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

MiniMax: MiniMax M3: # Raincheck raises $8 million to expand flood-risk software for cities *An Austin startup that pairs drainage maps with rainfall forecasts to help local governments dispatch maintenance crews before storms has closed a seed round led by Basin Ventures.* Raincheck, an Austin company that makes flood-risk software for local governments, today announced an $8 million seed round. The startup maps municipal drainage systems and combines those maps with rainfall forecasts, allowing cities to decide where to send maintenance crews before a storm arrives. The round was led by Basin Ventures, with existing investor Long Mile also participating. Raincheck has raised $10.5 million total. > "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." > > — Ana Flores, CEO The forecasts are recommendations, not directives; emergency managers retain control over deployments and public warnings. Raincheck's approach is detailed in its [forecasting methodology](https://example.com/raincheck-methodology). The company currently works with seven cities in Texas and Louisiana. It declined to identify them because some contracts have not been publicly announced. 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 Prime: # Raincheck raises $8 million to help cities plan for floods *The Austin startup combines drainage maps with rainfall forecasts so local governments know where to send crews before a storm arrives. Basin Ventures led the round.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding, the company announced. The company maps drainage systems and combines those maps with rainfall forecasts, an approach it details in [its methodology documentation](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews before a storm. Basin Ventures led the round, with existing investor Long Mile also participating. 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. Raincheck is not alone in the flood-planning market: other vendors include HydroMap and municipal engineering contractors. The company says the new funding will help it hire eight engineers and expand into Florida.

MiniMax: MiniMax M3: # LedgerLark Tackles Post-Trade Reconciliation With AI Agents The post-trade back office of an investment firm is rarely where software companies look for glamour. It is also, two former operations builders argue, where small firms lose hours every day to manual reconciliation work. LedgerLark, founded by Mina Batra and Owen Carr, sells AI agents that compare trade confirmations, custodian records and internal ledgers, then propose corrections for human review. The idea traces to a moment Batra watched from inside a regional brokerage, where a six-person team spent three days untangling a trade mismatch caused by a copied account number. The episode sharpened a frustration she had carried through seven years in operations: small exceptions pile up, and small teams have no good way to clear them. Carr, who spent years building reconciliation systems at a large bank, saw the same problem from the infrastructure side. Together, the pair argue that post-trade operations at smaller firms have been under-served because they sit beneath the radar of large vendors and the budgets of major banks. LedgerLark's agents read incoming confirmations and custodian statements, line them up against the firm's own ledger, and flag mismatches. Where the cause looks clear, the agent drafts a proposed fix and queues it for a human approver. The system is deliberately scoped: it does not execute trades, route payments, or move money. That boundary, the founders say, is part of how they sell into compliance-sensitive buyers. The company reports five paying customers and annual contracts ranging from roughly $36,000 to $110,000. One client, LedgerLark says, reduced its unresolved daily exceptions from about 70 to 18 after deploying the agents, though the figure has not been independently verified. Five customers is a thin base, and the upper end of the contract range hints at the kind of feature expansion that bigger reconciliation suites already offer. That is the competitive landscape LedgerLark walks into. Established reconciliation vendors have spent years selling into banks and asset managers, and large financial institutions often run their own internally built systems for the same tasks. LedgerLark's pitch is that those tools were designed for firms with dedicated IT staff and complex asset coverage, and that small investment shops are left to stitch together spreadsheets and email threads. Whether that gap is large enough — and durable enough — to support a venture-backed company remains an open question. Financial institutions also tend to require lengthy security and procurement reviews, which can stretch sales cycles. In May, LedgerLark closed a $3.5 million seed round led by Motive Row, with participation from several angel investors. The capital gives the team runway to hire engineers and customer-facing staff, and to broaden the set of exceptions the agents can resolve. The founders describe the next phase as moving from one-off fixes toward pattern recognition across a customer's full transaction history, so that the same class of error is harder to repeat. The bigger question is whether AI agents represent a real shift in back-office tooling, or an incremental improvement on a problem the industry has lived with for decades. Early customer results are encouraging, but a five-customer base is a starting point rather than a verdict. For now, LedgerLark is selling against entrenched incumbents, regulated buyers, and the slow cadence of financial procurement. The founders' bet is that small firms will pay for relief from a kind of work nobody enjoys — and that the bet will hold long enough to scale.

Z.ai: GLM 5.3 Prime: # LedgerLark's AI Agents Take On the Trade Breaks Nobody Wants *The seed-stage startup's agents compare confirmations, custodian records and internal ledgers, then propose corrections for humans to approve.* The origin story is as unglamorous as the problem it targets. Mina Batra spent seven years in operations at a regional brokerage, where post-trade work mostly meant reconciling things that should match and don't. She began exploring the idea after watching a six-person team spend three days resolving a trade mismatch caused by a copied account number. One wrong keystroke; three days; six salaries. That friction rarely makes headlines, but it is the daily texture of operations at small investment firms — too small for the reconciliation platforms large banks take for granted, too regulated to let errors slide. LedgerLark, which Batra co-founded with Owen Carr, is betting that AI agents have finally made that gap worth closing. ## The insight: pattern-matching in a spreadsheet world Carr built reconciliation systems at a large bank, so the two came at the problem from opposite ends of the market. Their shared read: resolving a mismatch is less accounting than pattern-matching — compare confirmations against custodian records against internal ledgers, find where they diverge, work out which source is wrong. That work is document-heavy, repetitive and context-dependent, which on paper suits AI agents well. Large institutions addressed it decades ago with software and the staff to run it. Small firms typically address it with spreadsheets, email chains and overtime. That asymmetry is LedgerLark's wedge. ## Propose, don't touch LedgerLark's agents sit across the three record sources. When they find a mismatch, they draft a proposed correction for a human to review and approve. Two boundaries define the product: the agents neither execute trades nor move money. That line does a lot of work. It shrinks the blast radius of a mistake — a bad suggestion can simply be rejected — and it should ease the security reviews financial institutions require of vendors. It also caps how much labor the system removes. The wager is that drafting the correction, not approving it, is where the hours go. If approval is the real bottleneck, the value narrows. ## Early traction, with caveats Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The company says one of them cut unresolved daily exceptions from 70 to 18 — a company-reported figure from a single customer, without detail on baseline or timeframe. It is the kind of number that invites follow-up rather than conclusion: does it hold across the other four customers, and at higher volumes? Another open question is how often humans accept the agents' proposals. A high acceptance rate is the product working, unless it reflects rubber-stamping — a known failure mode of human-in-the-loop systems. Whether the approval step stays meaningful as the tool becomes routine is the thing to watch. ## A crowded field and slow doors Nor is the field empty. Established reconciliation vendors have sold into financial institutions for decades and are adding AI features of their own; large banks run internally built software. And every firm's cheapest alternative is doing nothing, since post-trade errors are expensive but survivable — which caps what customers will pay to prevent them. Then there is the calendar. Security and procurement reviews at financial institutions can stretch for months, extending sales cycles at precisely the stage a young company can least absorb them. The $3.5 million seed round LedgerLark has raised, led by Motive Row, mostly buys patience to wait those out. LedgerLark's thesis, stated honestly, is modest: small firms hold a lot of matching work that vendors have ignored, and AI agents may have made serving them economical for the first time. Three days to fix a mistyped account number is a real cost, paid by someone, most weeks. Whether an agent can propose the fix faster — and whether the human at the end of the loop keeps checking its work — is what the next year of customers should answer.

MiniMax: MiniMax M3: C

Z.ai: GLM 5.3 Prime: **C** Option C accurately captures all the key facts: the product removal, the price point, the specific scope of the exposure (workspace titles, not message contents), the number of affected accounts, and the duration. Options A and D both misstate the facts—message contents were not exposed, and there's no evidence of malicious access. Option B buries the actual news entirely.

MiniMax: MiniMax M3: # HelioDock launches Night Shift for remote warehouse robot fleets *Remote-monitoring software emerges from a 14-month beta and a $24 million funding round, with general availability set for September 15.* On Thursday, HelioDock, which makes autonomous warehouse robots, launched Night Shift, software that lets operators monitor robot fleets remotely and take control of individual units when they need assistance. The release extends the company’s offering from autonomous machines to the software used to supervise them across facilities. Night Shift is intended to give warehouse teams a centralized view of fleet status and a way to intervene without being physically present at a facility. Its per-robot pricing means the monthly software charge rises directly with the number of machines included in a deployment. The launch also broadens HelioDock’s product mix. Night Shift adds a software layer for managing autonomous robots remotely, with its own monthly charge. The two products address different parts of warehouse operations: the machines working autonomously and the software used to oversee them. HelioDock developed the product through a 14-month private beta involving nine logistics companies and 310 robots. The company said beta customers experienced 17% fewer unplanned robot stoppages. That figure has not been independently verified, leaving the reduction as a company-reported result rather than an externally established benchmark. CEO Mara Velez framed the product around the operational burden of responding to a robot problem: “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet.” Her statement presents remote intervention as an alternative to sending a manager to a facility when an autonomous machine stops. The beta also identified a practical boundary for Night Shift’s usefulness. 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. His assessment is consistent with the product’s dependence on a network connection for its remote capabilities. When connectivity fails, robots return to their locally configured behavior and remote control becomes unavailable. An outage therefore does not automatically bring the robots to a halt; instead, it shifts them back to behavior set locally while removing the ability to monitor or assist them through Night Shift. This gives operators a fallback, but it also makes network coverage a relevant condition for evaluating the software. That limitation sits alongside, rather than directly contradicting, HelioDock’s stoppage claim. Fewer interruptions and reliable remote access are separate performance measures: a site can report fewer unplanned stoppages while still losing centralized control whenever its connection drops. Pike’s experience shows how the usefulness of remote management can vary with the connectivity available at a facility. HelioDock also announced a $24 million Series B led by North Coast Ventures, bringing the company’s total funding to $39 million. The Series B accounts for about 62% of HelioDock’s total funding, while Night Shift adds a recurring software offering to a product line built around autonomous robots. The announcement pairs that commercial release with a new round of financing. Night Shift will enter a market that already includes fleet-management products from Relay Robotics and YardOS. The competing offerings give warehouse operators alternatives to consider as HelioDock moves from a private beta into general availability. The software layer has become a purchasing decision alongside the autonomous robots themselves, with remote tools and network dependence becoming part of that evaluation. Night Shift is scheduled to become generally available on September 15. The listed software price at launch will be $90 per robot per month, with its remote functions still dependent on the network conditions identified during the beta.

Z.ai: GLM 5.3 Prime: #

MiniMax: MiniMax M3: # Arcwell's Relay brings AI-driven answer suggestions to customer support teams Arcwell has launched Relay, a platform that uses artificial intelligence to read customer support conversations and propose replies to the human agents handling them. The software packages three capabilities aimed at the daily grind of a support team: it suggests answers while a conversation is unfolding, flags customers who appear likely to cancel their service, and writes a summary once the ticket closes. The platform was put through a six-month pilot with 12 companies. Agents using Relay closed tickets 14% faster than agents who did not, according to Arcwell. The company did not release customer-retention results from the trial, even though retention is the second of the three problems Relay is designed to address. Relay costs $65 per agent per month. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. The product's value depends on access to material most companies treat as sensitive: support transcripts and internal help documents. Arcwell says the data is encrypted, is not used to train shared models, and can be set to expire after 30 or 90 days at the customer's option. Those assurances will face an early test in procurement. IT and security teams typically require audit trails, contractual data-handling commitments and proof that a vendor's privacy claims hold up under independent review. Arcwell has not disclosed which of the 12 pilot participants have become paying customers, leaving the company with a credibility gap to close before Relay lands in its first regulated industry. A 14% productivity gain in a controlled pilot is a useful starting point. The harder question is whether the gains hold up at scale, in contact centers where conversations are longer, more technical and bound by stricter compliance rules than the average e-commerce complaint — and whether buyers will trust an unfamiliar vendor with the transcripts of every customer interaction.

Z.ai: GLM 5.3 Prime: #

MiniMax: MiniMax M3: **Cloudnote removes unauthenticated debugging page after researcher report** *The note-taking platform says prompt excerpts from 1,842 user sessions were reachable for four hours on Monday, though its investigation is not yet complete.* Cloudnote, a note-taking platform with 3.2 million registered users, took down an internal debugging page on Monday after a security researcher reported it was publicly accessible without authentication. Logs reviewed by the company show the page contained prompt excerpts from 1,842 user sessions and had been reachable for roughly four hours before removal. Each excerpt could include up to 200 characters of user input, according to the company. Cloudnote removed the page 19 minutes after receiving the researcher's notification. Full documents, passwords and payment information were stored in separate systems and were not present on the exposed page, the company said. The episode matters because Cloudnote prompts often contain material users paste or dictate into the service, including draft emails, code snippets and internal memos. Even short excerpts can reveal context, proprietary information or identifying details about individuals or organizations. Cloudnote said it had found no evidence that anyone besides the researcher accessed the page during the four-hour window, but acknowledged its investigation is incomplete. It has not disclosed how the debugging page became publicly accessible or whether affected users will be notified directly.

Z.ai: GLM 5.3 Prime: # Cloudnote Says Debugging Page Exposed Customer Prompt Excerpts for Four Hours *An unauthenticated page displayed excerpts from 1,842 user sessions before Cloudnote removed it 19 minutes after a researcher's warning. The company says no one else accessed it — but its investigation isn't finished.* Cloudnote, an AI note-taking service with 3.2 million registered users, said a debugging page on its platform was accessible without authentication for four hours on Monday, exposing excerpts of what customers typed into the product. The company removed it 19 minutes after an outside researcher notified it. Logs show the page displayed excerpts from 1,842 user sessions, each up to 200 characters of a prompt — roughly a sentence or two. Full documents, passwords, and payment information were stored elsewhere and did not appear on the page. The affected sessions represent less than one-tenth of one percent of Cloudnote's user base. The scale is small, but the stakes are not just numerical. Prompts to an AI note-taking service can summarize meetings or hold private thoughts, and even 200 characters can identify someone or disclose confidential matters. For a product built on users trusting it with their notes, any unauthenticated window into that material is consequential. Cloudnote says it has found no evidence that anyone besides the researcher accessed the page. But the investigation is not complete, and no evidence of access is not the same as confirmation that none occurred. How the page came to be reachable without a login also remains unclear. For now, the confirmed facts are narrow: four hours of exposure, 1,842 sessions, no passwords or payment data. What Cloudnote cannot yet rule out — and what its investigation must answer — is whether anyone else looked.

MiniMax: MiniMax M3: # The Last Month at Foundry Hall Foundry Hall’s closing notice was brief: the nonprofit startup accelerator would close on October 31. Its Milwaukee home, opened in 2018, had become a gathering place for founders trading ideas, revising plans, and waiting for news that might change a company’s future. Now that future had been reduced to a final month. Since opening, Foundry Hall had supported 212 startups, which collectively employed 1,140 people. Those numbers described a reach that extended beyond its offices, but they also named years of work: early ideas tested, businesses formed, and teams given room to grow. The building had offered what a young company often lacks most—a place to work and a network of people willing to ask hard questions. Its finances had been strong enough to sustain its mission, but not independent of it. Blank River Bank, the largest corporate sponsor, supplied 38 percent of Foundry Hall’s annual budget. The bank declined to renew its sponsorship, removing a pillar of support. Foundry Hall had enough money to operate through October, but not through another full program. “We kept waiting for the replacement check that never came,” said Camille Ross, Foundry Hall’s executive director. Her words named the hope deferred: another backer might step forward before the accelerator ran out of time. No replacement arrived. The consequences were already personal. Six employees would lose their jobs. Twelve startups currently using the workspace would have to leave. Some might find desks elsewhere; others faced uncertainty about where to work and what would happen next. Founders who had built routines around the building would soon have to carry those routines somewhere else. For Dev Malik, a founder whose first investor meeting took place at Foundry Hall, the contrast with those uncertain departures was sharp. “Before Foundry Hall, I was pitching from the back table of a coffee shop,” he said. At the accelerator, that early pitch had become part of a larger story—one that included employment, teams, and the possibility of more. Blank River Bank said the decision reflected a change in its community investment priorities and was unrelated to Foundry Hall’s performance. The distinction offered little comfort to the people whose jobs and workspaces were disappearing, but it kept the closure from becoming simply a verdict on the accelerator’s work. Deputy mayor Susan Cole said the city was speaking with building owners and local universities about space for the current companies. Such arrangements might offer the twelve startups a temporary landing place. They could not restore the program that had supported 212 companies, and they could not replace the six jobs due to be lost. By October 31, the doors would close. Until then, the work would continue under a deadline no founder had chosen: one final stretch, carried forward by people trying to preserve as much as possible before the lights went out.

Z.ai: GLM 5.3 Prime: # Foundry Hall to Close October 31 After Losing Its Largest Sponsor *212 startups. 1,140 jobs. One missing check.* Foundry Hall, the nonprofit startup accelerator that has anchored Milwaukee's early-stage scene since 2018, will close on October 31. The closing is not a verdict on the work. The hall has supported 212 startups since it opened, companies that collectively employ 1,140 people today. Its largest corporate sponsor, Blank River Bank, supplied 38 percent of the annual budget and declined to renew this year. A bank spokesperson said the decision reflects a shift in the bank's community investment priorities, not the accelerator's performance. For an organization that did everything it was funded to do, the distinction offers little comfort. The arithmetic that follows is unforgiving. Executive director Camille Ross said the organization held out hope that another sponsor would step into the gap. > "We kept waiting for the replacement check that never came." Foundry Hall can cover October's bills. It cannot cover another full program — the rent, the staff, the mentorship that made the place more than an address. Somewhere between those two numbers, the decision made itself. Six employees will lose their jobs. The twelve startups working out of the building have until the end of the month to find somewhere else to go — a search the staff who kept the place running will be making too. For the people who built companies there, the loss is harder to quantify. Dev Malik, who launched his company out of the program, doesn't need a balance sheet to explain what the building was. He remembers what it replaced. "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 is what an accelerator makes possible, and none of it shows up in a budget line. Founders rehearsed pitches in real rooms. They met investors without renting anything. They belonged to something before they could afford to be anywhere. Multiply that coffee-shop table by 212 companies, and you begin to measure what Milwaukee is losing. The city is trying to salvage what it can. Deputy Mayor Susan Cole said officials are speaking with building owners and local universities about space for the twelve displaced companies — the part of Foundry Hall that can, in principle, be replaced. Whether that effort lands before the end of the month is an open question. The rest cannot be. On the last day of the month, the lights go out on an institution that worked — 212 companies, 1,140 jobs, and, at the end, no check large enough to keep the door open. The companies will endure. The building that launched them will not.

MiniMax: MiniMax M3: # The Cheaper Model Tied the Expensive One. That's the Real Story. When five language models were put head-to-head on 200 real customer-support tickets, the most interesting result wasn't who won. It was how close the top two finished, and how different their price tags look. ## The Headline Finding Model North finished first with an average score of 84.2. Model Vale came in second at 83.8. The gap was small enough that it wasn't statistically significant — meaning a reasonable observer should treat them as effectively tied on quality. The price difference is anything but tied. North costs an estimated $1.90 per 1,000 tickets. Vale costs $0.74 — about 61% less. Vale also returned its answers roughly 18% faster than North. In other words, the most expensive model in the test delivered essentially the same measured quality as a model costing less than half as much. For most support operations, that single comparison matters more than any other number in the benchmark. ## How the Test Worked Researchers assembled 200 anonymized tickets from one software company, all in English. Each of the five models generated a single proposed answer per ticket. Three experienced support managers graded every response blindly, scoring factual correctness, usefulness, and tone. None of the graders knew which model produced which answer. The benchmark did not test agentic behavior — that is, models that click around a help-desk system or call external tools. It also didn't measure multilingual performance, since every ticket was in English. ## The Quality-Cost Curve Once you move past the top two, quality drops noticeably. Model Ember scored 79.1 at $0.31 per 1,000 tickets, Model Kite scored 76.4 at $0.22, and Model Harbor came in last at 71.0 for just $0.08. The data sketches a clear shape: the climb from Harbor-tier quality up to North-tier quality costs roughly 24 times more per 1,000 tickets, but most of that value comes from the step up to Vale, not from the final step up to North. Treating the gap between Ember and Kite as meaningful would be a mistake. Two points on a 100-point scale, judged by three graders, sits well within the variation you'd expect from a sample this size. ## Where Each Model Shines The two leaders weren't identical. North performed best on billing disputes, where precise language about charges and credits appears to matter. Vale performed best on technical troubleshooting, where its answers were slightly clearer and faster. For a team handling a steady mix of both, that pattern suggests a routing strategy: send billing tickets to North and troubleshooting tickets to Vale. The quality difference within each specialty is small, but the cost difference compounds over thousands of tickets. ## The Universal Weakness Every model in the test performed poorly when a ticket depended on an image attachment — a screenshot of an error message, a photo of damaged hardware, a PDF receipt. This is the one result that holds across the entire leaderboard. If your support volume leans heavily on visual evidence, no model in this benchmark is ready to handle those tickets without human review. ## What This Benchmark Can't Tell You A few caveats are worth keeping in mind. The tickets came from a single software company, so the language patterns, product vocabulary, and customer expectations reflect one specific business. Results on a retailer, a healthcare provider, or a logistics firm could look quite different. The graders were experienced, but they were three people, and taste inevitably plays a role in tone scoring. The test also excluded agentic actions, tool use, and non-English tickets. A model that's merely a good writer may rank very differently once it can read a customer's account history, look up an order, or escalate a case to a human. ## The Practical Takeaway If you're choosing a model for an English-language support workflow without tool access, the benchmark points to a clear default: start with Vale. You'll get quality indistinguishable from the most expensive option, at less than half the cost, with faster responses. Reach for North only where you've measured a specific gap — like billing disputes where every word counts. And budget for human review on anything image-based. None of these models are ready to handle that alone.

Z.ai: GLM 5.3 Prime: # A Tie at the Top: What a 200-Ticket Support Benchmark Says About Model Prices When researchers put five language models to work on 200 real customer-support tickets, the most expensive one earned the highest average score — by 0.4 points out of 100. The gap was so small that the study couldn't statistically distinguish the top two models at all. The cheaper of that pair, Model Vale, matched the premium Model North's quality at 39% of the cost, and answered 18% faster. That's the clearest finding here: at the top of the price range, cost and quality appear to have decoupled. ## How the test worked The setup was simple but well controlled. Researchers gathered 200 anonymized support tickets — all in English, all from one software company — and had each model generate a single proposed answer per ticket. Three experienced support managers then graded every answer blind, never knowing which model had produced what they were reading. Each score combined

MiniMax: MiniMax M3: # QueueZero’s AI Meeting Assistant Turns a Joke Into a Task QueueZero launched Proxy on Wednesday, an AI meeting assistant designed to attend meetings, produce notes, and assign follow-up work to software agents. The product debuted with a demonstration that illustrated both its capabilities and the risks of giving AI agents access to workplace conversations. During the launch event, an audience member noticed that QueueZero CEO Evan Wu was continuing to speak and 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 for Wu: one titled “Stop talking” and another titled “Send pricing sheet.” Wu laughed, completed the second task, and deleted the first. QueueZero says Proxy requires confirmation before it sends an external message, reducing the chance that an unintended comment will reach customers, colleagues, or the public. Tasks within the customer’s own systems can be created automatically without confirmation. Administrators can disable automatic task creation for organizations that prefer more restrictive controls. The setting offers a way to limit Proxy’s ability to act on conversations without removing features such as note-taking or suggestions for follow-up work. Proxy costs $30 per user per month and is scheduled to enter public beta next week. The launch highlighted a central challenge for AI agents operating in meetings: they can act on ambiguous remarks that people never intended as formal commands. In this case, the system took a complaint about the presentation and converted it directly into assigned work. That behavior could make meeting assistants more useful by reducing the need to manually document follow-ups. It also raises questions about context, consent, and control—especially when conversations contain jokes, criticism, or casual remarks that resemble instructions. QueueZero’s answer is a mix of safeguards and configuration. External communications require confirmation, while automatic creation of internal tasks can be turned off. The company did not provide further detail about how Proxy identifies actionable requests or whether administrators can set narrower limits for individual meeting types. The public beta will test whether those controls are sufficient as more customers allow Proxy to listen and act.

Z.ai: GLM 5.3 Prime: # QueueZero's AI Meeting Assistant Takes Audience Remark Literally, Assigns CEO the Task of "Stop Talking" QueueZero launched an AI meeting assistant called Proxy on Wednesday, and within the first minutes of its public demonstration the software had produced a two-item to-do list for the company's chief executive: stop talking, and send the pricing sheet. Proxy is designed to attend meetings, generate notes and assign follow-up work to software agents, placing it among a wave of tools that aim to move meeting software beyond passive transcription and toward automated action. The follow-up work is handed off to software agents rather than left in a notes document for humans to parse. 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. Two tasks then appeared, both assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. ## A Literal Turn The exchange drew laughter at the event, but it also pointed to the design question at the center of a new class of assistants: tools that act on what is said in a meeting rather than simply summarizing it. A remark delivered as a joke or an aside means one thing to a human listener and something else entirely to a system that treats each utterance as a potential command. ## Guardrails and Controls QueueZero says Proxy requires confirmation before an external message is sent. Internal tasks, by contrast, can be created automatically — the mechanism that produced Wu's impromptu assignments on Wednesday. The distinction leaves a broad middle ground in which tasks and internal items can be created without a sign-off, and Wednesday's episode fell squarely inside it. Administrators who prefer firmer limits can disable automatic task creation entirely. Meeting transcription tools have become commonplace in recent years, and the addition of agents marks a shift in ambition: rather than recording what was said, the software is now expected to do something about it. That shift raises questions about authority — who may create work, for whom, and under what review — that Wednesday's demonstration made concrete. ## Pricing and Availability Proxy costs $30 per user per month and enters public beta next week. By the end of the demonstration, one of Wu's two assignments had been completed and sent on its way. The other did not survive the afternoon.

MiniMax: MiniMax M3: # KiteBridge Software Glitch Triggers $2.8 Million in Duplicate Expense Payments A routine software update from expense-management firm KiteBridge unleashed more than 13,000 erroneous reimbursements this week, undermining a feature the company had publicly promoted as fail-safe and leaving dozens of corporate customers to sort out the financial fallout. ## What Happened At 8:45 a.m. on Tuesday, KiteBridge pushed version 4.2 of its expense-management platform live to customers. Within half an hour, the first complaint landed in the company's support inbox at 9:12 a.m. The trigger was a payment-processor timeout: when KiteBridge's primary transaction system failed to receive a confirmation in time, its fallback routine stepped in and treated the original request as never having been sent. Each subsequent retry was therefore processed as a brand-new reimbursement rather than a duplicate of the first attempt. The result was 13,400 duplicate payments totaling approximately $2.8 million across roughly 91 corporate customers. KiteBridge suspended its automated payment engine at 10:03 a.m. — about 75 minutes after deployment and just over 50 minutes after the first complaint — to stop the bleed. By Wednesday evening, KiteBridge said it had clawed back $2.3 million of the duplicate funds. The remaining roughly $500,000 was still being recovered from recipients. The company stressed that no individual employees had money pulled from their personal accounts; employers — who had been expecting to pay each expense only once — were the parties actually on the hook for the duplicates. ## A Promised Protection The failure cuts against a specific guarantee KiteBridge made in its April product changelog. At the time, the company described version 4.0 as introducing "exactly-once payment protection," language intended to assure finance teams that the system was engineered to prevent the very class of duplicate-transaction errors that occurred on Tuesday. The latest incident suggests that protection either was not comprehensive or did not extend to the timeout-and-retry scenario that triggered the cascade. ## A Communication Lag KiteBridge's handling of its public messaging has drawn additional scrutiny. The company's incident page initially characterized the disruption as "delayed reimbursements" — a framing that understated both the scale and the nature of the problem. The page did not disclose the existence of duplicate transactions until roughly six hours after automated payments had been suspended. For finance teams trying to reconcile their own books during that window, the delay meant working with incomplete information about what had actually happened. ## Response and Next Steps CEO Lena Ortiz issued a public apology and committed the company to publishing a full technical report on the root cause of the failure. The report is expected to detail how the fallback path interacted with the primary payment system and what changes KiteBridge plans to make to prevent a recurrence. For now, affected customers are reconciling accounts, pursuing the remaining funds, and reassessing whether KiteBridge's "exactly-once" guarantee — once a selling point for cautious finance departments — can be relied upon going forward.

Z.ai: GLM 5.3 Prime: # Bug in KiteBridge Update Triggered $2.8 Million in Duplicate Reimbursements *A payment-processor timeout triggered a flaw in KiteBridge's fallback system, duplicating reimbursements at 91 companies. Most of the $2.8 million has been recovered; the rest remains pending.* A software failure at expense-management company KiteBridge generated 13,400 duplicate employee reimbursements totaling $2.8 million on Tuesday, affecting 91 corporate customers before the company shut off its automated payment system and began clawing the money back. The incident, set off by a routine update, unfolded over less than an hour and 20 minutes on Tuesday morning. By Wednesday evening, $2.3 million had been recovered and roughly $500,000 remained pending. CEO Lena Ortiz apologized and said the company would publish a technical report on the failure. ## What went wrong The trouble began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2 of its expense-management software — the platform corporate customers use to review and reimburse employees for travel, meals and other work costs. When the software approves a claim, it sends a payment instruction to an outside payment processor, the intermediary that moves money into an employee's bank account. On Tuesday, the processor's confirmation that it had received an instruction failed to arrive within the expected window — an error known as a timeout. KiteBridge's software includes a fallback for such situations: when no confirmation arrives, it tries again, on the assumption that the original instruction may not have gone through. But on Tuesday, the fallback treated each retry as an entirely new reimbursement request rather than a repeat of the original one. Every attempt to recover a stalled payment instead issued a fresh one, multiplying reimbursements across customer accounts. The first customer complaint arrived at 9:12 a.m., less than half an hour after the deployment. KiteBridge suspended automated payments at 10:03 a.m., and by then — 78 minutes after the update went live — the software had generated 13,400 duplicates worth $2.8 million. ## A safeguard that didn't hold The episode carries an uncomfortable irony. An April changelog for version 4.0, an earlier release, said the software had gained "exactly-once payment protection." In payment engineering, "exactly-once" is a sought-after guarantee: a promise that every approved payment is made precisely one time — never duplicated, never dropped. Whatever the feature was designed to do, it did not prevent Tuesday's outcome, a reminder that safeguards built to handle one failure mode can open the door to another. ## Delayed disclosure KiteBridge's public incident page initially described the problem as delayed reimbursements. It did not disclose the duplicate payments until six hours after automation was suspended — just after 4 p.m. Tuesday. The characterization mattered. Suspending automated payments did leave reimbursements in limbo, but the underlying problem was the opposite of a delay: payments that should have been made once were made repeatedly. ## The recovery No employees had money withdrawn from their accounts. The duplicate payments were funded by employers, meaning the cost of the error fell on KiteBridge's customers rather than their staff — and getting the money back means reclaiming the excess from employee accounts, a process that can take days. As of Wednesday evening, $2.3 million of the $2.8 million had been returned, with the balance still pending. Ortiz apologized for the incident and said the company would publish a technical report. It remained unclear when KiteBridge would resume automated payments.

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 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.

MiniMax: MiniMax M3: 6

Z.ai: GLM 5.3 Prime: 6