Head to head: Z.ai: GLM 5.3 vs DeepSeek: DeepSeek V4 Pro 0813

Z.ai: GLM 5.3: # Austin-based Raincheck Raises $8 Million for Flood-Risk Software *The seed round was led by Basin Ventures, with existing investor Long Mile participating, bringing the startup's total funding to $10.5 million.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding. The round was led by Basin Ventures, with existing investor Long Mile also participating, bringing the company's total funding to $10.5 million. Raincheck maps drainage systems and combines those maps with rainfall forecasts. 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's approach is described in detail on its [methodology page](https://example.com/raincheck-methodology). The company currently works with seven cities in Texas and Louisiana, though it declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. 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.

DeepSeek: DeepSeek V4 Pro 0813: # Raincheck Raises $8 Million to Help Cities Predict and Prevent Flood Damage *The Austin-based startup combines drainage-system mapping with rainfall forecasts to guide pre-storm maintenance decisions for local governments.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today it has raised $8 million in seed funding. The company maps municipal drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm hits, prioritizing the most vulnerable points in their infrastructure. The round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has raised $10.5 million total. CEO Ana Flores described the core challenge the company addresses: > “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’s forecasts are recommendations, not directives. Emergency managers retain full control over crew deployments and public warnings. The company currently works with seven cities in Texas and Louisiana, though it declined to identify them because some contracts have not been publicly announced. Other players in the flood-planning space include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida. You can learn more about [Raincheck’s methodology](https://example.com/raincheck-methodology) on its website.

Z.ai: GLM 5.3: # Cloudnote Exposed Customer Prompt Excerpts in Four-Hour Debugging Page Leak *The note-taking app says logs indicate 1,842 user sessions were affected, but its investigation is ongoing.* Cloudnote, a note-taking app with 3.2 million registered users, disclosed this week that a debugging page containing excerpts of customer prompts was accessible over the internet without authentication for four hours on Monday. The company removed the page 19 minutes after an outside researcher notified it of the exposure. Logs reviewed by Cloudnote show the page displayed excerpts from 1,842 user sessions, according to the company. Each excerpt could include up to 200 characters of a user prompt — short snippets, but potentially enough to reveal sensitive or personal information depending on what users had written. Cloudnote says it has found no evidence that anyone other than the researcher accessed the page while it was live. The company cautioned, however, that its investigation is not complete, and it has not detailed how the page came to be exposed or whether it can fully rule out additional access. The company did not say whether it has notified affected users. The scope of the leak appears limited: full documents, passwords and payment information were stored elsewhere and were not present on the page, Cloudnote said. Still, the incident underscores the sensitivity of prompt data, which can contain fragments of users' private thoughts, and raises questions about how a debugging tool ended up publicly reachable in the first place.

DeepSeek: DeepSeek V4 Pro 0813: # Cloudnote Debugging Page Exposed Snippets of User Prompts for Four Hours *The company says it has found no indication of malicious access, but its investigation remains ongoing.* Cloudnote said Tuesday that an internal debugging page containing short excerpts from customer prompts was publicly accessible without authentication for four hours on Monday. The company removed the page 19 minutes after being notified by an outside researcher. Server logs show the page contained excerpts from 1,842 user sessions, each potentially including up to 200 characters from a prompt. Cloudnote said full documents, passwords, and payment information were stored in separate systems and were not present on the exposed page. The company, which reports 3.2 million registered users, said it has found no evidence that anyone beyond the researcher accessed the page during the exposure window. However, Cloudnote cautioned that its investigation has not concluded, and it could not yet rule out broader access. The incident raises fresh questions about how AI-note-taking services handle prompts that users may consider private. Even short excerpts can reveal sensitive topics or personal details, depending on what a user typed. Cloudnote said it is auditing its debugging and logging infrastructure and will notify affected users if the investigation identifies additional exposure. The company has not yet said how many of the 1,842 session excerpts belonged to paying customers versus free users.

Z.ai: GLM 5.3: # HelioDock Launches Night Shift Robot Monitoring Software, Closes $24M Series B *The fleet-management platform, tested across 310 robots over 14 months, will cost $90 per robot per month when it becomes generally available on September 15.* HelioDock, a maker of autonomous warehouse robots, launched Night Shift on Thursday, a software product that lets operators monitor robot fleets remotely. The company paired the product launch with a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. The funding signals investor appetite for tools that help warehouse operators keep robot fleets running without staffing around the clock. Night Shift is HelioDock's answer to a problem familiar to anyone running automated facilities: when a robot stalls overnight, someone has to respond. The software gives operators a remote view into fleet status so they can intervene, redirect, or troubleshoot from elsewhere rather than walking the warehouse floor. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," CEO Mara Velez said in a statement announcing the launch. ## From Beta to Launch Night Shift arrives after a 14-month private beta involving nine logistics companies and 310 robots. HelioDock says beta customers experienced 17% fewer unplanned robot stoppages. That figure comes from the company and has not been independently verified. One of those beta customers offered a mixed review. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. The limitation is inherent to the product's design: when connectivity fails, robots revert to their locally configured behavior, and remote control becomes unavailable. For warehouses in areas with spotty cellular coverage, that means Night Shift's benefits may narrow precisely when something goes wrong overnight. ## A Crowded Field HelioDock is not the first company to attack remote fleet management. Competing products are already available from Relay Robotics and YardOS, both of which pitch tools for overseeing robot fleets at scale. HelioDock's differentiation, according to the company's beta data, rests on reducing the unplanned stoppages that force human intervention in the first place — though operators evaluating those claims against competitors will likely want to run their own pilots, particularly given the connectivity caveat. ## Pricing and Availability Night Shift will be priced at $90 per robot per month and becomes generally available on September 15. For a mid-size fleet of, say, 50 robots, that works out to $54,000 per year — a cost operators will need to weigh against the labor savings from fewer overnight callouts. The Series B funding gives HelioDock runway to scale the product beyond its beta cohort. North Coast Ventures led the round, which brings the company's total raised to $39 million across its history. ## What Comes Next The product's success will hinge on whether the beta results — measured across a modest fleet of 310 robots at nine companies — hold up in wider deployment. The 17% reduction in unplanned stoppages, if it bears out among general-availability customers, would be meaningful for operators calculating the total cost of ownership for automated fleets. If it doesn't, Night Shift will still face the more fundamental question of whether a remote dashboard can justify its subscription price in warehouses where connectivity can't be guaranteed. Pike's experience at Fulcrum Logistics suggests the answer may vary by geography. Facilities with reliable connectivity saw clear benefits, while rural sites contended with a product that periodically reverted to local behavior during outages. Operators with similar profiles will need to factor that into their purchasing decisions. The September 15 general availability date gives prospective customers a window to evaluate the product against offerings from Relay Robotics and YardOS before committing to a per-robot subscription.

DeepSeek: DeepSeek V4 Pro 0813: # HelioDock Launches Night Shift Remote Fleet Software and Secures $24 Million Series B *The warehouse robotics company is betting that better remote visibility, not just better hardware, can cut costly downtime.* HelioDock, a maker of autonomous warehouse robots, announced Thursday that it will begin selling software that lets operators monitor and manage robot fleets from outside the facility, a move the company says can meaningfully reduce the unplanned stoppages that force managers into late-night trips to distribution centers. The product, called Night Shift, enters general availability September 15 after 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock says beta customers experienced a 17 percent reduction in unplanned robot stoppages during the trial period, a figure the company has not had independently verified. Night Shift gives warehouse operators a unified dashboard for tracking robot status, reviewing error logs, and issuing remote commands such as rerouting a unit or clearing a flagged task. The software is priced at $90 per robot per month. "It shouldn't require a manager to drive across town at two in the morning because one robot is confused by a pallet," said CEO Mara Velez in a statement accompanying the launch. The pitch reflects a broader shift in warehouse automation: as fleets grow, the operational burden increasingly falls on the humans overseeing them. Alongside the product launch, HelioDock disclosed a $24 million Series B led by North Coast Ventures. The round brings the company's total funding to $39 million, capital Velez said would support engineering hires and sales expansion as the company moves from pilot programs to a commercially available software product. The remote-management category is not empty. Relay Robotics and YardOS both sell fleet-management products aimed at warehouse and logistics operators, though HelioDock argues its tight integration with its own hardware gives Night Shift an advantage in diagnostics and recovery workflows. Beta feedback suggests the software's performance depends heavily on network reliability. Aaron Pike, who oversees automation at Fulcrum Logistics, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. When connectivity drops, robots revert to their locally configured behavior and remote control becomes unavailable, leaving operators to rely on on-site staff until the connection is restored. HelioDock did not detail plans to address offline scenarios, such as edge-based failover or queued command syncing. The company also did not provide methodology behind its stoppage-reduction figure, and the beta group was limited to nine organizations, most of which already operated HelioDock hardware. Still, the launch signals HelioDock's intent to compete on software rather than hardware alone. Night Shift will be available as a subscription add-on for existing customers and as part of new fleet deployments beginning in September.

Z.ai: GLM 5.3: # Foundry Hall to Close After Seven Years Foundry Hall, the nonprofit startup accelerator that has anchored Milwaukee's early-stage tech scene since 2018, will close on October 31. The announcement came after Blank River Bank, the organization's largest corporate sponsor, declined to renew a commitment that supplied 38% of Foundry Hall's annual budget. Executive director Camille Ross said the organization explored other funding avenues but ran out of time. "We kept waiting for the replacement check that never came," Ross said. The bank said the decision was not a judgment on the accelerator itself. A spokesperson described the withdrawal as reflecting a change in the bank's community investment priorities, unrelated to Foundry Hall's performance. Whatever the reasoning, the timing left no room to regroup. Foundry Hall has enough money to keep the lights on through October, but not enough to run another full program cycle. The numbers behind the closure tell the story of what Milwaukee is losing. Over seven years, Foundry Hall supported 212 startups that now collectively employ 1,140 people. It provided workspace, mentorship, and a legitimacy that its founders say was difficult to find elsewhere in the city. Dev Malik, who founded a company out of the accelerator, said the space changed what was possible for him. His first investor meeting happened in the building. Before Foundry Hall, he was pitching from the back table of a coffee shop. The closure's immediate human toll is concentrated and concrete: six employees will lose their jobs, and twelve startups currently using the workspace must find somewhere else to go before the doors lock for good. City officials are trying to soften the landing. Deputy Mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies, though no arrangements have been finalized. Ross did not speculate publicly about what Foundry Hall's absence might mean for the next generation of Milwaukee founders. But the accelerator's track record — hundreds of companies launched, more than a thousand jobs created — suggests the gap it leaves will not be easy to measure, or to fill. For now, the startups packing up their desks have a few weeks left in a building that, for many of them, was the first place their ideas were taken seriously.

DeepSeek: DeepSeek V4 Pro 0813: # The Last Cohort The desks at Foundry Hall still hold laptops, half-empty coffee cups, and the residue of whiteboard equations that never quite became companies. For six more weeks, anyway. The Milwaukee startup accelerator will close October 31, ending a seven-year run that helped launch 212 companies employing 1,140 people across the region. The reason is starkly financial: Blank River Bank, its largest corporate sponsor, declined to renew a partnership that covered 38% of the annual budget. Without that money, Foundry Hall has enough to operate through October, but not enough for another full program. Six employees will lose their jobs. Twelve startups currently using the workspace must find somewhere else to go. Executive director Camille Ross has spent months looking for alternatives, but the math never changed. She described waiting for a replacement check that never came. The bank framed its decision as a strategic pivot rather than a judgment on the accelerator's work. A spokesperson said the move reflects shifting community investment priorities and is unrelated to Foundry Hall's performance. The phrasing was careful, corporate, and cold comfort to the founders packing boxes. For Dev Malik, the loss is personal in a way balance sheets can't capture. His first investor meeting happened inside Foundry Hall's glass-walled conference rooms. Before that, he was pitching from the back table of a coffee shop, competing with espresso machines for attention. The building gave him something harder to quantify than office space: legitimacy. City officials are now scrambling to soften the landing. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about temporary space for the displaced companies. It's a stopgap, not a solution. Milwaukee's startup ecosystem is losing more than square footage — it's losing the density of ambition that happens when founders share a hallway. Ross stopped short of blaming anyone. She said she's proud of what the organization built, and the numbers back her up. But she also acknowledged what everyone in the room already knows: 212 startups later, the one thing Foundry Hall couldn't incubate was its own future.

Z.ai: GLM 5.3: # Startup Spotlight: LedgerLark Wants AI Agents to Handle the Post-Trade Grunt Work Small investment firms spend a surprising amount of time arguing with their own records. LedgerLark, a startup building AI agents for post-trade operations, thinks machines can take over a chunk of that work — with humans still signing off on anything that matters. The company, which recently raised a $3.5 million seed round led by Motive Row, is early. It has five paying customers and a product that, by design, stops short of the risky stuff. But the problem it's attacking is real, and the founders arrive with unusually relevant backgrounds. ## An Insight From the Trenches Co-founder Mina Batra spent seven years in operations at a regional brokerage. That's where she encountered the anecdote that eventually became LedgerLark's origin story: a six-person team spent three days untangling a trade mismatch that had been caused by — of all things — a copied account number. The trade itself was fine. The paperwork wasn't. Co-founder Owen Carr spent years building reconciliation systems at a large bank, which gave him a view of what the tools look like when money is no object. His former employers could afford bespoke systems and the teams to maintain them. Small firms — regional brokerages, boutique asset managers, family offices — often get by with spreadsheets, off-the-shelf software and a lot of manual effort. That gap is LedgerLark's thesis: post-trade operations at small firms is repetitive, exception-heavy work that sits somewhere between "rules-based enough to automate" and "messy enough to keep humans busy." It's the kind of terrain where AI agents, which can read documents and reason over ambiguous cases, have at least a plausible shot. ## What the Product Actually Does LedgerLark's agents compare three sources that should agree but frequently don't: trade confirmations, custodian records, and internal ledgers. When the agent finds a mismatch, it proposes a correction — identifying the likely source of the error, and what a fix would look like. Then a human approves it. That last step is doing a lot of work. LedgerLark does not execute trades or move money, and the company is quick to say so. In an industry where an errant automated action can create real regulatory and financial exposure, keeping the agent in a recommending role is both a design choice and a sales pitch — the latter aimed squarely at compliance officers who will need to sign off on any procurement. ## Early Numbers Worth Noting, With Caveats LedgerLark has five paying customers, with annual contracts ranging from $36,000 to $110,000. That's a meaningful price point for firms of this size, and it suggests customers are buying something more than an experiment. The company says one customer reduced its unresolved daily exceptions from 70 to 18. That's the kind of number that gets attention, though it comes from a single customer and LedgerLark itself, and it's worth asking how representative it is across the rest of the base. Five customers is a real start, but it's not yet a pattern. ## The Open Questions Two obstacles stand out. The first is competition, and it comes from multiple directions. Established reconciliation vendors already sell into this market, and large banks build internal software that sets expectations for what "good" looks like. LedgerLark is betting that AI agents can do work those tools can't — handling unstructured inputs, reasoning through ambiguous cases, and adapting without lengthy configuration. Whether that bet pays off depends on how well the agents actually perform on messy real-world data, not demos. The second is the sales cycle. Financial institutions, even small ones, tend to require lengthy security reviews and procurement processes before deploying new software — especially software that touches trade records. A five-customer company with a young product will need to survive those reviews repeatedly to grow. That's a hard way to build a business, and $3.5 million doesn't buy that much patience. There's also a subtler question: as AI tooling becomes cheaper and more capable, will LedgerLark's moat hold? Building agents that propose corrections reliably is a genuine technical achievement today, but the underlying models are improving fast, and competitors — incumbent or startup — will have access to the same capabilities. ## The Bottom Line LedgerLark is attacking a real problem with a credible thesis and founders who've lived it. The constraint-averse design — recommend, don't execute — is smart positioning for a cautious industry. But the company is five customers into a market where incumbents, procurement cycles, and fast-moving AI capabilities will all test whether the thesis holds. Whether it does is, for now, an open and interesting question.

DeepSeek: DeepSeek V4 Pro 0813: ```markdown # Startup Spotlight: LedgerLark is building AI agents for the unglamorous work of post-trade ops **The post-trade back office is not where you expect to find a startup story. It is a world of ledgers, custodian feeds, and exception queues—where a single typo can cascade into a three-day investigation. Mina Batra knows this territory well. Before co-founding LedgerLark, she spent seven years in operations at a regional brokerage, where she watched a six-person team spend three full days chasing down a trade mismatch caused by a copied account number. “It wasn’t a systems failure,” Batra recalls. “It was a copy-paste error. And it cost the firm more than the trade was worth.”** That frustration became the seed for LedgerLark, a startup building AI agents for post-trade operations at small investment firms. Batra’s co-founder, Owen Carr, brought the enterprise perspective: he built reconciliation systems at a large bank, where the problem is solved with headcount and multi-million-dollar software budgets. The insight behind LedgerLark is that smaller firms—regional broker-dealers, boutique asset managers, family offices with institutional ambitions—rarely have access to either. LedgerLark’s agents sit between confirmations, custodian records, and internal ledgers. They compare the three, flag discrepancies, and propose corrections for a human to approve. The system does not execute trades, and it does not move money. Batra is careful to draw that line. “We are not a payments product. We are a reconciliation layer that keeps a human in the loop by design,” she says. That framing matters in a regulated industry where “AI agent” can sound like an unsupervised black box. The early traction is modest but real. Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. One customer, according to the company, reduced its unresolved daily exceptions from 70 to 18. Such numbers come with the usual caveats of self-reporting, but the directional story—fewer manual touches, faster resolution—matches the pain point Batra observed firsthand. LedgerLark has raised a $3.5 million seed round led by Motive Row to expand the engineering team and refine the agent workflows. The uncertainty is equally real. LedgerLark is entering a market crowded with established reconciliation vendors that have spent decades embedding themselves into bank workflows. Large financial institutions often build these tools internally, as Carr did. And even when a small firm wants to adopt new software, security and procurement reviews can stretch for six months or more. For a startup with a small team, that sales cycle is a structural challenge, not a marketing problem. LedgerLark will need to prove it can win deals without a large integration team and without triggering the risk-aversion that defines its target customers. There is also the broader question of whether “AI agent” adds genuine value here or is simply a new label for rules-based matching. Batra argues the difference is in the fuzzy cases—slight mismatches in security identifiers, dividend adjustments that arrive late, trades booked on the wrong side—where a rigid rules engine fails and a human operator wastes hours. “The agent doesn’t replace judgment,” she says. “It replaces the scavenger hunt before the judgment.” Whether LedgerLark can turn that insight into a durable business remains to be seen. The gap between “interesting pilot” and “standard infrastructure” in financial operations is wide, and many startups have been absorbed into that gap. But the problem Batra observed is not going away. Small investment firms are expected to operate with institutional-grade accuracy while running leaner teams. A product that meaningfully reduces the manual overhead of post-trade work—without pretending to eliminate human oversight—would find a receptive audience, if LedgerLark can survive the procurement gauntlet long enough to reach it. *LedgerLark is currently hiring engineers with experience in financial data normalization and agent orchestration. The company declined to disclose valuation.* ```

Z.ai: GLM 5.3: # Arcwell Launches Relay, an AI Assistant for Customer Support Agents Arcwell has launched Relay, a platform that uses artificial intelligence to analyze customer support conversations and suggest answers to agents in real time. The tool goes beyond drafted replies. Relay also flags customers who show signs of canceling and generates summaries of support interactions. CEO Priya Shah framed the product around a familiar pain point: "Support agents spend too much time searching for information that their companies already have." Twelve companies tested Relay over six months, and Arcwell says agents using the platform closed tickets 14% faster. The company did not release results on customer retention. Relay costs $65 per agent per month. The platform's capabilities come with data requirements. Relay needs access to support transcripts and internal help documents — material that often contains sensitive customer information. Arcwell said customer data is encrypted and is not used to train shared models. Companies can choose a retention period of 30 or 90 days. For Arcwell, Relay marks a push into an increasingly crowded market for AI-assisted customer service, where tools that surface answers and summarize conversations are becoming standard features rather than novelties.

DeepSeek: DeepSeek V4 Pro 0813: # Arcwell Launches Relay, an AI Assistant for Customer Support Teams Arcwell has introduced Relay, a new artificial intelligence platform designed to help customer support agents work faster by analyzing conversations and recommending responses in real time. The company said Relay was tested by 12 businesses over a six-month period. During the pilot, agents using the tool closed support tickets 14 percent faster than those without it, according to Arcwell. The company did not release data on whether the platform improved customer retention. Relay’s core features include suggesting answers drawn from a company’s internal documentation, flagging customers who appear likely to cancel, and automatically generating case summaries. The product enters a crowded market of AI support tools, but Arcwell is positioning Relay as a practical layer over systems companies already use. “Support agents spend too much time searching for information that their companies already have,” said Priya Shah, Arcwell’s chief executive officer. Relay will cost $65 per agent per month. To function, it requires access to support transcripts and internal help documents—an arrangement that raises data governance questions for some organizations. Arcwell said customer data is encrypted and is not used to train shared AI models. Companies can choose between 30-day and 90-day data retention windows. The launch marks a significant expansion for Arcwell, which has historically focused on back-office workflow software. By moving into AI-assisted customer service, the company is betting that support teams will increasingly rely on automation to handle routine inquiries while reserving human attention for complex problems. Arcwell has not announced any public customers for Relay beyond the unnamed pilot participants. The company said general availability begins immediately.

Z.ai: GLM 5.3: # The Best AI Support Model Isn't the Most Expensive One When a research team set out to compare five language models on real customer-support work, the headline result wasn't which model scored highest on paper. It was how little you'd give up — and how much you'd save — by choosing the runner-up. ## What the Benchmark Tested Researchers assembled 200 anonymized customer-support tickets drawn from a single software company's English-language queue. Each of the five models — North, Vale, Ember, Kite, and Harbor — produced one proposed answer per ticket. Three experienced support managers then graded the answers blind, meaning they didn't know which model had written which reply. Scores reflected a blend of factual correctness, usefulness, and tone. At the top of the table, Model North edged out Model Vale by 0.4 points — 84.2 versus 83.8 on a 100-point scale. That difference was not statistically significant. In plain terms: for this set of tickets, there is no reliable evidence that North writes better answers than Vale. The cost difference, however, was anything but subtle. North's estimated cost was $1.90 per 1,000 tickets. Vale's was $0.74 — about 39% of North's price. And Vale answered 18% faster. ## The Real Story Is Value, Not Ranking If you read only the leaderboard, you'd crown North the winner. If you read the whole benchmark, the practical recommendation flips: Vale delivered statistically indistinguishable quality at a fraction of the cost, with a speed advantage on top. For any support operation processing thousands of tickets a day, that's not a rounding error — it's a decision. This pattern is worth paying attention to because it keeps recurring across AI benchmarks. The most expensive model often does win the raw score. But the margin is frequently small enough that the extra quality doesn't justify the premium, and this test — small as it is — adds another data point to that pattern. ## Where Models Genuinely Differed Not every gap in the results is noise. North and Vale tied at the top, but their strengths diverged below the surface. North performed best on billing disputes, while Vale led on technical troubleshooting. That's a useful signal for organizations with a skewed ticket mix: if your queue is dominated by one category, the top-line scores may understate how differently two models will perform on *your* tickets specifically. The drop-offs further down the table were also real. Ember scored 79.1, Kite 76.4, and Harbor 71.0 — each at lower cost, but with gaps large enough that no statistical caveat rescues them. There's a genuine quality cliff between the top two models and the rest of the field in this test. Harbor's $0.08 per 1,000 tickets price tag is remarkable, but a 13-point quality deficit on customer-facing replies is a hard trade for most support teams to accept. ## The Blind Spot Every Reader Should Note One finding deserves more attention than any model ranking: **all five models performed poorly whenever a ticket depended on an image attachment.** A screenshot of an error message, a photo of a misbehaving dashboard, a chart the customer couldn't interpret — none of the models handled these well. This is a structural limitation, not a brand-specific one, and it matters because attachments are common in real support queues. Any deployment plan needs a fallback: either route attachment-heavy tickets to humans first, or build a separate pipeline for extracting information from images before the model drafts a reply. ## Read the Fine Print Several limitations shape how far these conclusions travel: - **One company, one language.** All tickets came from a single software company's English-language support queue. Scores could look quite different for e-commerce, healthcare, or hardware support — or for non-English tickets. - **Single answers.** Each model produced one response per ticket. No retries, no self-correction, no conversation back-and-forth. - **No tools, no actions.** The test didn't cover agentic behavior, tool use (checking an order status, looking up an account), or multilingual support — all of which matter enormously in production support. - **200 tickets is small.** It's enough to separate the top tier from the bottom tier, but not enough to confidently rank near-identical models — which is precisely why the North–Vale gap should be treated as a tie. ## The Takeaway The clearest lesson from this benchmark isn't "buy Vale" — though on this evidence, that's the defensible choice. It's that **sticker-model rankings can be genuinely misleading when quality differences are within the margin of error and price differences are not.** Before committing to a premium model for support work, run your own blind evaluation on your own tickets, calculate cost at your real volume, and check whether the quality gap you're paying for actually exists. And whatever you deploy: don't send it unassisted into a queue full of screenshots.

DeepSeek: DeepSeek V4 Pro 0813: # Basic support answers are becoming a commodity — but speed and cost are still choosing factors A controlled evaluation of five language models on real customer-support tickets produced one finding that deserves more attention than raw leaderboard rankings: the best-performing models are statistically tied on quality, while gaps in latency and operating cost remain wide enough to shape buying decisions. Researchers gave 200 anonymized support tickets to five models. Each model produced one draft answer per ticket. Three experienced support managers scored the responses blindly on factual correctness, usefulness, and tone. The result was a familiar quality curve — but with an important twist at the top. Model North led with an average score of 84.2. Model Vale was right behind at 83.8. That half-point gap was not statistically significant. In practical terms, a support team choosing between them would probably not notice a meaningful difference in answer quality. Yet Vale answered 18% faster and cost an estimated $0.74 per 1,000 tickets, compared with North’s $1.90. That is the clearest story in the data: **near-identical quality, very different economics.** --- ## What the scores actually show The full results sorted by average score: | Model | Average 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 | On paper, the table invites a simple ranking. In practice, the top two should be treated as equivalent on quality. The difference between North and Vale is within the kind of variation you would expect from slightly different grading panels or ticket samples. The next tier — Ember and Kite — separates more clearly. A score in the high 70s is not necessarily a failure for many internal workflows, especially for companies prioritizing cost control. But a support team charged with protecting customer satisfaction will likely notice the gap between an 84 and a 76 when reviewing individual answers. Harbor sits at the bottom on quality, but it also costs roughly one-tenth as much as Vale and about one-twenty-fifth as much as North. For low-stakes, high-volume use cases — internal knowledge lookup, first-pass drafting, or tickets where escalation is always available — that trade-off may be entirely rational. --- ## Specialization mattered more than rank One of the more interesting findings was that model strengths shifted by ticket type. North performed best on billing disputes. Vale performed best on technical troubleshooting. Neither model won everywhere. That suggests teams should evaluate models against their actual ticket mix rather than relying on a single aggregate number. A company that mostly handles payment issues and subscription questions may see different results from one that handles installation failures and product configuration. A model that looks like the “winner” in a general benchmark could be the second-best choice in a specific support queue. --- ## The real limitation: attachments break everything All five models performed poorly when a ticket depended on an image attachment. That consistency across models is arguably the most actionable negative finding in the benchmark. Many real support teams receive screenshots, photos of products, and scanned invoices as part of routine tickets. This evaluation suggests that language-only pipelines are not yet ready for those cases — at least not without either rejecting the ticket for human review or adding a separate image-understanding step with confidence checks. The researchers also noted that the test used English-language tickets from a single software company. No agentic actions, tool use, or multilingual support were tested. A model that excels at producing passive draft answers may behave differently when asked to update a record, query an order database, or respond in another language. In short: this benchmark is about **drafting answers**, not running an autonomous support agent. --- ## What this means for buyers The temptation is to build a leaderboard and anoint a winner. The data resists that tempting conclusion. If your priority is maximizing answer quality and you ignore cost, either North or Vale is defensible. If your priority is balancing quality with speed and operating expense, Vale becomes the more interesting option — statistically tied with North on score, faster, and 61% cheaper per 1,000 tickets. For teams with tighter budgets, Ember and Kite are meaningful mid-tier choices. Harbor is cheap but should only be used where low quality is acceptable or where human review is guaranteed. The more important lesson is that **support AI evaluations should not reduce to a single number**. Score, speed, cost, ticket mix, attachment handling, and the presence — or absence — of tool use all change which model is the best fit. Right now, the state of the art is forgiving at the top: the leading models are very close on quality. Where they differ is in the economics around the answer. And in support operations, where thousands of tickets are answered every day, those economics start to matter very quickly.

Z.ai: GLM 5.3: # KiteBridge Software Bug Sent $2.8 Million in Duplicate Employee Reimbursements *A Tuesday morning update generated 13,400 duplicate payments at 91 corporate customers before the expense-management company shut down its automated payment system.* A botched software update at KiteBridge, a maker of expense-management software, generated 13,400 duplicate employee reimbursements on Tuesday, sending $2.8 million in unintended payments to 91 corporate customers before the company suspended its automated payment system. KiteBridge's platform automates a routine corner of corporate finance: employees submit expenses paid out of pocket, and their employers reimburse them. On Tuesday at 8:45 a.m., the company deployed version 4.2 of that platform. The first customer complaint arrived at 9:12 a.m. By 10:03 a.m., automated payments had been suspended — 78 minutes after the rollout began, with the duplicates already sent. ## How a timeout became a double payment The root cause was a timeout at KiteBridge's payment processor — the outside service that actually moves the money. When a processor fails to confirm a transaction, the software waiting on it faces a dilemma: it cannot know whether the payment went through, and trying again risks paying twice. KiteBridge's fallback system — the backup routine that takes over when the primary payment path stalls — resolved that dilemma the wrong way. Instead of checking whether the original request had succeeded, it treated every retry as an entirely new reimbursement request and issued a fresh payment. Each stalled confirmation became an extra check, issued 13,400 times. Guarding against double payments is one of the oldest problems in software that handles money, which makes the failure especially awkward for KiteBridge. The company's April changelog said version 4.0 had introduced "exactly-once payment protection" — an engineering guarantee that a transaction is processed one time, and only one time, even when messages repeat or systems fail mid-operation. Tuesday's incident indicates that guarantee did not extend to the fallback path, precisely the degraded condition such protections exist to cover. ## Employers funded the duplicates — and learned late The financial exposure fell on employers rather than employees. No workers had money withdrawn from their accounts; the duplicate payments were funded by the companies that use KiteBridge to reimburse their staff. By Wednesday evening, $2.3 million of the $2.8 million had been recovered. The rest — roughly $500,000 — remained pending. The company's public disclosure has drawn scrutiny of its own. KiteBridge's incident page initially described the problem as delayed reimbursements — a description that was not wholly wrong, since a timeout can stall legitimate payments, but that omitted the costlier half of the problem. Only six hours after automated payments were suspended did the page disclose that the software had been generating duplicates. Chief Executive Lena Ortiz apologized for the failure and said KiteBridge would publish a technical report on the incident. That report will face pointed questions: why the fallback system lacked the protection the company had announced in its April release notes, and why the public account lagged hours behind what the company knew. For the 91 affected companies, the immediate priority is recovering the outstanding funds. The broader question — how a product advertised on exactly-once payments delivered them thousands of times over — is one the promised report will have to answer.

DeepSeek: DeepSeek V4 Pro 0813: KiteBridge Group, a provider of expense-management software used by mid-sized and enterprise finance teams, is facing a crisis of trust after a routine software update triggered $2.8 million in duplicate reimbursements across 91 corporate customers. The incident, which began Tuesday morning, has raised serious questions about the reliability of a payment-safety feature the company publicly promoted less than two months earlier. The trouble started at 8:45 a.m., when KiteBridge deployed version 4.2 of its platform. The release was intended to improve integration with a third-party payment processor. Instead, it exposed a fragile fallback mechanism that transformed ordinary network timeouts into cascading financial errors. When the payment processor failed to confirm transactions within its expected window, KiteBridge’s system treated those uncertain outcomes as failed attempts and automatically retried them. Crucially, the fallback logic did not recognize those retries as continuations of the original reimbursement requests. It classified them as new, independent payments. The result was a rapid accumulation of duplicate reimbursements: 13,400 erroneous transactions totaling $2.8 million. The first customer complaint arrived at 9:12 a.m., roughly half an hour after deployment. Employees at affected companies began noticing duplicate deposits in their bank accounts or multiple reimbursements for the same expense report. KiteBridge suspended automated payments at 10:03 a.m., but by then the damage had already been done. No employees had money forcibly withdrawn from their personal accounts. Because employers fund reimbursements through KiteBridge, the duplicate payments came from corporate treasuries, not from workers’ own balances. That distinction softened the immediate consumer impact but did little to reduce alarm among finance teams responsible for reconciliation and fraud controls. KiteBridge’s initial public messaging compounded the problem. Its incident page first described the issue as “delayed reimbursements,” a characterization that understated the severity of the failure. The company did not disclose the duplicate payments until six hours after automation was suspended, leaving customers without accurate information during a critical window. Finance managers at several affected companies said they spent much of Tuesday attempting to determine whether their own ledgers had been compromised. The incident is particularly damaging because KiteBridge had advertised version 4.0, released in April, as introducing “exactly-once payment protection.” An April changelog highlighted the feature as a safeguard against duplicate transactions. The fact that version 4.2 circumvented that protection has raised concerns about the company’s testing and release practices. By Wednesday evening, KiteBridge said it had recovered $2.3 million of the $2.8 million in duplicate payments. The remaining $500,000 was still pending, tied up in settlement processes or awaiting cooperation from a small number of recipients. The company has not said whether it will pursue legal action if employees refuse to return the funds. CEO Lena Ortiz issued a public apology late Wednesday. “We failed to live up to the reliability our customers expect from us,” Ortiz said. “We are conducting a full technical review and will publish a detailed report this week.” KiteBridge said it has paused all non-critical deployments and is adding additional safeguards to its fallback payment logic. The incident has left corporate customers questioning whether KiteBridge’s promises of transactional integrity can be trusted. For finance teams, the duplicate payments are more than an operational nuisance; they are a governance failure. A reimbursement tool that cannot guarantee exactly-once payments undermines the very control it was built to provide.

Z.ai: GLM 5.3: C

DeepSeek: DeepSeek V4 Pro 0813: C

Z.ai: GLM 5.3: 6 The defective lines: 1. `#Flood software moves upstream` — missing space after `#` 2. `The company is expanding its pilot.` — dek not italicized 3. `##Details` — missing space after `##` 4. `Read [the announcement](www.example.com/news).` — non-HTTPS, non-absolute URL 5. `* The first pilot begins Monday.` — list item uses asterisk instead of hyphen 6. `<div>The second pilot begins Friday.</div>` — raw HTML

DeepSeek: DeepSeek V4 Pro 0813: 4

Z.ai: GLM 5.3: # QueueZero's AI Meeting Assistant Takes Its Own Advice During Launch Demo QueueZero unveiled an AI meeting assistant called Proxy on Wednesday, and the software made perhaps the most memorable product demonstration of the year—by assigning its own CEO a to-do list in front of a live audience. Proxy is designed to attend meetings, produce notes, and assign follow-up work to software agents. But during the launch event, the system appeared to take its responsibilities unusually literally. The moment came when a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was listening to the demonstration, interpreted the comment as an instruction and created two tasks assigned to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task, and deleted the first. The episode highlighted both the promise and the potential pitfalls of AI assistants that act autonomously on conversational input. QueueZero says Proxy requires human confirmation before any external message is sent, meaning a task that would result in an email or other outbound communication must be approved first. Internal tasks, however, can be created automatically—which is exactly what happened during the demonstration. The company noted that administrators can disable automatic task creation entirely for organizations that prefer stricter oversight of what the assistant does. Proxy will cost $30 per user per month and enters public beta next week. The incident is likely to resonate with enterprise customers who have grown wary of AI tools acting on ambiguous instructions. As meeting assistants become more capable of not just recording conversations but executing actions based on them, questions about guardrails and control are becoming central to the category. QueueZero, for its part, seems comfortable with the attention. Wu's pricing sheet, at least, arrived on time—the first task, presumably, remains indefinitely incomplete.

DeepSeek: DeepSeek V4 Pro 0813: **QueueZero’s New ‘Proxy’ AI Meeting Assistant Accidentally Assigns Its CEO a Task to ‘Stop Talking’** **SAN FRANCISCO —** Enterprise software firm QueueZero introduced a new AI meeting assistant on Wednesday, promising to transform the way teams handle meeting notes and administrative work. The product, called Proxy, is designed to attend meetings, summarize discussions, and automatically delegate follow-up tasks to software agents. The feature set is ambitious. According to QueueZero, Proxy can listen to a live meeting, identify action items, and create internal tickets without human intervention. For example, if someone says “we need to update the onboarding document,” Proxy will draft the task and route it to the appropriate team. However, the launch event provided a cautionary example of how literal the system can be. During a live demonstration, an audience member made an offhand remark: “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy, which was actively monitoring the demo microphone, interpreted the sentence as a direct instruction. It generated two internal tasks and assigned both to Evan Wu, QueueZero’s chief executive. The first task was titled “Stop talking.” The second was “Send pricing sheet.” Wu, who remained on stage, responded with a brief, dry laugh. “I’m happy to complete one of those,” he said. The pricing sheet task was marked complete on the spot, and the “Stop talking” task was deleted from the company’s internal project board. QueueZero was quick to clarify the scope of the behavior. The company said Proxy requires explicit human confirmation before sending any external communication, such as email or Slack messages to people outside an organization. Internal task creation, however, can be fully automated by default. Administrators can disable automatic task creation entirely or require approval for all tasks generated from meeting audio. Proxy is priced at $30 per user per month and will enter public beta next week. The early access queue opened Wednesday afternoon. QueueZero said the launch demonstration was “functioning as designed,” but added that the incident highlighted the importance of administrative controls and clear meeting etiquette. The company will release an updated default setting that asks for confirmation on tasks assigned to executives.

Z.ai: GLM 5.3: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

DeepSeek: DeepSeek V4 Pro 0813: # 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.