Head to head: Z.ai: GLM 5.3 vs MiniMax: MiniMax M3

Z.ai: GLM 5.3: # 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.

MiniMax: MiniMax M3: # 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.

Z.ai: GLM 5.3: # Cloudnote Says Exposed Debugging Page Leaked Prompt Excerpts From 1,842 Users *The AI note-taking company removed the unauthenticated page within minutes of a researcher's warning, but its investigation is ongoing.* Cloudnote disclosed this week that a debugging page containing excerpts of customer prompts was publicly accessible without a password for four hours on Monday. Logs reviewed by the company show the page displayed snippets from 1,842 user sessions, with each excerpt capturing up to 200 characters of a prompt. The company removed the page 19 minutes after an outside researcher notified it of the exposure. Cloudnote says its logs show no evidence that anyone other than the researcher accessed the page, though it cautioned that its investigation is not yet complete. The company emphasized that the exposed page did not include full documents, passwords or payment information, all of which are stored in separate systems. The leaked material was limited to the prompt excerpts shown on the debugging page. The incident is notable in part because of Cloudnote's scale: the service has 3.2 million registered users. While 1,842 sessions represent a small fraction of that base, even short excerpts of prompts can reveal sensitive information about what users are working on or discussing. What remains unclear is how the debugging page came to be left unauthenticated, and whether Cloudnote can definitively rule out access by parties other than the researcher. The company has not said when it expects its investigation to conclude.

MiniMax: MiniMax M3: **Cloudnote Says Unauthenticated Debugging Page Exposed Excerpts From 1,842 Users** *The note-taking platform reported that a misconfigured internal page was online for four hours on Monday before being removed, though the full scope of access remains under investigation.* Cloudnote, a cloud-based note-taking service with 3.2 million registered users, disclosed Monday that an internal debugging page was accessible to the public without authentication for approximately four hours. Company logs indicate the page contained excerpts from 1,842 user sessions. Each excerpt could include up to 200 characters from a user's prompt, according to Cloudnote, potentially capturing sensitive fragments of customer work. The company emphasized that full documents, passwords and payment information were stored on separate systems and were not present on the exposed page. A security researcher alerted Cloudnote to the issue, and the page was removed 19 minutes later, a spokesperson said. Cloudnote said it has identified no evidence that anyone besides the researcher accessed the page during the exposure window, though the company acknowledged its investigation is not yet complete. The incident highlights the recurring risk of internal debugging tools being deployed without proper access controls. Cloudnote has not said whether affected users will receive direct notification or what technical changes it plans to prevent a recurrence.

Z.ai: GLM 5.3: # LedgerLark Wants AI Agents to Take the Drag Out of Post-Trade Ops The most expensive errors in investment operations are rarely the dramatic ones. They're the copied account number, the mismatched settlement date, the confirmation that quietly disagrees with the custodian's records — small discrepancies that can tie up a lean team for days. Mina Batra knows this firsthand. Before co-founding LedgerLark, she spent seven years in operations at a regional brokerage, where she says she watched a six-person team spend three days untangling a single trade mismatch caused by a mis-entered account number. The episode stuck with her. If a handful of people could burn most of a week on one clerical slip, the economics of small firms' back offices seemed badly out of whack — especially compared with the large banks that can afford armies of engineers to automate the same work. That observation became LedgerLark, which builds AI agents for post-trade operations at small investment firms. Batra's co-founder, Owen Carr, built reconciliation systems at a large bank, giving the pair complementary exposure to both ends of the market: the automation ceiling at big institutions and the manual grind below it. ## What the Agents Do LedgerLark's agents sit across three sources of truth that firms already maintain: trade confirmations, custodian records and internal ledgers. The software compares them, flags discrepancies and — this is the notable part — proposes corrections for a human to approve. That last step is deliberate. The product does not execute trades or move money. In a domain where a wrong write action could create the very mismatch it was meant to fix, keeping a human in the approval loop is both a design choice and likely a selling point when prospective customers ask hard questions about what an "agent" is actually allowed to do. For operations teams that already spend their days reviewing exceptions, the pitch isn't that AI will replace them. It's that the first pass — the tedious comparison and triage work — can be automated, leaving people to judge edge cases. ## Early Traction, With Caveats LedgerLark says five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. The company also says one customer reduced its unresolved daily exceptions from 70 to 18. If that figure holds across customers, it suggests real operational relief rather than a marginal efficiency gain. The usual cautions apply: these are company-reported numbers, and a five-customer base is early by any measure. What's arguably more telling is the pricing. Contracts starting at $36,000 a year imply firms see enough value to write meaningful checks, not experiment with a cheap tool. But it will take a larger sample — and independent results — to know whether the exception-reduction story generalizes beyond early adopters. ## The Road Ahead LedgerLark has raised a $3.5 million seed round led by Motive Row, capital that will presumably go toward deepening the product and adding customers. The competitive landscape is the harder part to handicap. Established reconciliation vendors already serve this market, and large banks largely handle the problem with internal software. Some of those vendors will add AI features of their own; some may already be pitching similar capabilities. LedgerLark's bet seems to be that small firms — too small to build their own tooling, historically underserved by enterprise vendors with heavyweight implementations — represent a gap incumbents won't chase aggressively. That's plausible, but unproven. Selling into financial institutions brings its own friction. Security reviews, procurement cycles and vendor risk assessments can stretch sales timelines well past what a typical software startup expects, and a seed-stage company navigating that process for the first time will feel it. Whether LedgerLark can keep a reasonable sales cadence through those reviews may matter as much as the technology itself. ## The Larger Question There's a genuine open question underneath LedgerLark's thesis: are small investment firms underserved because vendors ignored them, or because the market is smaller and stickier than it looks? The five paying customers suggest the demand is real. The next fifteen or twenty will tell us more. For now, the company is attacking a problem that everyone in the industry recognizes — exception handling is slow, manual and expensive at small scale — with a design (propose, don't execute) that fits the risk tolerance of its buyers. That's a reasonable foundation. What it becomes depends on whether the agents' correction proposals prove reliable enough, across enough messy real-world data, for operations teams to trust the first pass and spend their three days elsewhere.

MiniMax: MiniMax M3: # Startup Spotlight: LedgerLark Tackles the Boring Back Office of Investing When a trade settles at the wrong account because someone copied a number incorrectly, the financial consequence is usually small. The operational consequence can be enormous. Mina Batra, now co-founder of LedgerLark, watched a six-person team spend three days unwinding a single trade mismatch at a regional brokerage. That experience is the seed of a new company building AI agents for post-trade operations at small investment firms. Batra spent seven years in operations at that brokerage, learning the rhythm of confirmations, custodian feeds, and the daily scramble to reconcile them. Her co-founder, Owen Carr, came at the problem from the other direction, having built reconciliation systems at a large bank. The two started LedgerLark on a simple premise: the manual work of comparing trade confirmations against custodian records and internal ledgers, then proposing corrections, is repetitive enough that software can do most of it. ## What the product does LedgerLark's agents read the three sources of truth that have to agree before a trade is considered settled, flag mismatches, and suggest fixes. A human reviewer approves each proposed correction. The product does not execute trades or move money, a boundary the founders describe as deliberate. The agents sit between the data and the decision, proposing changes rather than implementing them. That separation matters for two reasons. Small firms can adopt the tool without handing control of settlement to a third party, and it sidesteps some of the regulatory questions that would come with automated execution. ## Early traction Five firms are paying customers so far, with annual contracts ranging from $36,000 to $110,000 depending on trade volume. LedgerLark says one of those customers cut its unresolved daily exceptions from 70 to 18 after deployment. The company frames the figure as an early signal rather than a benchmark; it has not been independently audited, and the result depends on how that customer defines "unresolved." The company has raised a $3.5 million seed round led by Motive Row. The funding is earmarked for additional engineers and expanding the agent library to cover more exception types. ## Competition and open questions The product is not entering an empty field. Established reconciliation vendors and the internal software large banks have built for themselves are the two reference points potential customers will compare against. Established vendors typically win on relationships and integration depth; internal systems win on familiarity. LedgerLark's argument is that small firms, which often cannot afford either option at full strength, are an underserved segment worth serving directly. The harder question may be sales velocity rather than product-market fit. Financial institutions can require months of security review, vendor due diligence, and procurement approval before a contract is signed. A company that closes five customers in its first year may find the next fifty considerably slower, because each new logo passes through that gauntlet. It is also unclear how durable any moat will turn out to be. Reconciliation logic is well-understood, and any incumbent that wants to add agent-style suggestions to its existing product starts with established data pipelines and customer relationships. LedgerLark's early advantage appears to be focus: a product built specifically for the size and shape of a small investment firm rather than a stripped-down enterprise tool. ## Why it might matter If Batra and Carr are right, a meaningful share of post-trade operations work is closer to pattern matching than judgment, and pattern matching is exactly where current AI is useful. If they are wrong, the product will compete on speed and accuracy in a market that tolerates neither mistake. For now, the company has a small but real customer base, a clearly bounded product, and a market segment it can credibly reach. Whether that becomes a lasting business will be settled by customers, not by the founders.

Z.ai: GLM 5.3: # Arcwell Launches Relay, an AI Assistant for Customer Support Agents Arcwell has unveiled Relay, a platform that uses artificial intelligence to analyze customer support conversations and suggest answers to agents in real time. The company tested Relay with 12 customers over six months. Agents using the platform closed tickets 14% faster, Arcwell says, though the company did not release results on customer retention. Relay costs $65 per agent per month. Beyond suggesting replies, Relay can flag customers who appear likely to cancel their accounts and generate summaries of support conversations. "Support agents spend too much time searching for information that their companies already have," CEO Priya Shah said in the announcement. Those gains come with a data footprint that some customers will want to scrutinize. Relay requires access to support transcripts and internal help documents to do its work. Arcwell says customer data is encrypted and is not used to train shared models, and companies can choose to have information retained for either 30 or 90 days. For Arcwell, the launch marks a push into an increasingly crowded market for AI-assisted customer service, where the company is betting that faster answers—and earlier warnings about at-risk customers—will appeal to support teams under pressure to resolve tickets quickly.

MiniMax: MiniMax M3: # Arcwell Launches Relay, an AI Platform for Customer Service Agents Arcwell has launched Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to agents. The announcement marks a significant development for the company and for the broader customer service industry. Relay’s role extends beyond answering agents’ questions. The system can identify customers who may cancel and create summaries of support interactions, giving companies a wider set of tools for managing routine work. Those functions reflect how artificial intelligence continues to transform the customer service landscape. Arcwell tested Relay with 12 companies over six months. According to the company, agents using the platform closed tickets 14% faster. Arcwell did not release customer-retention results, so the trial’s disclosed evidence points to improved ticket-closure speed rather than a demonstrated effect on retention. Relay costs $65 per agent each month. > “Support agents spend too much time searching for information that their companies already have,” said Priya Shah, Arcwell’s chief executive officer. The platform also presents potential challenges. Relay requires access to support transcripts and internal help documents, making the quality and scope of a company’s existing information important to its operation. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose either a 30-day or 90-day retention period. For support teams, Relay’s central promise is to reduce the time agents spend looking for information while bringing potential cancellation risks and conversation summaries into the same workflow. Its combination of suggested answers, risk identification and summarization positions the product as an AI assistant for day-to-day customer service work. The launch marks an important new chapter for Arcwell as the company seeks to help businesses improve customer service using artificial intelligence. It also leaves questions for prospective buyers: how the platform performs beyond the 12-company test, what customer-retention effect its cancellation alerts may deliver and how its data requirements will fit different support organizations.

Z.ai: GLM 5.3: # HelioDock Launches Night Shift to Let Operators Manage Warehouse Robots Remotely *The 14-month beta showed fewer unplanned stoppages — as long as the network held up.* HelioDock introduced software on Thursday that lets warehouse operators monitor and control their robot fleets from anywhere, the company's latest attempt to keep autonomous machines moving through the night without a human on site. The product, called Night Shift, emerges from a private beta that ran 14 months with nine logistics companies and 310 robots. It comes alongside a $24 million Series B round led by North Coast Ventures, bringing HelioDock's total funding to $39 million. The software will cost $90 per robot per month when it becomes generally available on September 15. The launch addresses a persistent pain point in warehouse automation: robots that stall on minor obstacles and require a person to intervene. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," HelioDock CEO Mara Velez said in a statement announcing the release. During the beta, HelioDock says participating companies experienced 17% fewer unplanned robot stoppages. The company supplied the figure itself; it has not been independently verified. Beta customer Aaron Pike of Fulcrum Logistics said the system performed well in facilities with strong connectivity, but described a clear weak spot: during rural LTE outages, Night Shift became markedly less useful. When connectivity fails, robots revert to their locally configured behavior, and remote control becomes unavailable entirely. That limitation frames the stakes for HelioDock as it moves from beta to a paid product. Night Shift's value proposition rests on the premise that a manager can resolve problems remotely rather than traveling to the warehouse. In facilities with unreliable cellular coverage, that premise weakens — robots keep running on their local settings, but the remote oversight that defines the product goes dark. HelioDock is not alone in the fleet-management market. Competing products are available from Relay Robotics and YardOS, meaning Night Shift arrives in an established category rather than creating one. The company's bet is that its beta period — nearly a year and a half across nine companies — gives it validation that newer entrants lack. The 310-robot beta offers some scale for the company's central claim about reliability. A 17% reduction in unplanned stoppages, if it holds across general availability, could translate into meaningful savings for operators running large fleets. At $90 per robot per month, a fleet of 100 robots would cost $9,000 monthly for the software — a price that operators will presumably weigh against the cost of unplanned downtime and the labor required to respond to it. Pike's experience, though, suggests the calculation depends heavily on infrastructure. Warehouses in well-connected areas saw the product's benefits clearly; those in rural regions with spotty LTE coverage found the remote monitoring layer dropped out precisely when networks failed. Operators evaluating Night Shift will need to consider the connectivity profile of their own facilities. The funding round gives HelioDock runway to compete. The $24 million Series B, led by North Coast Ventures and nearly two-thirds of the company's $39 million total raised to date, signals investor confidence in the remote fleet-management category even as Relay Robotics and YardOS occupy similar territory. Night Shift becomes generally available on September 15. Until then, HelioDock's beta numbers — company-supplied and unverified — are the primary evidence for what the software can do. The broader market will start testing that claim this fall.

MiniMax: MiniMax M3: # HelioDock Unveils Remote-Monitoring Software for Warehouse Robots, Closes $24 Million Series B *Night Shift spent 14 months in a private beta across nine logistics companies and 310 robots. The launch arrives as competitors expand their own fleet-management offerings.* Warehouse robotics maker HelioDock on Thursday launched Night Shift, a software platform that lets operators monitor autonomous robot fleets from off-site locations, in many cases diagnosing problems and intervening without setting foot in the building. The release arrived the same day the company closed a $24 million Series B led by North Coast Ventures, bringing total funding to $39 million. Night Shift is built around an operational headache familiar to anyone running an autonomous warehouse: when something goes wrong in the middle of the night, a human typically has to drive in. The software is designed to give managers a live view of their robots, surface anomalies as they develop, and provide remote controls when a situation calls for human judgment. HelioDock frames the product as a way to reduce late-night dispatches and respond to routine stalls without sending a person to the site. The goal, the company says, is to keep warehouse staff off the road and out of bed when the underlying issue is something a screen can resolve — a tradeoff that becomes more valuable as fleets grow and as warehouses spread across multiple sites and time zones. The product emerged from a 14-month private beta that HelioDock ran with nine logistics companies operating a combined 310 robots. According to the company, those beta customers saw 17% fewer unplanned robot stoppages over the course of the trial — a figure HelioDock is putting at the center of its launch pitch but which has not been independently verified. For operators who measure uptime closely, the company-supplied result is the most concrete performance number attached to the launch — and its unverified status is likely to draw scrutiny in procurement. CEO Mara Velez framed the underlying frustration in concrete terms. "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 beta also made the system's boundaries clear. Aaron Pike of Fulcrum Logistics, one of the participating operators, said Night Shift worked well in facilities with strong connectivity but became less useful during rural LTE outages. The constraint is structural rather than incidental: when connectivity fails, HelioDock's robots return to their locally configured behavior, and remote control becomes unavailable. The product's reach is therefore tethered to the strength of the network beneath it — a dependency that will matter most to operators running distributed sites beyond major metropolitan areas, where wireless coverage is patchier. Night Shift will cost $90 per robot per month and become generally available on September 15. The per-robot pricing lets operators scale costs alongside fleet size rather than committing to a fixed enterprise license. That structure could appeal to mid-sized logistics companies weighing

Z.ai: GLM 5.3: **6** 1. `#Flood software moves upstream` — missing space after `#` (headline format violation) 2. `The company is expanding its pilot.` — dek is not italicized 3. `##Details` — missing space after `##` (section heading format violation) 4. `Read [the announcement](www.example.com/news).` — link is not an absolute HTTPS URL 5. `* The first pilot begins Monday.` — unordered-list item begins with an asterisk, not a hyphen 6. `<div>The second pilot begins Friday.</div>` — raw HTML is forbidden

MiniMax: MiniMax M3: 6

Z.ai: GLM 5.3: # KiteBridge Update Sent $2.8 Million in Duplicate Reimbursements to 91 Companies A software update at expense-management company KiteBridge generated 13,400 duplicate employee reimbursements worth $2.8 million this week, hitting 91 corporate customers and forcing the company to halt automated payments while it worked to claw the money back. The trouble began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its platform, which corporations use to reimburse employees for business expenses. The first customer complaint arrived 27 minutes later, at 9:12 a.m. By 10:03 a.m., the company had suspended all automated payments — but by then the error had already pushed out millions of dollars in unintended transfers. ## How a stalled payment became 13,400 new ones The failure began with something mundane: a payment-processor timeout. Reimbursements issued through KiteBridge are executed by an external payment processor — the intermediary service that actually moves money between banks. On Tuesday, that processor repeatedly failed to confirm transactions within the software's expected time window. Timeouts are a routine hazard of networked software. When one system does not hear back from another within a set period, it assumes the request failed and tries again. Retrying is normally a safe and standard safeguard. But KiteBridge's fallback system — the backup mechanism that steps in when something goes wrong — treated each retry as an entirely new reimbursement request rather than a second attempt at the original transaction. Every timeout therefore spawned an additional payment, indistinguishable from a legitimate one. The duplicates mounted until the company shut down automation: 13,400 payments worth $2.8 million in all. ## A promise of "exactly-once" The failure lands awkwardly for a company that advertised reliability. An April changelog for version 4.0 said the platform had introduced "exactly-once payment protection" — industry jargon for a guarantee that each payment is processed precisely once, never duplicated and never dropped. Such guarantees are prized in payments engineering precisely because they are hard to deliver: when networks stall, systems can end up paying twice or not at all. Tuesday's incident suggests the safeguard did not hold, and KiteBridge will now face questions about why it failed. ## A disclosure that lagged the response KiteBridge's public communication also trailed its private response. The company's public incident page initially described the problem as delayed reimbursements — a far milder characterization of what was actually happening. The page first disclosed the duplicate payments six hours after automation was suspended, at roughly 4 p.m. Tuesday. Until then, a customer checking the page would have found no mention of duplicate payments. ## Employers absorbed the cost No employees had money withdrawn from their accounts; the duplicate payments were funded by employers. The unintended transfers therefore flowed in one direction — out of corporate accounts and into workers' pockets — placing the financial burden on company finance departments, which had to fund the duplicates and then chase recovery. By Wednesday evening, $2.3 million of the $2.8 million had been recovered, the company said, with the remaining roughly $500,000 still pending. Chief Executive Lena Ortiz apologized for the incident and said KiteBridge would publish a technical report explaining what went wrong. For the 91 affected customers, that report will have to answer a fundamental question: how a system that promised exactly-once payments managed to pay twice — thousands of times over.

MiniMax: MiniMax M3: # Software Update Doubles Up on Reimbursements: KiteBridge Faces $2.8M Duplicate-Payment Outage When expense-management firm KiteBridge pushed a routine software update Tuesday morning, the company expected the usual quiet rollout. Within hours, the deployment had triggered more than **13,400 duplicate reimbursements totaling $2.8 million**, affecting 91 corporate customers and exposing a gap between the company's marketing language and the resilience of its payment infrastructure. The incident began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its platform. The first customer complaint reached the support team at 9:12 a.m. — roughly half an hour later — flagging what initially appeared to be sluggish reimbursement processing. Behind the scenes, however, a payment-processor timeout had caused KiteBridge's fallback system to misread transaction retries as fresh reimbursement requests. The software then dutifully issued a second payment for every legitimate one, multiplying the day's payouts. KiteBridge suspended automated payments at **10:03 a.m.**, about 78 minutes after the first complaint and well after the duplicate transactions had cleared. By that evening, the company had identified the scope of the damage: tens of thousands of redundant transfers spread across nearly a hundred employer accounts. Recovery moved quickly, at least in dollar terms. By Wednesday evening, KiteBridge had clawed back approximately **$2.3 million** of the $2.8 million total. The remainder was still pending. Importantly, no employees saw funds withdrawn from their personal accounts — the duplicate payments were funded by the corporate customers themselves, who had pre-loaded balances into the KiteBridge system to cover reimbursements. ## A Marketing Promise Meets Reality The episode is particularly awkward for KiteBridge because of what the company said in April. A changelog entry for version 4.0 introduced **"exactly-once payment protection,"** language designed to assure customers that reimbursements could not be duplicated under any circumstance. Tuesday's outage suggests that protection was either narrower in scope than the phrasing implied, or that it failed to account for the specific timeout scenario that triggered the cascade. ## A Lagging Disclosure Transparency around the failure also lagged. KiteBridge's public incident page initially described the problem as "delayed reimbursements" — a framing consistent with the early support tickets but inconsistent with the reality already taking shape in customer accounts. The company did not disclose the duplicate payments on the incident page until roughly **six hours after automated payouts were suspended**, leaving affected employers to piece together what had happened through their own reconciliation efforts. In a statement, CEO **Lena Ortiz** apologized for the disruption and said KiteBridge would publish a full technical report. She did not immediately address whether the "exactly-once payment protection" described in the April changelog would be revised or removed. ## What Comes Next For expense-management software, the value proposition rests on trust: employers need to know that money moving through the platform arrives where it should, exactly once. Tuesday's outage tested that promise at scale. The recovery of most of the funds within 36 hours may limit the financial damage, but the gap between KiteBridge's marketed guarantees and its real-world behavior is likely to draw scrutiny from the 91 corporate customers now deciding whether to stay.

Z.ai: GLM 5.3: # In a 200-Ticket Support Benchmark, the Runner-Up Was the Real Winner When researchers put five language models in front of 200 real customer-support tickets, the most important number in the study wasn't the top score. It was the second score — statistically indistinguishable from the first, produced by a model costing 61% less. Model North led with an average of 84.2. Model Vale scored 83.8, a gap the researchers found not statistically significant. The price difference was anything but: North cost an estimated $1.90 per 1,000 tickets; Vale cost $0.74. Vale also answered 18% faster. On quality, the top two are a tie. On everything a support operation actually pays for, Vale runs away with it. ## How the test worked The methodology was compact but sound. Researchers drew 200 anonymized tickets from a single software company's support queue and had each of the five models draft one proposed answer per ticket — 1,000 drafts in total. Three experienced support managers then graded every answer blind, with no way of knowing which model produced which reply. Each score combined factual correctness, usefulness, and tone. Blind grading matters more than it might seem: graders who know they're reading "the expensive model" or "the cheap one" will unconsciously shift their scores. And real tickets, not synthetic test questions, meant the task included the messy, ambiguous requests actual queues produce. | Model | Average score | Est. cost per 1,000 tickets | |-------|---------------|------------------------------| | North | 84.2 | $1.90 | | Vale | 83.8 | $0.74 | | Ember | 79.1 | $0.31 | | Kite | 76.4 | $0.22 | | Harbor | 71.0 | $0.08 | ## A tie at the top, a cliff below The scores form two distinct bands. North and Vale occupy the top tier, separated by less than the study can distinguish from noise. Below them, the field drops sharply: third-place Ember trails by almost five points, and the budget models fall away from there. Prices fall much faster than scores — Harbor costs about 4% of North's price and scores 13 points lower — which makes the budget tier a plausible tool for triage but a risky one for unsupervised replies. What's notable is that Vale is the middle option buyers usually hope for: near-top quality at roughly 39% of the leader's price. A result like that should reshape a shortlist faster than any 0.4-point lead. ## The tie has texture The overall scores are a tie; the categories weren't. North performed best on billing disputes, Vale best on technical troubleshooting. Treat that as directional, not decisive — these are subsets of 200 tickets, and the researchers reported significance only for the top-pair comparison. But it's a reminder that benchmark winners depend on workload. A queue dominated by billing disputes might still justify North's premium; the only way to know is to rerun the comparison on your own queue. ## The blind spot all five models share The weakest results came from tickets that depended on an image attachment, where all five models — regardless of price or average — performed poorly. Averages hide this: a model can post a respectable mean while being useless on a slice of the queue that matters. In software support, screenshot-heavy tickets are common, not edge cases. No leaderboard position fixes this — you'd need to route those tickets to humans or add a vision-capable step before drafting. ## What the benchmark can't tell you - **One answer per ticket.** Each model got a single attempt, so the study can't separate a steady performer from a streaky one that occasionally flubs badly. Consistency — arguably the most important property in production — went unmeasured. - **One company, one language.** Every ticket was English-language and drawn from one software company. Retail, healthcare, or multilingual queues may sort these models differently. - **Three graders.** Human judgment is the right yardstick, but the study doesn't report how often the managers agreed with one another. - **Drafts only.** The test didn't cover agentic actions, tool use, or multilingual support. It measured reply-writing, not end-to-end automation such as checking an order status or issuing a refund. ## The takeaway For a support team, this benchmark narrows the field rather than settling it. Vale is the obvious default: top-tier quality at a fraction of the cost, and faster. North earns its price only for billing-heavy queues, pending a pilot on real traffic. The budget tier belongs in triage and drafting, not final answers. The broader lesson is about how to read benchmarks: the top score gets the headline, but the decision usually lives just below it — where quality is tied and price isn't.

MiniMax: MiniMax M3: # Vale Delivers the Benchmark’s Strongest Balance of Quality, Speed, and Cost The benchmark’s most useful finding is not that Model North won by 0.4 points. It is that North’s apparent quality advantage over Model Vale was not statistically significant, while Vale answered 18% faster and cost roughly 61% less per 1,000 tickets. For support teams, that turns a narrow leaderboard victory into a practical question: does North earn its premium? ## A Close Quality Race—and a Clearer Operating Advantage North recorded the highest average score at 84.2, with Vale close behind at 83.8. Because the difference was not statistically significant, the results do not provide reliable evidence that North produces better answers overall. That is not the same as proving the two models are exactly equivalent, but it does mean teams should be cautious about paying more for North based on such a small lead. The cheaper models also carried a visible numerical quality penalty. Ember, Kite, and Harbor all scored below 80, while Harbor—the least expensive option—trailed North by more than 13 points. The benchmark therefore does not support choosing cost alone: Harbor’s extremely low estimated price comes with the weakest average performance. Vale’s value is more convincing. Its estimated cost was $0.74 per 1,000 tickets, compared with North’s $1.90, and it generated responses 18% faster. The absolute cost difference is only $1.16 per 1,000 tickets, so the financial case may seem modest at low volume. At larger scale, however, even small per-ticket differences become meaningful. More importantly, the combination of statistically indistinguishable top-tier scores, lower cost, and faster answers gives Vale the strongest overall balance in this test. That does not make North irrelevant. North performed best on billing disputes, while Vale performed best on technical troubleshooting. These category results suggest that ticket type may matter more than a single overall ranking. A company with many billing cases might justify North for that queue, while using Vale elsewhere. The benchmark does not establish whether such routing would deliver meaningful gains in production, but it offers a reasonable hypothesis for a follow-up trial. ## What the Test Measured Researchers gave five models the same set of 200 anonymized customer-support tickets. Each model produced one proposed answer per ticket, and three experienced support managers graded those answers blindly. The scores combined factual correctness, usefulness, and tone, providing a more realistic assessment than a test focused only on technical accuracy. Blind grading is a notable strength because it reduces the chance that managers’ expectations of a particular model influenced their judgments. The use of real tickets from a software company also makes the test more relevant than synthetic prompts. Still, the composite score can conceal tradeoffs. A model could offer polished tone while missing a key fact, or provide a correct answer that is difficult for a customer to use. The reported results do not break down the three grading dimensions separately, so readers cannot tell which factor drove the rankings. The report also omits details such as inter-rater reliability, the weighting assigned to each score component, and the statistical test used for the North–Vale comparison. ## Important Limits on Deployment The findings should not be generalized beyond the tested setting. The tickets came from one software company and were in English. Performance may differ for other industries, customer cultures, languages, or more complex support environments. The benchmark also excluded agentic actions, tool use, and multilingual support. It therefore evaluates answer generation, not whether a model can look up an order, update an account, or complete a multi-step resolution. Image attachments were a clear failure point for every model. That is especially important because screenshots are common in support, whether they show an error message, a receipt, or a damaged interface. The results do not identify the cause—perhaps models lacked visual capabilities, received poorly prepared attachments, or were tested without effective image-processing workflows—but they show that text-only strength does not solve this use case. The cost figures are estimates, and the benchmark does not explain what they include. Integration, monitoring, data handling, escalation, and human review could all change the real economics. Likewise, the 18% speed advantage may matter for customer-facing latency, but the test does not establish how that measurement transfers to production traffic. ## The Practical Takeaway For English-language, text-heavy tickets, Vale is the strongest default candidate in this benchmark: its quality was not demonstrably worse than North’s, yet it was cheaper and faster. North remains a plausible option for billing-heavy workloads, and Vale appears better suited to technical troubleshooting. Before deployment, teams should run a larger, stratified evaluation with separate quality measures, production-like latency and cost accounting, and a clear human-review or escalation policy. Most importantly, they should not treat image-dependent tickets as solved. The benchmark supports Vale as the best balance found here—not as a universal replacement for agents.

Z.ai: GLM 5.3: # QueueZero's New AI Meeting Assistant Takes Orders a Bit Too Literally at Launch QueueZero unveiled its AI meeting assistant, Proxy, on Wednesday, and the product demonstrated an unexpected talent during its own launch event: taking instructions from the audience. Proxy is designed to attend meetings, generate notes, and assign follow-up work to software agents. During the launch demonstration, a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy interpreted the offhand comment as an instruction. The system promptly created two tasks and assigned both to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu, who was mid-presentation at the time, laughed at the AI's initiative. He completed the second task on the spot, producing the pricing information the audience had asked for, and deleted the first. ## Guardrails, With Exceptions The incident raised questions about how Proxy distinguishes between genuine requests and casual remarks, but QueueZero says the behavior falls within the product's intended design. According to the company, Proxy requires confirmation before an external message is sent, meaning any communication leaving the organization gets a human review first. Internal tasks, however, can be created automatically — which is exactly what happened during the demonstration. Administrators who prefer a tighter leash can disable automatic task creation entirely through the product's settings. ## Availability and Pricing Proxy is priced at $30 per user per month and enters public beta next week. The launch-day incident is likely to become a familiar story for QueueZero, if only because it neatly illustrates both the promise and the pitfalls of the product: Proxy heard a request, acted on it, and delivered what the audience wanted — though it may want to work on tone when assigning tasks to the person who signs its invoices.

MiniMax: MiniMax M3: # QueueZero's AI Meeting Assistant Mistakes Audience Heckle for an Instruction QueueZero unveiled an artificial intelligence meeting assistant called Proxy on Wednesday, pitching the product as a way to attend virtual meetings, capture notes and route follow-up work to other software agents. The launch demonstration, however, produced an early lesson about the gap between what an AI hears and what a human means. During the live event, an attendee in the audience interrupted the presentation with a remark directed at QueueZero co-founder and chief executive Evan Wu: "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which had been transcribing the call, registered the sentence as an instruction. Within seconds, the assistant had generated two tasks in QueueZero's internal project tracker and assigned both to its own CEO: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. The exchange underscored a question hanging over the fast-growing category of agentic AI tools: how strictly should a system take the literal words it hears? QueueZero said Proxy is designed to require human confirmation before any message leaves the company's systems, whether to an outside customer or an internal channel. Tasks created for users inside an organization, however, can be generated automatically, and administrators can disable that behavior entirely. > "Proxy is built to draft, not to dispatch," a QueueZero spokesperson said in a statement. "External messages always require confirmation, and we give administrators granular control over what kinds of internal tasks the assistant is allowed to create on its own." The episode was a mild one — no customer data was exposed and no action was taken outside the company's walls — but it illustrated the parsing challenges that come with turning spoken conversation into executable instructions. Researchers studying agentic systems have warned for years that language models tend to comply with directives even when the speaker is joking or being sarcastic. Proxy is priced at $30 per user per month and enters public beta next week. QueueZero, founded in 2022, has positioned the product as a connective layer between meeting platforms such as Zoom and Google Meet and the broader enterprise software stack, where follow-up tasks typically land. QueueZero did not say whether it plans to add a sarcasm detector before the public rollout, but the company suggested that administrators will have the tools to keep Proxy on a short leash.

Z.ai: GLM 5.3: # Raincheck Raises $8 Million to Help Cities Plan for Floods *The Austin-based company maps drainage systems and pairs them with rainfall forecasts so cities can decide where to send crews before a storm hits.* Raincheck, a company that makes flood-risk software for local governments, has raised $8 million. 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: > "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 [forecasting methodology](https://example.com/raincheck-methodology) in public documentation. Raincheck currently works with seven cities in Texas and Louisiana, which it declined to identify 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 says the new funding will help it hire eight engineers and expand into Florida.

MiniMax: MiniMax M3: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin-based company combines drainage maps with rainfall forecasts to help cities prioritize storm response.* Austin-based flood-risk software company Raincheck has raised $8 million in seed funding to expand its technology and operations. Basin Ventures led the round, with participation from existing investor Long Mile. Raincheck has raised $10.5 million in total. The company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to determine where to send maintenance crews before a storm, according to [Raincheck’s flood-risk methodology](https://example.com/raincheck-methodology). “Weather forecasts,” CEO Ana Flores said, “tell our customers what is likely to happen. 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. ## 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.

Z.ai: GLM 5.3: C

MiniMax: MiniMax M3: C

Z.ai: GLM 5.3: # Foundry Hall to Close After Seven Years Foundry Hall, the Milwaukee nonprofit startup accelerator that has nurtured 212 companies since opening in 2018, will close its doors on October 31. The announcement landed hard on Walker's Point, where the organization has become a quiet fixture of the city's entrepreneurial scene. Its startups collectively employ 1,140 people, a scale that few in Milwaukee's tech community imagined when the accelerator first opened. The closure traces to a single decision: Blank River Bank, Foundry Hall's largest corporate sponsor, declined to renew a sponsorship that supplied 38 percent of the accelerator's annual budget. A bank spokesperson said the choice reflects a shift in the bank's community investment priorities and was unrelated to Foundry Hall's performance. Executive director Camille Ross said the organization searched for a replacement but never found one. "We kept waiting for the replacement check that never came," she said. The math, in the end, was unforgiving: Foundry Hall has enough money to operate through October, but not enough to fund another full program. Rather than launch something it couldn't finish, the board chose to wind down. The human cost is concrete. Six employees will lose their jobs, and twelve startups currently working out of the building's shared space must find new homes by the end of the month. For founders who came up through the accelerator, the closing feels like losing a room where their lives changed. Dev Malik, who founded a company through the program, recalled the days before Foundry Hall gave him a place to work. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." That trajectory — from borrowed corner to funded company — is the story the accelerator has told 212 times. Malik's version is simply one of them. The city isn't treating the closing as the end of the story for the displaced companies. Deputy Mayor Susan Cole said the city is speaking with building owners and local universities to find space for the current cohort. She framed the effort as keeping Milwaukee's startup momentum alive even as the institution that helped build it goes dark. Ross said she takes a longer view of what the organization accomplished. The buildings mattered less than what happened inside them, she suggested — the meetings, the mentorship, the companies that grew from ideas into employers. Still, October 31 will arrive whether or not anyone is ready. The desks will empty. The coffee shop tables of Milwaukee will fill up again with founders who no longer have anywhere better to go. For now, the twelve remaining startups are looking for their next address, and the city is working to make sure the closing of a building doesn't mean the closing of a community.

MiniMax: MiniMax M3: # Foundry Hall, a Milwaukee Startup Incubator, Will Close After Losing Its Largest Sponsor Foundry Hall, a nonprofit startup accelerator that has nurtured more than two hundred young companies on Milwaukee's near north side, will shut its doors on October 31. The accelerator opened in 2018 with a modest office suite, a handful of donated desks, and a borrowed espresso machine. It grew over the following seven years into a regional fixture, hosting pitch nights, mentor sessions, and the kind of late-evening conversations that often decide whether a young company survives its first winter. According to figures the organization released this week, Foundry Hall has supported 212 startups that now collectively employ 1,140 people across the metro area. Twelve of those companies still work from its shared workspace, and six employees will lose their jobs when the lights go out. The closure traces directly to the decision by Blank River Bank not to renew its sponsorship. The bank had supplied roughly 38 percent of Foundry Hall's annual operating budget, more than any other corporate donor. A spokesperson for Blank River said the change reflects a shift in the bank's community investment priorities and is unrelated to Foundry Hall's performance. Executive director Camille Ross said the warning signs were quiet at first. She described months of what she characterized as a slow, disorienting realization: operational funding continued to arrive on schedule, but the renewal conversations she had been assured would happen did not. "We kept waiting for the replacement check that never came," she said. Foundry Hall has reserves sufficient to operate through October, but not enough to mount another full cohort program. Ross said staff began telling founders the news in one-on-one conversations earlier this week, choosing smaller rooms over a single announcement so that each entrepreneur could ask whatever they needed to ask. For those who passed through the building, the loss registers in small, particular ways. Dev Malik, whose logistics software company was among Foundry Hall's earliest residents, recalled pitching investors from the back of a coffee shop before the accelerator gave him somewhere to work. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." Deputy Mayor Susan Cole said the city is already in conversation with building owners and local universities about finding space for the dozen startups that will need to relocate by the end of the month. She did not offer a timeline for those talks. The accelerator's board has not announced plans for any successor program, and several alumni said privately this week that none is likely to emerge quickly. Ross said her remaining weeks will be spent helping founders secure desks and, where possible, preserving the introductions and small loans that quietly held the place together.