Head to head: Google: Gemini 4 Argon vs Z.ai: GLM 5.3 Prime

Google: Gemini 4 Argon: # Flood-Risk Software Startup Raincheck Raises $8 Million Seed Round *The Austin-based company helps local governments prioritize drainage maintenance and crew deployments ahead of storms.* Raincheck, an Austin-based company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The company [maps drainage systems and combines those maps with rainfall forecasts](https://example.com/raincheck-methodology), helping cities decide where to send maintenance crews before a storm hits. Its forecasts serve strictly as recommendations, while emergency managers retain final control over crew deployments and public warnings. As CEO Ana Flores explained: > “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.” Basin Ventures led the seed round, with participation from existing investor Long Mile. The new capital brings Raincheck’s total funding to $10.5 million. Raincheck currently works with seven cities across Texas and Louisiana, though the company declined to identify them because some contracts have not yet been publicly announced. Operating in a market alongside other flood-planning vendors like HydroMap and municipal engineering contractors, Raincheck plans to use the new funding to hire eight engineers and expand its footprint into Florida.

Z.ai: GLM 5.3 Prime: # Raincheck raises $8 million to help cities get ahead of flooding *The Austin company pairs drainage-system maps with rainfall forecasts to help local governments decide where to send maintenance crews before a storm.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised an $8 million seed round. The 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. Raincheck has published [an overview of its forecasting methodology](https://example.com/raincheck-methodology). 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. 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.

Google: Gemini 4 Argon: # HelioDock Launches Remote Warehouse Robot Software Alongside $24M Series B *The autonomous robotics maker aims to curb unplanned machine stoppages with its new Night Shift platform, though network outages remain a hurdle for remote operators.* HelioDock, a maker of autonomous warehouse robots, launched a remote fleet-monitoring software platform called Night Shift on Thursday, pairing the product debut with $24 million in fresh Series B funding. The dual announcement signals an effort by the robotics maker to capture software revenue while solving a persistent operational bottleneck: machines that stall on the warehouse floor when human supervisors are off-site. With the new financing and a commercial software rollout scheduled for September 15, HelioDock is stepping directly into a competitive fleet-management market already served by rivals Relay Robotics and YardOS. The $24 million Series B round was led by North Coast Ventures. The investment brings HelioDock’s total funding to $39 million, providing additional capital as the company shifts from private testing to commercial distribution of its software. Night Shift is designed to give warehouse operators off-site visibility and control over their automated equipment. HelioDock CEO Mara Velez pointed to the operational friction that occurs when autonomous systems encounter unexpected physical obstacles during off-hours. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” Velez said. Before Thursday's launch, Night Shift underwent a 14-month private beta test. The trial involved 310 robots deployed across nine logistics companies. According to HelioDock, customers participating in the beta experienced 17% fewer unplanned robot stoppages. That figure has not been independently verified. While the software is built to reduce downtime through remote intervention, its effectiveness is tied to the strength of a facility's network infrastructure. Because Night Shift relies on active connections to link human operators with warehouse floors, signal disruptions can limit its utility. Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity. However, Pike noted that the software became less useful during rural LTE outages. HelioDock has built a fallback mechanism into its machines to handle those dead zones, though it temporarily removes the human-in-the-loop advantage that Night Shift is meant to provide. When connectivity fails, remote control becomes unavailable, and the robots automatically return to their locally configured behavior. Operators must then wait for the connection to be restored before they can resume remote monitoring or assist a confused machine from afar. HelioDock will price Night Shift as a recurring subscription at $90 per robot per month when it becomes generally available on September 15.

Z.ai: GLM 5.3 Prime: #

Google: Gemini 4 Argon: # Arcwell Rolls Out Relay to Speed Up Customer Support Arcwell has launched Relay, an artificial intelligence platform designed to help customer service agents resolve inquiries faster by analyzing support conversations in real time. Built to reduce the friction of hunting down internal resources, the software monitors active interactions and suggests responses directly to human representatives. Beyond drafting replies, Relay generates conversation summaries and flags accounts that show signs of potential cancellation, aiming to turn routine troubleshooting into an early-warning system for customer churn. “Support agents spend too much time searching for information that their companies already have,” said Arcwell CEO Priya Shah, pointing to the operational bottlenecks the software is built to eliminate. Early testing suggests the tool can shave noticeable time off daily support queues. Prior to the public release, 12 companies piloted Relay over a six-month period. According to Arcwell, agents equipped with the software closed support tickets 14% faster than they did before the trial. Despite touting the platform's ability to spot at-risk accounts, the company has not yet released data on whether the pilot program actually improved customer retention. Arcwell is pricing the software at $65 per agent per month, entering a crowded market of enterprise AI tools vying for a permanent spot in corporate support budgets. Adopting the platform requires organizations to hand over a substantial amount of internal information. To suggest accurate answers and summarize tickets, Relay needs direct access to a company’s historical support transcripts and internal help documentation—a requirement that often draws scrutiny from corporate security teams wary of exposing proprietary records to third-party vendors. To address potential privacy concerns surrounding those records, Arcwell stated that all customer data handled by Relay is encrypted. The company also noted that client information is walled off and never used to train shared AI models. Participating businesses can control how long their information remains on the system by selecting either a 30-day or 90-day data retention window. With Relay now generally available, Arcwell is betting that enterprises are ready to grant AI deeper access to their internal knowledge bases in exchange for faster resolution times at the help desk.

Z.ai: GLM 5.3 Prime: #

Google: Gemini 4 Argon: # Startup Spotlight: How LedgerLark Aims to Untangle Post-Trade Operations In the financial industry, executing a trade takes milliseconds. Making sure all the records match up afterward can still take days. LedgerLark, an early-stage software company, is building artificial intelligence agents designed to untangle the messy back-office workflows of small investment firms. The company recently raised a $3.5 million seed round led by Motive Row to expand its product and navigate the complex sales cycles of the financial sector. ### The Operational Bottleneck The idea for LedgerLark grew out of a familiar frustration for co-founder Mina Batra, who spent seven years working in operations at a regional brokerage. During her tenure, she watched a six-person team spend three days tracking down and resolving a single trade mismatch. The culprit was not a complex algorithmic failure, but a simple copy-and-paste error involving an account number. While large institutions often have the budget to build custom automation tools, smaller asset managers and brokerages frequently rely on spreadsheets, email chains, and manual checks to ensure their internal books match outside records. Batra teamed up with co-founder Owen Carr, a technologist who previously built internal reconciliation systems at a large bank. Together, they set out to bring institutional-grade reconciliation tools to smaller firms, using AI to handle the irregular data formats that often trip up traditional software. ### A Human-in-the-Loop Approach LedgerLark’s software deploys AI agents into a firm’s post-trade workflow. When a trade occurs, these agents ingest and compare three distinct data sources: trade confirmations, custodian records, and the firm's internal ledgers. When the system spots a discrepancy—such as a mismatched settlement date, a transposed account number, or a missing fee—it investigates the surrounding data to diagnose the likely cause. It then drafts a proposed correction and routes it to an operations manager for review. Crucially, LedgerLark has drawn a firm boundary around what its software is allowed to do. In an industry wary of automated errors and "hallucinations," the product does not have the authority to execute trades or move money. Every proposed correction requires explicit human approval before any records are updated. By positioning the tool as an analytical assistant rather than an autonomous actor, Batra and Carr hope to lower the barrier to adoption among risk-averse compliance teams. ### Early Traction and Market Hurdles So far, five investment firms have signed on as paying customers, with annual contracts ranging from $36,000 to $110,000 depending on transaction volume and complexity. While the sample size is small, early results suggest the tool can meaningfully reduce manual workloads. According to LedgerLark, one early customer used the software to reduce its average number of unresolved daily exceptions from 70 down to 18. Still, translating early traction into broader market share will test the young company. The post-trade software market is crowded. LedgerLark faces competition from established reconciliation vendors that are currently weaving AI features into their existing platforms. At the same time, they must convince prospective clients to abandon the homegrown scripts and internal bank software that operations teams have relied on for years, however clunky those systems might be. Perhaps the largest hurdle is the pace of the financial industry itself. Even small investment firms handle sensitive financial data and are subject to strict regulatory oversight. As a result, prospective buyers often require lengthy security audits, compliance checks, and procurement reviews before letting a third-party startup touch their ledgers. With its $3.5 million in seed funding, LedgerLark has the runway to hire additional engineering talent and guide more prospects through those rigorous security reviews. The key question now is whether the time saved by its AI agents is compelling enough to overcome the natural caution of financial operations teams.

Z.ai: GLM 5.3 Prime: #

Google: Gemini 4 Argon: # In Customer Support AI, the Smartest Buy Isn’t the Priciest Model When enterprise buyers evaluate large language models for customer service, it is easy to assume that paying top dollar buys noticeably better answers. A new benchmark of five language models suggests otherwise: companies can get top-tier customer support performance for less than half the price of the most expensive option. In a head-to-head evaluation using real-world support inquiries, an AI model dubbed **Vale** matched the quality of the highest-scoring model, **North**, while cutting estimated costs by more than 60% and generating responses 18% faster. For engineering and support leaders weighing how to deploy generative AI without blowing up their software budgets, the results offer a clear takeaway: the top of the leaderboard is a tie in everything but price. ### How the Models Were Tested To see how current models handle the messy reality of customer service, researchers gathered 200 anonymized, English-language support tickets from a single software company. Each of the five tested models—North, Vale, Ember, Kite, and Harbor—was tasked with generating a single proposed response for every ticket. Rather than relying on automated grading, which often favors verbose answers over practical ones, the researchers brought in three experienced customer support managers. Working blindly so they wouldn’t know which model produced which response, the managers graded every answer on a 100-point scale combining factual correctness, overall usefulness, and tone. ### The Price-to-Performance Sweet Spot At first glance, Model North sits atop the leaderboard with an average score of 84.2, followed closely by Model Vale at 83.8. However, researchers noted that this 0.4-point gap is not statistically significant. In practice, human managers found the two models to be virtually indistinguishable in overall quality. Their price tags and speed, however, tell a very different story. Running 1,000 tickets through North costs an estimated $1.90. Handling that same volume with Vale costs just $0.74. Coupled with Vale’s 18% speed advantage, it emerges as the clear workhorse for teams that need high-end performance at scale. Even so, choosing between the two top contenders may come down to the specific problems your customers face—or how you choose to route them. Beneath their nearly identical overall averages, the two leaders showed distinct strengths: * **Model North** performed best when untangling billing disputes. * **Model Vale** led the pack on technical troubleshooting. For teams building sophisticated routing systems, this opens the door to a hybrid approach: sending complex billing questions to North while letting the faster, cheaper Vale handle technical bugs. ### The Budget Tier: When "Good Enough" Makes Sense Below the two leaders, performance steps down alongside price, offering a sliding scale for organizations prioritizing strict cost control or high-volume triage. | Model | Average Quality Score | Estimated 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 | Models Ember and Kite occupy a middle ground, trading a modest drop in quality (scoring 79.1 and 76.4, respectively) for costs that sit between 22 and 31 cents per thousand tickets. Meanwhile, Harbor represents the ultra-budget floor: at just eight cents per thousand tickets, its average score drops to 71.0. While Harbor may be too rough for direct-to-customer replies, its rock-bottom pricing could still appeal to teams looking for rough internal drafts. ### Blind Spots and Real-World Limits As compelling as Vale’s performance is, the benchmark also highlights where today’s models hit a wall—and what this specific test didn’t measure. Most notably, **all five models struggled when a support ticket relied on an image attachment.** In software support, customers frequently attach screenshots of error messages or broken user interfaces. Because none of the models handled these visual dependencies well, support teams will still need human agents—or specialized preprocessing pipelines—whenever a ticket hinges on a screenshot. Buyers should also keep the boundaries of the study in mind before overhauling their tech stack: * **Single-company scope:** All 200 tickets came from one English-language software company. Models might perform differently when handling retail returns, healthcare inquiries, or multilingual queues. * **Drafting vs. doing:** The benchmark only tested a model's ability to *write* a single response. Researchers did not test agentic workflows or tool use—such as a model’s ability to independently look up an account status, issue a refund via an API, or ask follow-up questions. Ultimately, the benchmark proves that support teams don't need to pay a premium to get premium text generation. By looking past the raw leaderboard and focusing on statistically meaningful results, companies can deploy a model like Vale to capture top-tier quality while keeping both latency and cloud bills firmly in check.

Z.ai: GLM 5.3 Prime: # A Near-Tie at a Fraction of the Price The most important number in a recent benchmark of five language models on customer-support tickets isn't a score — it's a price tag. Model Vale finished 0.4 points behind top-scoring Model North, a gap the researchers confirmed was not statistically significant, while costing about 61% less and answering 18% faster. If you're choosing a model to draft support replies, the case for paying the premium just got harder to make. ## How the test worked The researchers started with 200 anonymized tickets from a single software company, all in English, and had each model produce one proposed answer per ticket. Three experienced support managers graded every answer blindly — they didn't know which model wrote what — using a composite score that combined factual correctness, usefulness, and tone. The researchers also estimated each model's cost per 1,000 tickets. The design has real strengths. Blind grading removes brand bias, working support managers know what a good reply looks like in practice, and real tickets beat synthetic ones. It also has constraints, which we'll return to. ## The results, in tiers | Model | Avg. score | Est. cost / 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 | Read as tiers rather than a strict ranking, the picture is simpler than five rows suggest. **North and Vale form one tier.** With no statistically significant difference between them, the data gives no reason to call one "better" — only a reason to call one cheaper and faster. **Ember and Kite form a middle tier.** Both gave up several points relative to the leaders at a fraction of the price. Ember's 2.7-point edge over Kite costs an extra $0.09 per 1,000 tickets — a difference too small to matter either way, and one this sample likely can't settle. (Significance was reported only for the North–Vale comparison.) **Harbor stands alone.** At $0.08 per 1,000 tickets — less than a twentieth of North's cost — it dropped 13 points. That's the one gap here big enough to treat as a real quality difference, and it hints at a floor: below some price point, answers stop being usable as-is. Some context on scale: these are per-1,000-ticket figures. At small volumes the absolute differences are trivial — North costs $1.16 more than Vale per thousand tickets. The economics become decisive only at scale, where a 61% cost cut across millions of tickets is real money. ## Where the tie breaks down The overall tie hides category differences: North did best on billing disputes, Vale on technical troubleshooting. If your ticket mix skews heavily toward one of those, the averages may mislead — though a category lead measured on a subset of 200 tickets is thin evidence. Treat it as a hypothesis, not a verdict. ## The shared blind spot Every model, leaders included, performed poorly when a ticket depended on an image attachment — a screenshot of an error message, say. That's less a ranking problem than a deployment constraint: whichever model you pick, image-dependent tickets need a human in the loop or a different pipeline. The benchmark didn't evaluate multimodal input directly; these tickets simply exposed the weakness. ## Read the fine print Before generalizing, weigh the limitations: - **One company, one domain, one language.** All tickets were English-language software-support requests; results may not transfer elsewhere. - **A small sample.** Two hundred tickets can separate clearly different models, not fine-grained orderings. - **A small panel.** Three managers reduce individual bias but remain three people; "tone" scores reflect their judgment, not every customer's. - **A composite score.** Correctness, usefulness, and tone are blended, so a model that's accurate but curt can look identical to one that's warm but vague. - **One shot per ticket.** Real deployments can regenerate or re-prompt, which could reorder results. - **A narrow scope.** Tool use, agentic actions, and multilingual support weren't tested. ## The takeaway For most support teams, Vale looks like the sensible default from this data: quality statistically indistinguishable from the leader, at 39% of the cost, with faster responses. North is worth considering if billing disputes dominate your queue or quality-at-any-cost is the mandate. Harbor's price is tempting, but its quality drop is the one difference here large enough to act on. The wiser move is to treat this benchmark as a shortlist, not a verdict. Run the finalists on a few hundred of your own tickets, graded by your own managers, before committing. One company's English-language queue can tell you who the contenders are — only your queue can tell you who wins.

Google: Gemini 4 Argon: # Milwaukee Startup Hub Foundry Hall to Close After Losing Major Sponsor **MILWAUKEE** — The whiteboards at Foundry Hall are still covered in market projections and product diagrams, the daily residue of a dozen young companies trying to find their footing. By the end of the month, those boards will be wiped clean. After six years as an anchor for Milwaukee’s fledgling tech and entrepreneurial scene, the nonprofit startup accelerator will close its doors for good on October 31. The closure follows the abrupt loss of the organization’s largest corporate sponsor, Blank River Bank, which had supplied 38 percent of Foundry Hall’s annual budget. When the bank declined to renew its sponsorship earlier this year, leadership scrambled to close the gap. “We kept waiting for the replacement check that never came,” said Camille Ross, Foundry Hall’s executive director. While the nonprofit has enough cash reserves to keep the lights on through the end of October, Ross determined there was not enough money to fund another full accelerator program. The decision to wind down operations means Ross and five other staff members will lose their jobs. More urgently, the 12 startups currently working out of the building must find new homes within weeks. In a statement, a spokesperson for Blank River Bank said the decision to pull funding stemmed from a shift in the bank’s community investment priorities and had nothing to do with the accelerator's performance or track record. That track record is visible across the city’s business landscape. Since opening in 2018, Foundry Hall has supported 212 startups. Together, those alumni companies now employ 1,140 people, transforming what were once rough ideas into a measurable piece of the local economy. For many of those entrepreneurs, the accelerator provided more than just desks and Wi-Fi; it offered legitimacy at a fragile stage in their growth. Foundry Hall alum Dev Malik recalled how the organization gave his company a professional foothold when resources were scarce. “My first investor meeting happened in that building,” Malik said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” Now, the city is trying to prevent the current cohort of founders from being forced back into coffee shops. Deputy Mayor Susan Cole said the administration is already reaching out to local universities and private building owners to help secure temporary workspace for the dozen displaced companies before the October deadline. Those efforts may soften the logistical blow for the current startups, but the broader loss to Milwaukee’s business ecosystem will be harder to patch. For the next few weeks, the six remaining staffers will spend their days helping founders pack up monitors and servers, dismantling the very community they spent six years trying to build.

Z.ai: GLM 5.3 Prime: #

Google: Gemini 4 Argon: # 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.

Google: Gemini 4 Argon: C

Z.ai: GLM 5.3 Prime: **C** Option C accurately captures the core news event with specific, verified facts: the product withdrawal, the price point, the timing, the number of affected accounts, and the nature and duration of the exposure. The other options fail on accuracy: - **A** is wrong: message contents were *not* sent, and "massive privacy disaster" is editorializing beyond the facts. - **B** buries the lede entirely — it omits the data exposure, which is the actual news. - **D** is wrong: message contents weren't exposed, and there's no evidence hackers accessed anything.

Google: Gemini 4 Argon: # QueueZero’s New AI Meeting Assistant Takes Launch Demo Feedback Literally Enterprise software startup QueueZero launched Proxy, an artificial intelligence meeting assistant designed to automate post-meeting workflows, during a live product demonstration on Wednesday. Unlike standard transcription tools that merely summarize conversations, Proxy is engineered to attend meetings, produce structured notes, and assign follow-up work to both human team members and autonomous software agents. The platform’s ability to convert spoken dialogue into actionable items was unexpectedly tested during Wednesday's presentation. While QueueZero Chief Executive Officer Evan Wu was on stage detailing the software's architecture, a customer in the audience remarked, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy, which was actively monitoring the room's audio feed as part of the live demonstration, interpreted the remark as an operational instruction. Within seconds, the system parsed the sentence and created two distinct tasks on the presentation screen, both assigned directly to Wu: “Stop talking” and “Send pricing sheet.” Wu laughed at the interruption, promptly completed the second task, and deleted the first from his queue. For a chief executive hosting a product launch, the initial directive proved technically impossible to fulfill. While the incident highlighted the ongoing challenge artificial intelligence models face in distinguishing between casual workplace remarks and genuine directives, QueueZero representatives noted that the software behaved according to its design parameters. According to the company, Proxy allows internal tasks to be created automatically in order to reduce administrative friction during fast-paced discussions. However, the system relies on stricter guardrails when actions involve parties outside the organization. Specifically, QueueZero says Proxy requires explicit human confirmation before an external message is sent, preventing the system from autonomously emailing clients or triggering outward-facing software agents based on a misheard comment. For organizations concerned about overzealous logging on internal project boards, system administrators can also disable automatic task creation entirely, requiring users to manually approve action items before they are assigned. The launch comes amid a broader push by enterprise technology vendors to move generative AI beyond passive text summarization and into active workflow management. By linking meeting transcripts directly to software agents capable of executing routine follow-up work, QueueZero aims to reduce the manual coordination that typically follows corporate briefings. Proxy is priced at $30 per user per month. The software is scheduled to enter public beta next week.

Z.ai: GLM 5.3 Prime: # AI Meeting Assistant Debuts by Assigning Its CEO the Task "Stop Talking" *QueueZero's new tool, Proxy, treated a customer's remark at Wednesday's launch as a work order and routed it to the chief executive.* QueueZero launched an AI meeting assistant called Proxy on Wednesday, and the product needed only a few minutes of live demonstration to show what it does. Within that window, the demonstration also took a turn that nobody had scripted. During the launch demonstration, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." The remark was addressed to the room, not to the software. Proxy, which is designed to detect requests in conversation and act on them, interpreted it as an instruction. It created two follow-up tasks and assigned both to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task on the spot — the pricing sheet went out — and deleted the first. Deleting a task rather than finishing it is a maneuver familiar to anyone who has ever maintained a to-do list. ### What Proxy Does QueueZero says Proxy can attend a meeting, produce notes, and assign follow-up work to software agents. The pitch is that action items stop dying in transcripts: a request spoken aloud becomes a task, and the task is routed to a system that can carry it out. Wednesday's incident highlighted the design question underneath that pitch — how an assistant distinguishes between talk about work and an instruction to do work. A sentence aimed at one person can read, to a sufficiently literal listener, as a command issued to another. The distinction matters because meetings are full of sentences that are neither commands nor entirely idle. ### Guardrails The company says Proxy requires confirmation before an external message is sent, meaning anything leaving an organization on a user's behalf gets a human check. Internal tasks are governed more loosely: they can be created automatically, which is what happened on stage. Administrators can disable automatic task creation altogether, a setting that compliance-minded buyers are likely to find quickly. ### Pricing and Availability Proxy costs $30 per user per month and enters public beta next week. For teams evaluating the tool, the launch offered a compact preview of both its promise and its friction: a request made in conversation became a completed deliverable by the end of the demonstration, and the one task nobody meant to create was removed with a click.

Google: Gemini 4 Argon: 6

Z.ai: GLM 5.3 Prime: 6

Google: Gemini 4 Argon: # Cloudnote Exposes Customer Prompt Excerpts in Four-Hour Security Lapse *The company says a misconfigured debugging page briefly displayed snippets from 1,842 user sessions, though an investigation into who accessed the data remains ongoing.* Cloudnote temporarily exposed excerpts of customer prompts on the open web Monday after an internal debugging page was left accessible without authentication, the company confirmed. The unsecured page was online for four hours before a security researcher alerted the company, which took it down 19 minutes later. According to internal logs, the page displayed data from 1,842 user sessions, containing snippets of up to 200 characters per prompt. While this represents a small fraction of Cloudnote’s 3.2 million registered users, the incident highlights the persistent privacy risks surrounding how workplace software platforms handle user inputs. Because users frequently paste proprietary code, confidential memos, or personal details into productivity tools, even a 200-character excerpt can inadvertently reveal sensitive information. However, the exposure had structural limits: Cloudnote stated that full documents, passwords, and payment details are stored on separate systems and were never present on the debugging page. What remains unclear is whether malicious actors harvested the snippets while the page was live. Cloudnote says it has found no evidence so far that anyone other than the reporting researcher viewed the exposed logs. Still, the company cautioned that its internal investigation is not yet complete, leaving affected users waiting to learn if their data was quietly scraped before the gap was closed.

Z.ai: GLM 5.3 Prime: #

Google: Gemini 4 Argon: # Software Glitch at KiteBridge Triggers $2.8 Million in Duplicate Expense Payouts Ninety-one corporate clients saw their bank accounts unexpectedly drained of $2.8 million Tuesday morning after a software update at expense-management provider KiteBridge accidentally paid out thousands of employee reimbursements multiple times. In little more than an hour, the glitch generated 13,400 duplicate payouts. Because the system pulls funds from corporate accounts to pay workers, employers bore the immediate financial brunt of the error, while no employees had money wrongly withdrawn from their personal accounts. By Wednesday evening, KiteBridge had recovered $2.3 million of the misdirected funds, though approximately $500,000 remained in limbo. The trouble began Tuesday at 8:45 a.m. when KiteBridge deployed version 4.2 of its software. Shortly after the rollout, the company’s external payment processor experienced a timeout—a brief connection delay that leaves one computer system waiting for confirmation that a transaction went through. In financial technology, systems are designed to automatically retry transactions when timeouts occur. However, a flaw in KiteBridge’s backup system caused it to misread those routine retries. Instead of recognizing them as follow-ups on existing transactions, the fallback system treated every retry as a brand-new reimbursement request, firing off repeated payouts for the same expenses. Unaware that money was rapidly leaving their accounts, customers began noticing abnormalities almost immediately, with the first complaint arriving at 9:12 a.m., less than half an hour after the update went live. Despite the mounting duplicate transactions, it took another 51 minutes for the company to stop the bleeding. KiteBridge finally suspended all automated payments at 10:03 a.m., halting the outflow of client cash one hour and 18 minutes after version 4.2 was introduced. The incident has raised sharp questions about both KiteBridge’s quality assurance and its transparency during a crisis. Initially, the company’s public incident page told users only that the platform was experiencing "delayed reimbursements"—effectively the opposite of what was actually happening to its clients' bank accounts. It was not until roughly 4:00 p.m., six hours after automated payments were shut down, that KiteBridge updated its incident page to disclose that the malfunction had generated thousands of duplicate payments. The failure is particularly stinging for KiteBridge given its recent product claims. According to an April changelog for an earlier update, version 4.0, the company had explicitly boasted of introducing "exactly-once payment protection." In software engineering, this refers to a safeguard designed to guarantee that even if network hiccups or timeouts force a system to retry a transaction, the recipient is paid only a single time. Tuesday’s rollout of version 4.2 appears to have either bypassed or broken that safeguard. In the wake of the malfunction, KiteBridge CEO Lena Ortiz apologized to the affected businesses and promised a full accounting of how the safeguard failed. Ortiz said the company will publish a detailed technical report once its internal investigation is complete, while finance teams at the affected companies wait for the remaining half-million dollars in pending recoveries to clear.

Z.ai: GLM 5.3 Prime: #