lets operators monitor robot fleets remotely. The company also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. The announcement brings together a product launch and a financing round at the point where Night Shift moves from private testing to general availability. The software is meant to reduce the need for overnight, on-site responses to robot faults. HelioDock’s performance figures are company-supplied, and the product’s usefulness depends on network connectivity. Night Shift spent 14 months in private beta. The testing involved nine logistics companies and 310 robots, giving the company a multi-customer trial before commercial release. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages. The figure has not been independently verified. Fulcrum Logistics participated in the beta. Its account provides an example of how the software performed in live logistics facilities, including both strong connectivity and rural LTE failures. The company did not provide additional details showing how the 17% reduction was calculated. HelioDock supplied the figure as part of its launch announcement. The claim is limited to unplanned robot stoppages. Chief Executive Mara Velez described the product as a way to avoid sending staff to a facility for a single confused robot. “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. Night Shift is intended for warehouses where robots operate after regular staffing hours or across multiple shifts. Remote monitoring lets operators watch fleet activity from another location. Where connectivity is available, the software offers a way to supervise robots without being physically present. Beta testing also exposed a constraint. Aaron Pike of Fulcrum Logistics said Night Shift worked well in facilities with strong connectivity but became less useful during rural LTE outages. When connectivity fails, robots return to locally configured behavior, and remote control becomes unavailable. Pike’s comments identify LTE outages as a point where Night Shift’s remote functions lose value. During a failure, operators cannot use remote control, and the robots rely on their local configuration. That makes network reliability a central part of the product’s performance. In facilities with stable connections, operators can maintain remote oversight. In rural sites or areas with intermittent LTE, the software may not remain useful during an outage, and robots will follow their local configuration. The limitation does not remove Night Shift’s monitoring function, but it defines the conditions under which remote control can operate. Facilities with weak or unstable wireless connections may experience periods when operators cannot intervene remotely. North Coast Ventures led the $24 million Series B. The round brings HelioDock’s total funding to $39 million. The company did not specify how the new capital will be used. The $24 million round makes up the majority of the $39 million total funding figure. Night Shift will cost $90 per robot per month and become generally available on September 15. The price applies to each robot in a monitored fleet, making the software a recurring cost tied to fleet size. Fleet-management products are also available from Relay Robotics and YardOS. HelioDock’s launch places Night Shift against those competing offerings as logistics companies choose how to supervise autonomous fleets.

grok-4.7: # HelioDock Starts Selling Remote Oversight for Its Warehouse Robots *Night Shift leaves a 14-month private beta as the company prices monitoring at $90 a robot a month and says a North Coast Ventures-led Series B has lifted total funding to $39 million.* HelioDock on Thursday launched Night Shift, software that lets operators monitor fleets of its autonomous warehouse robots from somewhere other than the building floor, and said it had raised $24 million in a Series B led by North Coast Ventures. The round brings HelioDock’s total funding to $39 million, according to the company. Night Shift, the product paired with that announcement, will cost $90 per robot per month and become generally available on September 15. It spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock says those beta customers experienced 17 percent fewer unplanned robot stoppages. That figure has not been independently verified, so the reduction remains a company-supplied result from a closed trial rather than a measurement an outside party has confirmed. HelioDock makes the robots Night Shift is meant to watch. Chief executive Mara Velez cast the software as a substitute for showing up in person when a machine stalls. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” she said. The line is the company’s clearest statement of what remote monitoring is for: an operator reaches a confused robot without making the trip. It does not describe how often that confusion occurred in the beta, or whether the 17 percent drop in unplanned stoppages came from fewer of those pallet problems specifically. HelioDock has tied the percentage only to unplanned stoppages among beta customers, not to a particular failure type. Distance only helps while the link holds. Aaron Pike of Fulcrum Logistics, a beta customer, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. HelioDock describes the same boundary in operational terms. When connectivity fails, robots return to their locally configured behavior, and remote control becomes unavailable. Night Shift does not keep a session alive through an outage or invent a new instruction set on the spot. The robot falls back to whatever behavior was already configured locally, and the person watching from elsewhere loses the ability to direct it. Pike’s comment and the company’s failure description point at one constraint from two sides: usefulness tracks the network, and rural LTE is one place that tracking breaks. The beta’s scale is the main evidence HelioDock has attached to the launch. Nine logistics companies and 310 robots over 14 months is a defined trial, not an open deployment, and it was private. The stoppage claim sits inside that boundary. HelioDock attributes the 17 percent reduction to beta customers; it has not presented the figure as independently verified, and Pike did not supply a separate count for Fulcrum Logistics. His remarks address connectivity, not the stoppage rate. A reader can know that one participating logistics operator found Night Shift less useful when rural LTE failed, and can know that HelioDock says unplanned stoppages fell across the beta. The notes of the launch do not show whether Fulcrum’s sites were part of the 17 percent calculation, or how the result would change at facilities where the connection drops often enough to push robots back onto local behavior. Price and availability are stated without that caveat attached. At $90 per robot per month, Night Shift is a subscription layered on HelioDock’s robots, with general availability set for September 15. The company has not published a different rate for larger fleets, a minimum number of robots, or a discount tied to the beta. The only fleet size it has given is the beta’s 310 robots, across nine logistics companies. Until September 15, the product remains what the prior 14 months made it: software in private use, not the version offered to buyers at large. The funding announcement arrived the same day, with North Coast Ventures named as the lead of the $24 million Series B. HelioDock says the round brings total funding to $39 million. The company has not itemized other investors in the round or assigned portions of the $24 million to Night Shift, to robot hardware, or to anything else. The capital figure and the software launch share a Thursday dateline; they are not, on the information released, a budget for the product. Other companies already sell tools in the same category. Competing fleet-management products are available from Relay Robotics and YardOS. HelioDock did not publish prices, outage behavior, or stoppage figures for those products, and neither competitor’s results appear alongside the 17 percent claim. The comparison available from the launch is limited to presence: Night Shift enters a field where fleet-management software already exists, at a stated HelioDock price of $90 per robot per month, after a private beta of the size and length the company has disclosed. What an operator gives up when the network fails is the part of Night Shift that Velez’s overnight drive does not cover. Remote control becomes unavailable, and the robot returns to locally configured behavior. In a facility Pike would recognize as strongly connected, the manager in her example might handle a pallet confusion without crossing town. During a rural LTE outage of the kind he described, that remote option is gone, and the local configuration is what the robot still has. The beta ran 14 months with nine logistics companies and 310 robots. HelioDock says unplanned stoppages among those customers fell 17 percent. The figure is still unverified by anyone outside the company, and the software’s own fallback says the remote shift ends where the connection does.

Qwen: Qwen3.8 Max Prime: # LedgerLark Looks at the Paperwork After the Trade When a trade is executed, the attention often goes to price, position and market risk. What follows can be less visible: matching confirmations, checking custodian records and fixing the small errors that keep an internal ledger from agreeing with outside records. LedgerLark is building AI agents for this stage of the work at small investment firms. The startup’s product compares documents and records from confirmations, custodians and internal ledgers, then proposes corrections for a person to review and approve. It does not execute trades or move money. The idea came from co-founder Mina Batra’s experience in operations at a regional brokerage, where she spent seven years. In one episode described by the company, a six-person team spent three days resolving a trade mismatch caused by a copied account number. That kind of delay is not dramatic by market standards, but it can become expensive when it ties up staff, leaves positions unclear and requires repeated follow-up. Batra founded the company with Owen Carr, who previously worked on reconciliation systems at a large bank. Their combined background points to a familiar problem in financial operations: many firms still rely on manual checks, spreadsheets and email chains even when trading itself has become highly electronic. LedgerLark’s agents are aimed at that gap. The software reviews post-trade data across different sources and flags items that do not match. Rather than asking operations teams to start with a blank screen or a long queue of exceptions, it presents proposed fixes that a human must approve. The company describes the approach as assisted work, not autonomous control. That distinction matters in an industry where errors can be costly and regulators expect firms to maintain strong oversight. By keeping corrections subject to human sign-off, LedgerLark appears to be addressing a practical concern for financial institutions that may be wary of software taking direct action over financial records. Early traction is modest but specific. The company says five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. It also reports that one customer reduced unresolved daily exceptions from 70 to 18. Since that result comes from the company and likely reflects a particular client’s workflow, it is best read as an early indication rather than a broad guarantee. The funding is also early-stage. LedgerLark has raised a $3.5 million seed round led by Motive Row. The capital is likely to be used to deepen product integration, support more clients and navigate the security reviews that commonly come with selling software to financial institutions. Those reviews are one of the main obstacles the company faces. Even small investment firms may need to examine data access, permissions, audit trails and vendor risk before adopting a new tool. Lengthy procurement and security processes can slow sales, especially for a young company selling into a cautious sector. Competition is another question. LedgerLark is not entering an empty field. Established reconciliation vendors already serve parts of the market, and some larger institutions have internal software built over many years. The startup will need to show that its agents are easier to deploy, more adaptable to smaller teams, or more useful in handling messy post-trade data than existing options. The opportunity depends on how much pain remains in these workflows. If LedgerLark can reduce repetitive checking without adding new operational risk, it may earn a place in the daily routine of small asset managers and brokerages. If not, customers may decide that incumbent tools, even imperfect ones, are easier to keep. For now, LedgerLark presents a straightforward proposition: use AI to find and explain mismatches after trades, then let people decide what to change. In financial operations, that restraint may be as important as the automation itself.

grok-4.7: # Startup Spotlight: LedgerLark Looks at the Cleanup After the Trade Investment software often treats the trade as the main event. LedgerLark starts later, after the order is done and the paperwork begins to disagree with itself. The company builds AI agents for post-trade operations at small investment firms. The agents compare confirmations, custodian records and internal ledgers, then propose corrections for human approval. They do not execute trades or move money. That limit is part of the design: the product is meant to sit inside a process that already has controls, not to replace the people who approve a change. Co-founder Mina Batra spent seven years in operations at a regional brokerage. She began exploring the idea after seeing a six-person team spend three days resolving a trade mismatch caused by a copied account number. The mistake was ordinary. The time was not. That gap—between a small copying error and days of staff attention—is the problem LedgerLark is organized around. Co-founder Owen Carr built reconciliation systems at a large bank. LedgerLark’s premise is that smaller investment firms face a similar mismatch between records, without the headcount or the in-house systems a large institution can keep. That is a hypothesis, not yet a market map. It explains why the company is starting with small firms rather than trying to replace the reconciliation stacks already running inside big banks. Early revenue is small and specific. Five firms are paying customers. Annual contracts range from $36,000 to $110,000. The spread matters. A $36,000 contract can be a trial with a budget attached; a $110,000 contract has to survive a harder internal conversation. Neither figure, across five customers, says how long those firms will stay or whether the next sale will look like these. The company has raised a $3.5 million seed round led by Motive Row. For a product with a narrow workflow and a handful of contracts, that is enough capital to keep building and to learn where buyers stall. It is not evidence that those obstacles have been cleared. In this market, one of those points is procedural. Financial institutions may require lengthy security and procurement reviews, and a review can outlast the quarter in which a deal was first discussed. One figure the company has offered should stay attributed. The company says one customer reduced its unresolved daily exceptions from 70 to 18. If the definition of an unresolved exception was stable, that is a sharp change in a daily queue. It is also one customer, reported by the company, and it is not presented here as an audited result. Exception counts can fall because software helped, because volume changed, because staff changed, or because the firm started counting differently. Until more customers report the same kind of before-and-after, the number is a lead, not a conclusion. Competitors are not hypothetical. Established reconciliation vendors already occupy parts of this work, and internal bank software covers other parts for institutions that built it. LedgerLark’s practical difference, so far, is an agent that compares confirmations, custodian records and internal ledgers and stops at a proposed correction. That may suit a small firm that wants help drafting fixes and does not want software that can send an order or move cash. It may also be a thin difference if incumbents add a similar draft-and-approve step, or if buyers decide the break is better handled inside tools they already pay for. Five paying firms, and a reported drop from 70 unresolved daily exceptions to 18, are reasons to keep looking. They are not yet a pattern. The insight is specific: a small copying error can consume days. Whether an agent can reliably take part of that work, without moving money, is still the open test.

Qwen: Qwen3.8 Max Prime: 6

grok-4.7: 6

Qwen: Qwen3.8 Max Prime: # Raincheck Raises $8 Million to Help Cities Prepare for Flooding *The Austin software company maps drainage systems and pairs them with rainfall forecasts to guide pre-storm maintenance.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in a seed round led by Basin Ventures. Existing investor Long Mile also participated, bringing Raincheck’s total funding to $10.5 million. The company maps drainage systems and combines those maps with rainfall forecasts, helping cities decide where to send maintenance crews before a storm. 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 describes its approach in [Raincheck’s methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana, though it 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.

grok-4.7: # Raincheck Raises $8 Million for Flood-Risk Software *The Austin company maps drainage systems and rainfall forecasts so local governments can send maintenance crews before a storm.* Raincheck, which makes flood-risk software for local governments, announced today that it has raised $8 million. The Austin company maps drainage systems and combines those maps with rainfall forecasts, an approach set out in its [flood-risk methodology](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews before a storm. The seed round was led by Basin Ventures. 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. 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.

Qwen: Qwen3.8 Max Prime: # For Support AI, Vale Matches North at Lower Cost In a benchmark of five language models on customer-support tickets, the most important result is not that one model swept the field. It is that Model Vale came within a whisker of Model North’s score while costing far less and responding faster. North averaged 84.2, Vale averaged 83.8, and researchers found the difference was not statistically significant. For teams choosing a support-drafting model, that near-parity matters more than the tiny gap at the top. The clearest takeaway is value. Vale’s estimated cost was $0.74 per 1,000 tickets, compared with $1.90 for North. Vale was also 18% faster. If two systems produce answers that human graders cannot reliably distinguish, the cheaper and faster option has a strong claim for production use. That does not make Vale universally better: North has advantages in particular areas, and support workflows are rarely one-dimensional. But many organizations may not need to pay a premium for North-level quality across routine tickets. The test was narrow enough to interpret carefully. Researchers used 200 anonymized customer-support tickets from one software company. Each model produced one proposed answer per ticket. Three experienced support managers graded the answers blindly, without knowing which model wrote them. Scores combined factual correctness, usefulness, and tone—three qualities that matter in support, where an accurate reply can still fail if it is abrupt. The grading approach reflects practical customer-service judgment, not just automated accuracy metrics. The results are best read as a comparison of draft quality for human review, not proof that any model can operate without oversight. The models formed a rough quality ladder. North and Vale led, with Ember and Kite in the middle and Harbor trailing. But the details matter less than the pattern: the strongest models were close to each other, while cheaper models showed larger quality drops. Ember at 79.1 and Kite at 76.4 may be acceptable for some low-risk queues, especially with human editing. Harbor’s 71.0 score and $0.08 cost point to a different trade-off: it may suit teams where budget is tight and error tolerance is high. The benchmark does not say those models are useless; it says their savings come with more visible quality risk. Specialized strengths also complicate a simple ranking. North performed best on billing disputes, where precision, policy awareness, and careful wording can matter greatly. Vale performed best on technical troubleshooting, which often involves reasoning through symptoms and guiding customers toward fixes. A support organization dominated by billing issues might prefer North despite Vale’s overall value. One focused on product troubleshooting might find Vale even more attractive. Routing different ticket types to different models could be worthwhile, though it adds operational complexity and requires its own evaluation. One limitation deserves emphasis: all five models performed poorly when a ticket depended on an image attachment. This is a major caveat for real support environments, where screenshots, error dialogs, receipts, and annotated images are common. A model that looks strong on text-only tickets may fail when the customer’s core evidence is visual. Until image understanding is tested and improved, the benchmark should not justify automation for attachment-heavy cases. Human review remains especially important there. The study also leaves questions unanswered. It used English-language tickets from one company, so results may not transfer to other products, customer bases, or support styles. It did not test agentic actions, tool use, or multilingual support. A model that writes well may still be unsuitable if it must query a knowledge base, update a ticket, issue refunds, or follow escalation rules. The benchmark measures proposed answers, not full workflow safety. For buyers, the practical lesson is to match the model to the queue. Vale looks like the strongest general-purpose starting point based on comparable quality, lower cost, and faster responses. North remains compelling for billing-heavy or high-stakes text support. Lower-cost models may work with stronger human oversight. And any deployment should be tested against real tickets, including attachments and policy edge cases, before managers trust it with customers. The clearest finding is not that Vale beat North. It is that the gap was small enough to make cost, speed, and ticket mix the deciding factors.

grok-4.7: # Near-Equal Quality, Unequal Cost: What a Support-Ticket Test Actually Shows The clearest result in this benchmark is not that one model “won.” It is that the two strongest models were, for practical purposes, tied on quality, while one of them was substantially cheaper and faster. Model North scored 84.2 and Model Vale 83.8 on a combined grade of factual correctness, usefulness, and tone. That gap was not statistically significant. Vale, however, was estimated at $0.74 per 1,000 tickets against North’s $1.90, and it answered 18 percent faster. For a support team choosing a default model, that is the decision the data actually supports: paying more than twice as much for North did not buy a measurable quality advantage on this workload. ## How the test was run Researchers gave five language models the same set of 200 anonymized customer-support tickets from a single software company. Each model produced one proposed answer per ticket. Three experienced support managers graded those answers blindly, so they did not know which model wrote which reply. Scores combined factual correctness, usefulness, and tone rather than treating any one of those as the whole job. That design is enough to compare draft quality under a shared rubric. It is not a full simulation of a support desk. The models were not asked to take actions, call tools, look up accounts, or handle languages other than English. Each ticket received a single proposed answer, so the test does not show how a model behaves across a back-and-forth thread or after a human edit. ## Quality holds, then falls Read as a cost curve rather than a ranking table, the results split into two stories. At the top, North and Vale are interchangeable on the overall score. Below them, cheaper models give up real ground. Ember, at $0.31 per 1,000 tickets, averaged 79.1. Kite, at $0.22, averaged 76.4. Harbor, at $0.08, averaged 71.0. Those are not rounding errors. A team that moves from the top pair toward the cheapest option should expect noticeably weaker drafts, not a free lunch with a smaller bill. The interesting inefficiency is between North and Vale, not between Vale and Harbor. Vale’s price is well under half of North’s, and its latency advantage is large enough to matter in a queue. Harbor is an order of magnitude cheaper than North, but the score drop of more than 13 points is the kind of gap managers are likely to feel in factual misses, thin answers, or tone that needs rewriting. Cost savings only count if the remaining human work does not eat them back. ## Strength depends on the ticket Overall parity also hides a split by ticket type. North did its best work on billing disputes. Vale did its best work on technical troubleshooting. A buyer who mostly handles invoices, refunds, and plan changes has a reason to look harder at North even if the headline scores match. A buyer whose queue is dominated by setup problems, errors, and how-to questions has a reason to prefer Vale beyond price and speed. The benchmark does not publish the size of those category gaps, so they should be treated as a routing hint, not as proof that one model should own a whole queue. The shared failure is clearer. All five models performed poorly when a ticket depended on an image attachment. That is not a small-model problem or a cheap-model problem. Screenshot-heavy support—error dialogs, broken layouts, hardware photos—remains a weak spot across the set. Any rollout that assumes the model can “see” what the customer attached will overstate what this test measured. ## What the numbers do not license Several limits should travel with the finding. The tickets were English-language and came from one software company, so the ranking may not hold for another product, another tone of voice, or another market. Managers graded proposed answers, not resolved tickets, customer satisfaction, or handle time after a human review. Estimated cost per 1,000 tickets is a useful comparison, not a production invoice: real spend will move with prompt length, retries, and whatever review step sits in front of the customer. The study also did not test agentic actions, tool use, or multilingual support. A model that drafts a good reply may still be a poor choice if the job is to issue a refund, query an order system, or answer in another language. Small differences should not be promoted into a winner. North’s 0.4-point edge over Vale is the kind of gap this design cannot distinguish from noise. Larger separations—from the top pair down through Ember, Kite, and Harbor—are the ones that justify a quality-versus-cost conversation. ## The practical reading For this company’s English support tickets, Vale is the default that the evidence favors: essentially the same graded quality as North, at a much lower estimated cost, with faster answers, and a relative strength on technical troubleshooting. North remains the better candidate where billing disputes dominate and the extra spend is acceptable. Cheaper models are a deliberate trade, not a tie. And none of the five should be trusted, on this evidence, to carry tickets whose meaning lives in an image. The next useful test is not another overall average. It is a routed trial—billing to one model, troubleshooting to another, images kept with a human—measured on resolution, not just on a blind draft score.

Qwen: Qwen3.8 Max Prime: C

grok-4.7: C

Qwen: Qwen3.8 Max Prime: # The Last Program at Foundry Hall Foundry Hall will close on October 31, ending a run that began in 2018 with a simple promise: give Milwaukee founders a place to build. The nonprofit startup accelerator has supported 212 companies, which now employ 1,140 people. By Halloween, its rooms will be empty, its program calendar finished, and six staff members will be out of work. The shutdown is not the result of a slow decline. Foundry Hall has enough money to operate through October, but not enough to fund another full cohort of startups. The gap appeared when Blank River Bank, its largest corporate sponsor, declined to renew a sponsorship that had supplied 38% of the organization’s annual budget. The bank said its decision reflected a shift in community investment priorities and was not tied to Foundry Hall’s performance. That distinction may matter on paper, but inside the building it feels abstract. Executive Director Camille Ross described the final months simply: “We kept waiting for the replacement check that never came.” The sentence has the flatness of someone who has already spent months making contingencies and now has only the fact left to say. For founders, the loss is personal. Dev Malik remembered taking his first investor meeting at Foundry Hall after years of pitching from the back table of a coffee shop. His story is a version of the hall’s larger argument: that early companies need more than capital. They need a desk, a conference room, someone who knows whom to call, and the credibility that comes from being part of a place. Twelve startups currently using the workspace must leave. Some are small enough to scatter into home offices and shared tables. Others have employees, inventory, client meetings, and deadlines that do not pause for a landlord’s timeline. Deputy Mayor Susan Cole said the city is talking with the building owners and local universities about space for the affected companies. That effort may soften the blow, but it cannot recreate what Foundry Hall had accumulated: routine, trust, and a common language for starting something uncertain. The staff will spend the final weeks doing the quiet work of closing. They will answer questions, pack files, introduce founders to whoever might help, and try to make the last program feel less like an ending than a transfer of energy. When the doors lock at the end of October, Milwaukee will still have its startups. The question is whether the city will replace the room where so many of them first learned to stand.

grok-4.7: The lights in Foundry Hall still came on at seven. Someone had forgotten, or refused, to change the timer. On the second floor of the old Milwaukee warehouse, desks faced the river the way they had since 2018, when the nonprofit accelerator first opened its doors to people who arrived with ideas and not much else. By October 31 those lights would go dark for good. Camille Ross, the executive director, had spent the summer doing the arithmetic in public and in private. Blank River Bank, the hall’s largest corporate sponsor, had supplied 38 percent of the annual budget. When the bank declined to renew, the missing share was not a line that could be trimmed. It was the difference between another year and a closing date. “We kept waiting for the replacement check that never came,” Ross said. What remained would carry operations through October. It would not carry another full program. The bank described the decision as a shift in its community investment priorities, unrelated to the hall’s performance. The statement was careful and complete, and it left out the six employees who would lose their jobs, and the twelve startups still working in the shared space who would have to pack up and go. Over seven years the hall had supported 212 startups. Together those companies employed 1,140 people—an abstract total until you stood in the common room and heard someone explain a payroll delay, or watched a founder tape a revised deck to the glass. Dev Malik, whose company had come through early, remembered his first investor meeting in that building. Before Foundry Hall, he said, he had been pitching from the back table of a coffee shop. The room had not made the company. It had made the meeting possible, which was a smaller claim and, to him, the one that mattered. The city had begun looking for somewhere else. Deputy mayor Susan Cole said officials were talking with building owners and local universities about space for the companies still inside. No one was promising the same rent, the same printers, or the same person at the front desk who knew which founders took their coffee black. A room could be found. A place was harder. In the last weeks people packed without much ceremony. Whiteboards were wiped and then, sometimes, written on again, as if a sentence left behind might still be useful to whoever came next. Nameplates came down. A couch that had outlasted three cohorts was claimed by no one and left where it was. The river outside kept its own schedule. Foundry Hall had always been a bet that a city could make room for beginnings. The bet had worked often enough to be counted in jobs and companies and years. It had not worked well enough, in the end, to be renewed. Ross walked the floor in the late afternoon, checking locks she would soon hand over. She did not give a speech. The hall had never been the kind of place that needed one. What it had needed was a check, and a reason to stay open past the end of the month. Neither arrived.

Qwen: Qwen3.8 Max Prime: # QueueZero Launches AI Meeting Assistant Proxy After Demo Assigns Tasks to CEO QueueZero on Wednesday launched Proxy, an artificial intelligence assistant designed to attend meetings, take notes and assign follow-up work to software agents, extending the company’s push into the crowded market for workplace automation. The product was demonstrated during a launch event, where a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy, apparently treating the remark as a command rather than a joke, generated 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, according to the company. The moment gave QueueZero an unexpected opportunity to explain how Proxy handles instructions. The company said the assistant requires confirmation before any external message is sent, a safeguard intended to prevent embarrassing or premature communications. Internal tasks, however, can be created automatically, reflecting a common design in AI workplace tools that prioritize speed while leaving sensitive actions under human control. QueueZero said administrators can disable automatic task creation, allowing companies to set the level of autonomy that fits their workflows. The setting is likely to be a central consideration for businesses adopting AI agents, particularly in environments where informal comments, side conversations or offhand remarks could be misread as operational directives. The incident also highlighted a broader challenge facing developers of meeting software: distinguishing between conversation and command. As AI assistants become more active participants in calls and conference rooms, vendors are being forced to define how language becomes work, and what guardrails are needed when an employee makes a casual statement that a machine might interpret as a request. Proxy will enter public beta next week and will cost $30 per user per month. That pricing places it in the midrange of AI workplace tools, where buyers increasingly expect not only transcription and summaries, but some ability to trigger action after a meeting ends. For QueueZero, the launch mishap may have been useful evidence. The company can point to a live example of what its software can do, what it might misunderstand, and how much control administrators retain. If nothing else, the debut showed that an AI assistant can be very attentive to the wrong sentence.

grok-4.7: # QueueZero’s Proxy Turns an Audience Remark Into Tasks for Its CEO QueueZero launched Proxy, an AI meeting assistant, on Wednesday. The product can attend a meeting, produce notes, and assign follow-up work to software agents. Its debut also showed how quickly a spoken aside can become an item on someone’s list. 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 and assigned both to QueueZero chief executive Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task, and deleted the first. The moment was the clearest illustration yet of a rule the company is now stressing in public. Proxy requires confirmation before an external message is sent. Internal tasks can be created automatically. Administrators can disable automatic task creation if they do not want remarks converted into assignments without review. The incident did not appear to interrupt the rest of the introduction. After dismissing the first task, Wu used the second to distribute the pricing information the audience had asked for. Proxy costs $30 per user per month and enters public beta next week. QueueZero is pitching Proxy as more than a transcription tool. Meeting software has spent the past two years adding summaries and searchable transcripts. Proxy’s claim is that notes should lead directly to work: an agent can be told to draft a follow-up, update a record, or open a ticket rather than leaving an action item in a document no one reopens. That design assumes the system can tell an instruction from a comment, and that organizations will accept some assignments appearing before a person has reviewed them. Wednesday’s demonstration tested that assumption in front of customers. The sentence was not addressed to Proxy, and it mixed a joke at the chief executive’s expense with a practical request. The assistant did not distinguish between the two. Both became tasks. Only Wu’s decision to delete one of them kept “Stop talking” from remaining on his list. The assistant, unlike the room, showed no sign it had heard a joke. QueueZero’s confirmation requirement is meant to limit what can leave the company without a person in the loop. An internal task can be created on its own. A message to someone outside cannot. In this case, both outputs stayed inside that boundary. One was discarded. The other was a pricing request Wu chose to fulfill himself, rather than an email Proxy sent on its own. The company did not describe how Proxy decides that a remark is an instruction, or how it assigns work when the speaker does not name the product or an agent. Those questions will matter once the public beta begins. Customers will be able to turn automatic task creation off. They will also be able to leave it on, and then discover which comments the system treats as orders.

Qwen: Qwen3.8 Max Prime: # KiteBridge Update Turns Retries Into $2.8 Million in Duplicate Reimbursements When KiteBridge deployed version 4.2 of its expense-management software Tuesday morning, a failure in its payment-handling system turned legitimate reimbursement retries into fresh payouts, generating 13,400 duplicate reimbursements worth $2.8 million and affecting 91 corporate customers. The sequence began at 8:45 a.m., when KiteBridge deployed version 4.2. The problem surfaced in the pathway that connects reimbursement approvals to actual payments. Within minutes, the system encountered a timeout from an outside payment processor. In such cases, software commonly uses a fallback mechanism to decide whether a transaction should be attempted again or marked unresolved. KiteBridge’s fallback failed in a consequential way. Rather than recognizing payment retries as attempts to complete an existing reimbursement, the system treated them as separate, new reimbursement requests. That distinction mattered: once the software accepted the retries as valid new claims, it authorized duplicate payments for expenses that had already been settled. In payment systems, this is the kind of error that can multiply quickly, because each retry can create another record that looks legitimate to automated approval tools. The first customer complaint reached KiteBridge at 9:12 a.m., less than half an hour after deployment. Over the next hour, the company worked to determine whether the issue was limited to delayed transactions or had already produced unauthorized payments. At 10:03 a.m., KiteBridge suspended automated payments, halting further reimbursements while engineers investigated. Publicly, the company’s messaging lagged behind the technical reality. KiteBridge’s incident page initially described the issue as delayed reimbursements, a characterization that did not alert customers to duplicate payments. The page disclosed duplicate payments only six hours after automated payments were suspended, leaving affected businesses to piece together what had happened. By Wednesday evening, KiteBridge said it had recovered $2.3 million of the duplicated funds. The remaining amount was still pending, as the company worked with payment processors and customers to reverse or reconcile the transactions. KiteBridge said employees did not have money withdrawn from their accounts. Instead, employers funded the duplicate payments, leaving corporate finance teams to recover the excess amounts or correct reimbursement records. The failure is particularly notable because KiteBridge had previously advertised safeguards designed to prevent this kind of duplication. In an April changelog, the company said version 4.0 had introduced “exactly-once payment protection,” a phrase suggesting the software should allow a payment to be processed one time, even if network problems force a retry. In distributed payment systems, that guarantee is difficult to maintain when external processors time out, because the platform must know whether a transaction completed, failed, or never started. Tuesday’s incident raised questions about how that protection worked with version 4.2 and the fallback logic that handles payment-processor interruptions. Chief Executive Lena Ortiz apologized and said the company would publish a technical report. For now, the episode is a reminder that financial automation requires careful handling of failures, retries, and edge cases, where one mistaken assumption can turn a routine recovery step into a costly duplicate transaction.

grok-4.7: # KiteBridge Software Error Sent $2.8 Million in Duplicate Reimbursements A software failure at expense-management company KiteBridge generated 13,400 duplicate reimbursement payments totaling $2.8 million on Tuesday, briefly undoing a safeguard the company had advertised as protection against exactly this kind of mistake. Ninety-one corporate customers were affected. By Wednesday evening the company had recovered $2.3 million; the remaining funds were still pending. The trouble began with a routine release. At 8:45 a.m. Tuesday, KiteBridge deployed version 4.2 of its software, which employers use to approve employee expenses and send money back to workers who paid out of pocket. Less than half an hour later, at 9:12 a.m., the first customer complaint arrived. What looked at first like a delay was in fact a cascade of repeated payments. The technical cause was a timeout at a payment processor—the outside service that actually moves money between bank accounts. When that processor was slow to confirm a transfer, KiteBridge’s fallback system was supposed to try again safely. Instead, it treated each retry as a brand-new reimbursement request. In ordinary payment systems, a timeout does not always mean the money failed to move; it can mean the confirmation simply arrived late. A well-designed fallback checks whether the original request already went through before sending another. KiteBridge’s did not. Each unanswered attempt was recorded as a fresh obligation, and the software kept paying. Automated payments continued until 10:03 a.m., when KiteBridge suspended them. In that window the system issued thousands of duplicates. Employees did not lose money from their own accounts. Employers fund reimbursements, so the duplicate charges hit company accounts, not workers’ personal balances. That limited the harm to individuals, but it left finance teams at 91 companies sorting unexpected outflows on a Tuesday morning. Recovery moved quickly once the automation was stopped. By Wednesday evening, $2.3 million of the $2.8 million had been clawed back or otherwise recovered. The rest remained outstanding, pending further reconciliation with customers and payment partners. KiteBridge has not said how long that remaining work will take. The failure is especially awkward because of a promise the company made only months earlier. An April changelog—the public note that accompanies a software update—said version 4.0 had introduced “exactly-once payment protection,” a standard industry goal meaning a payment should be sent once and only once, even if a network hiccup forces a retry. Version 4.2, released Tuesday, was a later update on that same product line. The incident shows that the protection either did not cover this timeout path or did not survive the newer release. KiteBridge has not yet explained which. Communication lagged the facts. The company’s public incident page initially described the problem as delayed reimbursements, a framing that suggested money was slow to arrive rather than being sent too often. The page first disclosed the duplicate payments six hours after automated payments were suspended—well after customers had already begun reporting the charges. For finance teams trying to freeze accounts or warn banks, those hours mattered. Chief executive Lena Ortiz apologized and said the company would publish a technical report. That report will need to explain how a fallback meant to recover from a timeout instead multiplied payments, why the exactly-once protection advertised in April did not stop it, and why the first public description of the outage omitted the duplicates. Until then, the clearest measure of the failure remains the one customers already have: $2.8 million sent twice, most of it since recovered, and a safeguard that did not hold.