Head to head: Mistral: Mistral Large 4 vs Z.ai: GLM 5.3 Prime

Mistral: Mistral Large 4: # The Last Cohort The fluorescent lights in Foundry Hall's main bay will stay on until the end of the month, illuminating a space that has grown gradually quieter over the past six months. Six years after opening in a converted warehouse on Milwaukee's near south side, the nonprofit accelerator is emptying out. Twelve startups are packing servers and whiteboards into cardboard boxes, while six employees clear out the small offices lining the mezzanine, their last day set for October 31. Since 2018, the organization has incubated 212 startups—companies that now collectively employ 1,140 people across the region. Running in twelve-week cycles, the program paired early-stage founders with mentors from Milwaukee's established manufacturers and tech firms, creating a pipeline that turned the warehouse into a factory for new businesses. But the infrastructure supporting that work depended on a single pillar: Blank River Bank, which supplied 38 percent of the annual budget. When the bank declined to renew its sponsorship this spring, the math became unsolvable. The accelerator has enough cash to operate through October, but not to fund another full program cycle, leaving it to close rather than continue diminished. "We kept waiting for the replacement check that never came," said Executive Director Camille Ross, describing the months of quiet desperation as she reached out to donors who did not return her calls. For Dev Malik, whose logistics software company graduated from the program in 2021, the building represents a before-and-after. "My first investor meeting happened in that building," he said, standing recently in the lobby where founders once waited nervously for mentors. "Before Foundry Hall, I was pitching from the back table of a coffee shop." The distinction matters: the difference between a concept and a company often hinges on having a room, a legitimacy, a door that opens to capital. A spokesperson for Blank River Bank said the withdrawal reflected "a change in our community investment priorities" and was "unrelated to Foundry Hall's performance." But spreadsheets do not capture what the twelve current residents are losing—not just desks, but the daily proximity to mentors who have navigated the same regulatory mazes, the same hiring droughts, the same 2 a.m. anxieties about runway. Deputy Mayor Susan Cole said the city is working with building owners and local universities to find space for the current companies, a last-minute effort to prevent a scattering of talent before the doors close. On the final morning, the hallways will echo differently, the scuffed concrete floors bare where twelve desks once stood. The 1,140 jobs will continue in their own offices across the city. The 212 companies will persist in the market, some thriving, some folded, all carrying the mark of their origin. But the specific gravity of this place—the particular alchemy of founders sharing coffee at dawn, the whiteboards covered in half-erased roadmaps, the sound of a door opening to possibility—will disperse into Milwaukee's wider ecosystem. What remains is a building waiting for new tenants, and a city reckoning with the cost of losing a room where the future used to practice being real.

Z.ai: GLM 5.3 Prime: #

Mistral: Mistral Large 4: **KiteBridge Software Glitch Triggers $2.8 Million in Duplicate Reimbursements** A routine software deployment Tuesday morning escalated into a significant financial incident for expense-management company KiteBridge, whose platform processed 13,400 duplicate reimbursement payments totaling $2.8 million across 91 corporate customers before engineers could halt the automated system. The sequence began at 8:45 a.m. Tuesday when KiteBridge released version 4.2 of its software. By 9:12 a.m., the first customer complaints had arrived, yet the company initially described the disruption on its public status page as "delayed reimbursements"—a characterization that persisted for hours despite the growing chaos behind the scenes. At the technical root of the failure was a breakdown in how the system handled payment retries. When an upstream payment processor experienced timeouts—a common occurrence in high-volume financial networks—KiteBridge's fallback infrastructure failed to distinguish between a retry of an existing transaction and a new request. Instead of recognizing idempotency keys that would mark a retry as a continuation of the original payment, the software treated each automated retry as a fresh reimbursement authorization. In distributed payment systems, such safeguards rely on unique identifiers to prevent duplicate processing during network instability, but version 4.2's logic circumvented these protections, transforming a temporary network hiccup into a mass-duplication event. The irony deepened when examining KiteBridge's own documentation. An April changelog for version 4.0 had prominently featured "exactly-once payment protection"—a technical safeguard specifically engineered to ensure that each reimbursement was processed precisely one time, even during system failures or network interruptions. Tuesday's incident demonstrated that this protection failed precisely when it mattered most. KiteBridge suspended all automated payments at 10:03 a.m., roughly 78 minutes after the first complaint and nearly 90 minutes after the faulty deployment began propagating through customer systems. However, the company waited until approximately 4:00 p.m.—six hours after automation was suspended—to update its incident page and disclose that the problem involved duplicate payments rather than mere delays. The financial burden landed on employers rather than individual employees. While no workers experienced unauthorized withdrawals from their personal bank accounts, the 91 affected corporations funded the erroneous duplicate payments, effectively reimbursing employees twice for single expenses. For these companies, the error created immediate accounting reconciliation challenges and temporary cash-flow disruptions during a critical period for many businesses processing expense cycles. By Wednesday evening, KiteBridge reported recovering $2.3 million of the $2.8 million in erroneous disbursements, though $500,000 remained outstanding as the company worked with banking partners to reverse the transactions and restore correct balances. CEO Lena Ortiz apologized for the incident and pledged to publish a comprehensive technical report detailing how the failure occurred. The episode has drawn scrutiny to the company's incident-response protocols, particularly the delay in accurately characterizing the problem publicly and the failure of safeguards marketed as preventing exactly this type of error.

Z.ai: GLM 5.3 Prime: #

Mistral: Mistral Large 4: 6

Z.ai: GLM 5.3 Prime: 6

Mistral: Mistral Large 4: # Arcwell Launches Relay, an AI Platform to Speed Customer Service Resolutions Arcwell has launched Relay, an artificial intelligence platform that analyzes customer support conversations to suggest answers to agents, generate case summaries, and flag customers who may be at risk of canceling. During a six-month pilot involving 12 companies, agents using the tool closed tickets 14% faster, according to the company. The launch represents a significant development for Arcwell and the broader customer service industry, where artificial intelligence is rapidly transforming how companies handle support inquiries. The tool positions Arcwell in a competitive market of vendors racing to prove that machine intelligence can augment human agents rather than replace them, a distinction that has become crucial as enterprises grapple with the limitations of fully automated support. Unlike customer-facing chatbots, Relay operates as an agent-assist platform, working behind the scenes to surface relevant information from internal help documents and historical support transcripts as conversations unfold. The platform's capabilities extend beyond drafting suggested responses. Relay scans dialogue for sentiment shifts that might indicate impending churn, and automatically compiles detailed summaries once cases are resolved. The service costs $65 per agent each month. "Support agents spend too much time searching for information that their companies already have," said CEO Priya Shah. The tool's architecture requires deep access to sensitive corporate data, including private support transcripts and proprietary internal documentation. To address privacy concerns, Arcwell states that customer information is encrypted and is not used to train shared AI models, ensuring that proprietary business data remains siloed rather than contributing to a collective intelligence pool. Companies can configure data retention periods of either 30 or 90 days, giving them flexibility over how long conversation data remains stored. While Arcwell did not release specific customer-retention results from the pilot, the company is betting that faster resolution times will translate into lower churn rates. Relay marks an important new chapter for Arcwell as it seeks to help companies improve customer service through artificial intelligence, signaling a continued industry shift toward automated agent assistance.

Z.ai: GLM 5.3 Prime: #

Mistral: Mistral Large 4: # HelioDock Raises $24 Million, Debuts Night Shift Software for Remote Warehouse Robot Monitoring *The fleet management platform enters general availability September 15 after a 14-month beta involving nine logistics companies and 310 robots, though rural connectivity limitations persist for some users* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, a software platform that enables remote monitoring of robot fleets, and announced a $24 million Series B funding round led by North Coast Ventures. The investment brings the company's total funding to $39 million and coincides with the product's transition from a 14-month private beta to general availability on September 15. Night Shift addresses a persistent operational challenge in automated logistics facilities, where robot stoppages traditionally require physical intervention from on-site personnel, sometimes during overnight hours when staffing is minimal. The software allows warehouse operators to monitor fleet activity from remote locations, providing visibility into robot status and operations without requiring managers to travel to facilities when autonomous units encounter obstacles they cannot navigate independently. The platform enables oversight of robot fleets from outside the warehouse environment. The platform spent 14 months in a private beta involving nine logistics companies and a combined fleet of 310 robots. This extensive testing period allowed HelioDock to evaluate Night Shift across a substantial deployment of autonomous equipment before making the software available to the broader market. The beta participants operated the software within their existing warehouse workflows, allowing the company to assess performance across different operational settings. HelioDock claims that beta participants experienced 17% fewer unplanned robot stoppages while using Night Shift, though the company has not provided independent verification of this performance figure. The company reported the statistic following the 14-month evaluation period, indicating improved robot availability during the testing phase. Unplanned stoppages occur when robots encounter situations they cannot resolve autonomously, requiring human intervention to resume normal operations. "We want to remove the burden of physical intervention from warehouse management," said CEO Mara Velez. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet." The statement reflects the company's focus on addressing scenarios where autonomous robots require human assistance to resume operations, particularly during hours when on-site staffing is typically reduced. The system's effectiveness varies significantly based on connectivity infrastructure, according to beta participants. Aaron Pike of Fulcrum Logistics, one of the nine companies involved in the private beta, reported that Night Shift worked well in facilities with strong network coverage but became less useful during rural LTE outages. When connectivity fails, the robots automatically return to their locally configured behaviors, rendering remote control unavailable. During these outages, operators lose the ability to monitor or direct fleet activity until the network connection is restored, creating a gap in oversight that could affect warehouse operations in areas with unreliable cellular service. The $24 million Series B round, led by North Coast Ventures, brings HelioDock's total funding to $39 million. The investment provides capital for the company as it transitions Night Shift from beta testing to commercial availability, supporting the expansion of its software business alongside its existing robotics hardware operations. HelioDock enters a market already served by competing fleet management products from Relay Robotics and YardOS, which offer similar remote monitoring capabilities for autonomous warehouse vehicles. Night Shift will be priced at $90 per robot per month when it becomes generally available to customers on September 15. The pricing structure introduces a recurring revenue model for HelioDock, charging customers for each robot connected to the software platform alongside its existing hardware business.

Z.ai: GLM 5.3 Prime: #

Mistral: Mistral Large 4: # LedgerLark Bets Small Investment Firms Will Buy AI for the Boring Work After a Trade The trade didn't fail. The account number did. Mina Batra watched a six-person operations team at her regional brokerage spend three days tracing a mismatch caused by a single copied digit. The fix took minutes; the investigation took days. It was this kind of Sisyphean labor—human beings hunting discrepancies across PDFs and spreadsheets—that convinced Batra the post-trade world needed a different kind of automation. Batra, who spent seven years in brokerage operations, and Owen Carr, who previously built reconciliation systems at a large bank, founded LedgerLark to address that gap. The company, which recently raised a $3.5 million seed round led by Motive Row, builds AI agents that sit between trade confirmations, custodian records, and internal ledgers. The agents flag discrepancies and propose corrections, but they do not execute trades or move money. A human still signs off. This narrow scope appears deliberate. While much of the current AI frenzy focuses on autonomous agents that act independently, LedgerLark is targeting a compliance-sensitive workflow where mistakes carry regulatory and financial consequences. The founders argue that small investment firms—those without the budgets of bulge-bracket banks—desperately need help with exception handling but cannot afford the legacy reconciliation software built for institutions with thousands of employees. Early traction suggests some demand. LedgerLark reports five paying customers, with annual contracts ranging from $36,000 to $110,000. The company claims one client reduced its unresolved daily exceptions from 70 to 18, though the sample size remains small and the duration of that improvement unclear. These figures are modest by enterprise software standards, yet they indicate that at least a handful of firms are willing to pay for automation in an area typically dominated by manual processes or expensive incumbents. The competitive landscape is crowded. Established reconciliation vendors have served banks for decades, and many large institutions maintain internal software teams that build bespoke solutions. LedgerLark must convince potential customers that its AI agents can outperform both commercial incumbents and in-house builds—without triggering the lengthy security and procurement reviews that can stall financial services sales for months. There are reasons for caution. With only five customers, it remains unclear whether LedgerLark's approach scales beyond its initial beachhead. The company's reliance on human approval for corrections, while prudent, may limit its appeal to buyers seeking full automation. And the $3.5 million seed round, while sufficient for early development, will not sustain a long sales cycle against entrenched competitors with deeper pockets. What LedgerLark offers is a test of whether AI agents can earn trust in the unglamorous middle of financial infrastructure. If the company can navigate the procurement gauntlet and prove its agents reduce exception queues without introducing new risks, it may find a foothold in a market that larger vendors have neglected. For now, it is watching whether those first five customers renew—and whether the next fifty will sign.

Z.ai: GLM 5.3 Prime: # LedgerLark Wants AI Agents to Clear Post-Trade Exceptions — With Human Sign-Off *The startup, founded by a brokerage operations veteran and a reconciliation engineer, compares trade records and proposes fixes for human approval.* ## The three-day break The idea for LedgerLark traces back to a copied account number. Mina Batra, then working in operations at a regional brokerage, watched a six-person team spend three days resolving a single trade mismatch caused by one mistyped field. Seven years in brokerage operations had taught her that the incident wasn't exceptional — it was the texture of the job. Exception queues, the daily pile of trades that don't line up across confirmations, custodian records, and internal books, are mostly made of small breaks: an account number entered wrong, a quantity off, a settlement date a day late. What makes them expensive is the labor. Each break requires someone to pull the records, locate the divergence, work out which side is right, and draft a correction. Co-founder Owen Carr spent years building reconciliation systems at a large bank, where that labor is absorbed by big engineering teams and licensed vendor platforms. The gap the two saw: small investment firms have the same error rates without the same tooling. Enterprise reconciliation software is priced and configured for institutions with large operations staffs; building internally requires engineers small firms don't have. ## Proposals, not permissions LedgerLark's agents sit between a firm's confirmations, custodian records, and internal ledgers, comparing them line by line. When they spot a mismatch, they flag the divergence and propose a specific correction, which a human then approves or rejects. The scope is deliberately narrow. The agents don't execute trades or move money — they're a proposal layer, not a hands-on-keyboard one. That looks as much like a sales decision as an engineering one: in post-trade, where errors compound quickly, "an AI suggests, a person decides" is far easier to defend to a compliance officer than autonomy would be. That narrowness concentrates rather than eliminates the risk. A proposed correction is only as good as the agent's matching, and a busy operator clicking through approvals could rubber-stamp a subtly wrong fix. Whether the proposals hold up against messier data, and whether humans stay attentive in the approval loop at volume, are the open technical questions. ## Five customers, one headline number Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. LedgerLark says one customer cut its unresolved daily exceptions from 70 to 18 — a self-reported figure from a single firm, but suggestive if it replicates. The company has raised a $3.5 million seed round led by Motive Row. The growth math deserves skepticism more than celebration. Post-trade buyers are financial institutions, which tend to put vendors through lengthy security and procurement reviews before anything gets signed. Five customers could take a long time to become twenty, and while the contracts are real, they're modest next to enterprise reconciliation deals. ## Patient incumbents LedgerLark isn't inventing reconciliation. Established vendors have sold matching engines into the industry for a long time, along with integrations, audit trails, and the regulatory familiarity that comes with incumbency. Large banks run internal software tuned to their own workflows. The startup's bet is that those players optimized for big institutions and left small firms behind — and that AI agents change the economics of serving them, because agents can adapt to whatever formats and quirks a small shop actually uses rather than requiring a standardized implementation. If the bet is right, incumbents will likely bolt similar AI features onto platforms that already have customers. If it's wrong, small firms may keep absorbing exception queues as a cost of doing business, much as they have for years. Either way, the signal to watch is an unglamorous one: whether exception counts keep falling at more firms, and whether any agent-proposed fix turns out to be wrong in ways nobody caught.

Mistral: Mistral Large 4: # Cloudnote Exposes Prompt Fragments from 1,842 Users in Unsecured Debug Page *The AI note-taking service confirmed that excerpts of up to 200 characters were accessible without authentication for four hours on Monday, though full documents, passwords and payment data were stored separately.* Cloudnote has confirmed that a debugging page containing excerpts of customer prompts was accessible without authentication for approximately four hours on Monday. The company, which serves 3.2 million registered users, said the exposure involved up to 200 characters from user prompts across 1,842 separate sessions. The exposed content consisted solely of prompt fragments, not the complete documents users may have stored in the service. Cloudnote stated that full documents, passwords and payment information were stored in separate, secure systems and were not present on the exposed page. The company has not specified what types of information appeared in the visible excerpts. A security researcher discovered the unauthenticated page and notified Cloudnote, which removed the page 19 minutes after receiving the report. The company said its preliminary review of access logs found no evidence that anyone other than the researcher viewed the page during the exposure window, though it emphasized that its investigation remains ongoing. The incident highlights ongoing concerns about how AI services handle user data during development and debugging processes. While Cloudnote maintains that the exposure was limited in scope and duration, the company has not yet determined whether any of the 1,842 affected users will be notified individually, or what specific safeguards will prevent similar exposures in the future.

Z.ai: GLM 5.3 Prime: #

Mistral: Mistral Large 4: C

Z.ai: GLM 5.3 Prime: **C** Option C accurately captures the key facts—who (ByteBarge), what (pulled its AI assistant), when (Tuesday), why (workspace titles exposed), and scale (4,860 accounts, 11 hours)—without editorializing. The other options fail on accuracy: A claims "private messages" were leaked, but message contents were not sent. D falsely claims everything users wrote was visible to hackers, contradicting both the fact that contents weren't exposed and the absence of evidence of malicious access. B buries the news entirely and reads like promotional copy.

Mistral: Mistral Large 4: # QueueZero Launches AI Meeting Assistant Proxy, Which Immediately Manages Its CEO QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant designed to attend virtual conferences, generate detailed minutes, and automatically assign follow-up work to software agents. The tool represents the latest entry in a crowded market of AI-powered productivity software aimed at reducing administrative overhead for remote and hybrid teams, promising to eliminate the need for human note-takers in routine business discussions while integrating with existing enterprise workflow platforms. The product's debut, however, offered an unplanned demonstration of its literal interpretation capabilities. During a live demonstration before an audience of prospective customers and industry analysts, a participant remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." The comment drew laughter from the audience before the system's response appeared on the projected screen. Proxy processed the statement as a direct command rather than rhetorical commentary. Within moments, the system had generated two internal tasks assigned directly to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu responded with visible laughter at the development, completed the request for the pricing documentation, and subsequently deleted the instruction to cease speaking. The episode highlighted the ongoing technical challenge of contextual understanding in artificial intelligence systems, which frequently struggle to distinguish between casual conversational remarks and actionable directives in professional settings. It marked a rare instance of a software demonstration successfully managing its own chief executive. Industry observers noted that the incident underscored both the sophistication and the current limitations of natural language processing in enterprise software. QueueZero emphasized that Proxy requires explicit user confirmation before transmitting any external communications or emails to parties outside the organization. Internal task creation, however, operates automatically by default without additional approval gates. Administrators retain the ability to disable automatic task creation entirely, allowing organizations to calibrate the level of automation according to their specific workflow requirements and risk tolerance for autonomous agent behavior. The company stated that these safeguards are designed to prevent unauthorized data exposure while maintaining workflow efficiency. The company priced Proxy at $30 per user per month, positioning it competitively against similar enterprise productivity tools from established vendors. The software will enter public beta next week, with broader general availability expected to follow later this quarter pending feedback from early adopters and integration testing with major calendar platforms.

Z.ai: GLM 5.3 Prime: # QueueZero Launches Proxy, an AI Meeting Assistant That Acts on What It Hears QueueZero on Wednesday launched Proxy, an artificial-intelligence assistant designed to attend meetings, produce notes and assign follow-up work to software agents. The launch demonstration included an unscheduled test of the system's judgment. During the presentation, a customer in the audience said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was monitoring the event as it would an ordinary meeting, treated the remark as an instruction. It created two tasks and assigned both to Evan Wu, QueueZero's chief executive: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task — the pricing sheet was sent — and deleted the first. He did not, in the end, stop talking. The exchange was brief and good-natured, but it illustrated the design question at the center of Proxy's pitch. Meeting assistants have become one of the most crowded categories in workplace software, with tools that transcribe conversations, summarize them and extract action items. Proxy's differentiator is the final step: rather than leaving action items in a document, it hands them to software agents as assignments. That requires deciding, in real time, which statements in a conversation count as instructions — a judgment people make without thinking, and one that software does not always make the same way. QueueZero says Proxy includes guardrails, though they are unevenly distributed. The assistant requires confirmation before an external message is sent, meaning nothing reaches customers, partners or other outside parties without a person approving it first. Internal tasks face no such barrier: they can be created automatically, which is how the two items assigned to Wu appeared without anyone signing off. The distinction reflects the differing costs of error — an errant external message can leak information or confuse a customer, while an unwanted internal task is, at worst, clutter. Administrators who want tighter control can disable automatic task creation entirely. By those rules, Wednesday's incident carried little risk. Both tasks were internal, and the only one with an outside consequence — the pricing sheet — was completed by Wu himself. The episode, at minimum, demonstrated the behavior Proxy is selling: the assistant heard an unscripted remark, parsed it as a request and assigned it to a named owner, exactly as designed. Proxy costs $30 per user per month and enters public beta next week.

Mistral: Mistral Large 4: # 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.

Mistral: Mistral Large 4: # The Cost of Perfection: Why the Priciest Support AI Isn't Worth the Premium When a software company set out to find the best AI for customer support, the results defied the usual assumption that higher price tags buy better answers. In a blind benchmark of five language models handling 200 real support tickets, the most expensive option delivered virtually identical quality to a model costing less than half as much—while running slower. Model North topped the leaderboard with an average score of 84.2 out of 100, but Model Vale followed closely at 83.8. Statistically, the two are indistinguishable; researchers found no significant difference between their performances. Yet the financial gap is stark: North costs an estimated $1.90 per 1,000 tickets processed, while Vale costs just $0.74. To make matters more lopsided, Vale generated responses 18% faster than its pricier rival. For a support operation handling thousands of tickets monthly, that combination of parity in quality, lower cost, and greater speed represents a compelling value proposition. **How the Test Worked** The benchmark was designed to mimic real-world support conditions while minimizing evaluator bias. Researchers fed 200 anonymized customer-support tickets—drawn from actual English-language inquiries at a single software company—to each of the five models. Every model produced one proposed answer per ticket, which three experienced support managers then graded blindly, without knowing which AI had generated the text. This blind review process prevents the halo effect where graders might unconsciously favor a well-known or expensive brand. The grading rubric combined three dimensions: factual correctness, practical usefulness, and tone appropriateness. This composite score aimed to capture not just whether the answer was factually accurate, but whether a human agent could actually send it to a customer without heavy editing or risk of embarrassment. **The Sweet Spot in the Middle** The results reveal a clear value hierarchy that challenges procurement instincts. While North and Vale occupy the quality tier, three other models—Ember (79.1), Kite (76.4), and Harbor (71.0)—offer progressively cheaper alternatives at $0.31, $0.22, and $0.08 per 1,000 tickets, respectively. The question for support leaders becomes where quality degradation becomes unacceptable. Notably, the top two models excel in different domains. North performed best on billing disputes, suggesting particular strength with structured, policy-heavy inquiries where precision matters. Vale dominated technical troubleshooting, indicating superior handling of complex, procedural problem-solving. This specialization implies that sophisticated deployments might route tickets by category—sending billing issues to North and technical problems to Vale—potentially outperforming any single model used universally. **The Budget Tier Trade-off** The three lower-cost models present a harder calculus. Harbor, at eight cents per 1,000 tickets, is roughly 24 times cheaper than North, but its 71.0 average score suggests a meaningful drop in answer quality. For high-volume, low-complexity deflection—password resets, basic FAQs, order status checks—this trade-off might be acceptable. For nuanced technical issues or sensitive billing questions, the savings likely evaporate in the additional human review time required to catch errors. **Important Caveats** Readers should treat these results as a snapshot, not a universal verdict. The test used tickets from only one software company and exclusively in English, limiting generalizability to other industries or global markets. Most strikingly, all five models performed poorly when tickets depended on image attachments—a critical gap for modern support workflows where customers frequently upload screenshots to illustrate technical problems. Researchers explicitly excluded agentic capabilities, tool use (such as checking account balances or resetting passwords programmatically), and multilingual support from the evaluation. These limitations matter because real-world AI support increasingly involves more than text generation. The models tested here produced static answers; they did not navigate backend systems, update tickets automatically, or chain multiple actions together. A model that scores well in a blind text comparison might falter when required to actually resolve a ticket end-to-end. **The Bottom Line** For organizations evaluating AI support tools, the data suggests a counterintuitive strategy: skip the premium tier. Model Vale offers statistically equivalent quality to the top performer at 61% lower cost and faster response times. The decision between North and Vale should likely hinge on specific use cases—billing-heavy queues versus technical troubleshooting—rather than overall scores. Meanwhile, the dramatic price differences among Ember, Kite, and Harbor invite experimentation with tiered deployment, reserving human agents for the complex scenarios where the cheapest models falter. In customer support, as in many technology decisions, the most expensive option is not automatically the best. Sometimes, the smart money goes to the model that knows when to be good enough.

Z.ai: GLM 5.3 Prime: #

Mistral: Mistral Large 4: # Raincheck Raises $8 Million for Flood-Risk Software *The Austin company maps drainage systems and combines them with rainfall forecasts to help local governments decide where to send crews before storms arrive* Raincheck, an Austin-based company that makes flood-risk software for local governments, announced today that it has raised $8 million. The company maps drainage systems and combines those maps with rainfall forecasts to help cities decide where to send maintenance crews before a storm. The seed round was led by Basin Ventures, with participation from existing investor Long Mile. The funding brings Raincheck's total raised to $10.5 million. > "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 of Raincheck The company has published its [methodology](https://example.com/raincheck-methodology) for mapping flood risk. Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. ## The Competition Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Z.ai: GLM 5.3 Prime: # Raincheck Raises $8 Million to Help Cities Get Ahead of Floods *The Austin company's software combines drainage maps with rainfall forecasts to show local governments where to send crews before a storm hits.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised an $8 million seed round. Basin Ventures led the round, which was announced today, with existing investor Long Mile also participating. The funding brings Raincheck's total raised to $10.5 million. The company maps cities' drainage systems and combines those maps with rainfall forecasts, an approach it lays out in [its forecasting methodology](https://example.com/raincheck-methodology). Cities use the software to decide where to send maintenance crews before a storm. 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, which the company declined to identify because some contracts have not been publicly announced. The company's 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.