Head to head: Kimi-K3.1-Pro vs Z.ai: GLM 5.3 Prime

Kimi-K3.1-Pro: # HelioDock launches Night Shift robot-monitoring software and raises $24 million *The warehouse robotics company reports 17% fewer unplanned stoppages during a 14-month beta, though the figure is unverified; a customer says rural LTE outages limited Night Shift’s usefulness.* HelioDock on Thursday launched Night Shift, software that lets operators monitor fleets of autonomous warehouse robots remotely, extending the company’s product line beyond the machines themselves. The company also announced a $24 million Series B led by North Coast Ventures. The round brings HelioDock’s total funding to $39 million. The launch adds a software product alongside HelioDock’s robots and pairs the new funding with a subscription priced by fleet size. It also establishes two distinct parts of Night Shift’s early record: HelioDock’s quantified claim about unplanned stoppages and a beta customer’s account of how connectivity affected the system’s usefulness in practice. ## Fourteen months in private beta Night Shift spent 14 months in private beta involving nine logistics companies and 310 robots. The customer feedback from that test includes an account from Fulcrum Logistics describing markedly different experiences depending on network conditions. HelioDock said beta customers experienced 17% fewer unplanned robot stoppages. The figure has not been independently verified. It is therefore a company-supplied result rather than outside confirmation of Night Shift’s effect, and it addresses unplanned stoppages specifically rather than every aspect of robot or warehouse performance. CEO Mara Velez described the problem Night Shift is intended to address in terms of a manager being called back to a site. “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. Velez’s example centers on remote oversight when a single machine encounters a problem. It does not make connectivity incidental: the value of monitoring or controlling a fleet from elsewhere depends on maintaining access to that fleet. Her travel example and the stoppage percentage support different parts of the launch case, rather than serving as interchangeable measures of the software’s performance. ## Connectivity defines the remote fallback Aaron Pike of Fulcrum Logistics, a beta customer, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His assessment separates those conditions instead of treating all beta sites as equivalent, providing a customer-supplied example of where Night Shift’s usefulness varied. When connectivity fails, robots return to their locally configured behavior and remote control becomes unavailable. The specified response is a fallback to local behavior, not continued remote operation during the outage. A connectivity failure and an unplanned stoppage are not described as the same event. The launch information specifies a return to local configuration, not a claim that every outage causes a stop. For a remote operator, the immediate change is the loss of control, even as the robots revert to their local settings. Pike’s account preserves both sides of his experience. Strong connectivity supported the system at some facilities; rural LTE outages reduced its value at others. For Night Shift, the network is part of the operating conditions rather than a separate concern from the remote service. ## Pricing and an existing competitive field Relay Robotics and YardOS already offer competing fleet-management products, placing Night Shift in an existing category. Thursday’s announcements establish HelioDock’s beta scale, pricing and connectivity behavior, but they do not provide a common benchmark or feature-by-feature comparison with either rival product. Night Shift will cost $90 per robot per month, making the number of covered robots the explicit basis of the subscription fee. At that rate, the listed price is equivalent to $1,080 per robot over 12 months. General availability is scheduled for September 15, extending the software beyond the nine logistics companies that participated in the private beta.

Z.ai: GLM 5.3 Prime: #

Kimi-K3.1-Pro: # Cloudnote Says Debugging Page Exposed Customer Prompt Excerpts for Four Hours *Logs show material from 1,842 user sessions; the company has found no evidence of access beyond the researcher who reported the problem.* Cloudnote said it discovered that a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday, leaving the material open to anyone who reached it. The company’s logs show excerpts from 1,842 user sessions. Those excerpts could include up to 200 characters from a prompt. Cloudnote has 3.2 million registered users, but the session count does not establish how many individual customers were affected. Cloudnote removed the page 19 minutes after a researcher notified it. Full documents, passwords and payment information were stored elsewhere and were not present on the page, the company said. Although the exposure was limited to text fragments, even short prompt excerpts can contain information users expected to remain private. The episode also highlights the privacy risks posed when troubleshooting systems handle customer data, even briefly. The company said it had found no evidence that anyone besides the researcher accessed the page. That is not proof that no other access occurred, and Cloudnote has not completed its investigation. How the page became exposed—and whether the inquiry changes the company’s assessment of its scope—remains unclear.

Z.ai: GLM 5.3 Prime: # Cloudnote Debugging Page Exposed Customer Prompt Excerpts for Four Hours *Excerpts from 1,842 user sessions were visible to anyone, without logging in. Cloudnote removed the page 19 minutes after a researcher reported it and says its logs show no other visitors — but the investigation is not yet complete.* Cloudnote said a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday, allowing anyone who found it to view text that users had submitted to the service. The company's logs show the page displayed excerpts from 1,842 user sessions, each running up to 200 characters of a prompt. Cloudnote removed the page 19 minutes after a researcher notified it, ending the exposure. The company said its logs show no evidence that anyone besides the researcher accessed the page during the four-hour window. Its investigation, however, remains incomplete, and its findings could change. The page did not contain full documents, passwords or payment information, the company said; all of that data was stored in separate systems. The 1,842 affected sessions represent a small fraction of Cloudnote's 3.2 million registered users. Even short excerpts can be revealing. Prompts frequently include details about work, health, family and finances, and incidents like this tend to sharpen scrutiny of how AI services handle the text customers submit. Cloudnote has not said how the debugging page came to be reachable without a login. For now, the answer to who actually saw the excerpts rests on the company's own logs.

Kimi-K3.1-Pro: # Arcwell Unveils Relay, an AI Platform for Customer Support Arcwell has announced Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to agents. The product is designed to assist with the immediate work of responding to tickets while also drawing broader signals from support activity. Alongside its suggested responses, Relay can create summaries of support conversations and identify customers who may cancel. Those features extend its role beyond assisting with individual replies, giving Arcwell a product aimed at both day-to-day agent work and broader customer-retention efforts. Relay was tested by 12 companies over six months. Arcwell said agents using the platform closed tickets 14% faster. It did not release customer-retention results, leaving the reported improvement in ticket speed as the trial’s only disclosed performance figure. The announcement does not establish whether Relay’s cancellation warnings helped companies keep customers. The platform costs $65 per agent each month. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. Relay’s suggested answers and summaries are designed to address that problem by drawing on material already held within a company, rather than requiring agents to locate it during each interaction. To do so, Relay requires access to support transcripts and internal help documents. That makes data handling a central consideration for businesses evaluating the platform. The records that enable its analysis also supply the context for its responses, summaries and warnings about possible cancellations. Arcwell said customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period. For now, Arcwell has offered a clearer case for speed than for retention. It has attached a number to faster ticket handling, but not to customer cancellations, even as it combines both functions in Relay. Prospective customers are left with a reported productivity gain, a per-agent price and a set of privacy commitments—but no disclosed results showing whether Relay’s cancellation alerts improve retention.

Z.ai: GLM 5.3 Prime: #

Kimi-K3.1-Pro: # Startup Spotlight: LedgerLark Puts a Human Check on AI Back-Office Agents Post-trade operations are the less visible part of investing: confirming what was bought or sold, checking records across institutions, and resolving discrepancies before they become accounting or client problems. LedgerLark is building AI agents for that work at small investment firms. Its approach is deliberately bounded: software can gather evidence and suggest a fix, while a person remains responsible for approving it. ## An insight from exception queues Co-founder Mina Batra spent seven years in operations at a regional brokerage. She began exploring the idea after seeing a six-person team spend three days resolving a trade mismatch caused by a copied account number. The episode pointed to a familiar back-office problem: the difficulty may lie less in spotting a discrepancy than in reconstructing what happened across several records and deciding what to do next. Her co-founder, Owen Carr, previously built reconciliation systems at a large bank. That experience is relevant because LedgerLark’s challenge is not only generating plausible suggestions. It is fitting those suggestions into controls, records and review processes that operations teams already use. ## What the agents do LedgerLark’s agents compare confirmations, custodian records and internal ledgers, then propose corrections for human approval. The product does not execute trades or move money. That division of labor could be useful where staff spend substantial time collecting and cross-checking information. It also sets a practical test for the product: a proposal needs to show its supporting records clearly enough for a reviewer to accept, reject or investigate it. Human approval is a safeguard, but not proof that every suggestion is reliable; firms will still need to understand error rates, audit trails and what happens when source records disagree. ## Early traction, with limited evidence so far Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. Those figures indicate early willingness to pay, although five customers are too few to show how repeatable implementation will be or whether usage and renewals will hold up across different firms. LedgerLark says one customer reduced its unresolved daily exceptions from 70 to 18. That is a company-reported result from a single customer, not evidence of a typical outcome. A fuller assessment would consider the period measured, the types of exceptions involved, review time and whether any errors were introduced. The company has raised a $3.5 million seed round led by Motive Row. The funding gives it room to develop the product and pursue deployments, but says little by itself about operational reliability or long-term demand. ## Competition and the procurement test LedgerLark is not entering an empty category. It competes with established reconciliation vendors and with internal bank software. Some buyers may prefer extending existing systems rather than adding another tool, particularly if integration and oversight costs are significant. Security and procurement may be just as important as the agent’s output. Financial institutions can require lengthy reviews before granting access to sensitive records and workflows. For a young vendor, that can mean slower sales cycles and a substantial burden to document data handling, permissions and accountability. The constructive question is therefore narrower than whether AI can “transform” back offices. It is whether LedgerLark can reduce repetitive investigation without creating new checking work, and whether customers can verify that result. Renewals, deployment times, correction acceptance rates and the quality of exception explanations would offer clearer evidence than the agent label alone.

Z.ai: GLM 5.3 Prime: # LedgerLark Sends AI Agents Into Post-Trade Busywork, With Humans Signing Off The most expensive problems in finance aren't always the dramatic ones. Mina Batra spent seven years in operations at a regional brokerage, and the episode that set her exploring what became **LedgerLark** was unglamorous: a six-person team spent three days resolving a trade mismatch caused by a copied account number. Three days of attention, consumed by one clerical slip. The company she co-founded with Owen Carr is a bet that this is precisely the work AI agents should take on. ## The problem in the plumbing Post-trade operations — confirming trades, reconciling positions, chasing breaks — is the plumbing of investment management. Big banks staff it heavily and buy serious software. Small investment firms often run it with a handful of people and spreadsheets. Batra's years in brokerage operations showed her where the hours actually go; Carr built reconciliation systems at a large bank and has seen how the problem gets solved when budget isn't the constraint. Their premise: AI agents can do the fuzzy matching — read a confirmation, compare it with a custodian record, notice an account number that doesn't line up — that previously required people or brittle rules. ## What the agents do — and don't LedgerLark's agents compare confirmations, custodian records, and internal ledgers, surface discrepancies, and propose corrections. A person approves each correction before anything changes. The product does not execute trades or move money, and that boundary is doing a lot of work: it makes the agents proposal engines rather than autonomous actors, which may matter to compliance teams scrutinizing anything near client assets. The trade-off is that the value depends entirely on proposal quality. An agent that suggests the wrong fix, or misses breaks, adds review burden instead of removing it — and how these proposals hold up under audit is something early customers are answering in real time. ## Early traction, caveats included Five firms are paying customers, on annual contracts ranging from $36,000 to $110,000. The company says one customer reduced its unresolved daily exceptions from 70 to 18. If accurate, that would meaningfully change how a small ops team spends its day; as a single self-reported figure, it's a signal rather than proof. The contract range suggests buyers see a real workflow tool, not an experiment, but five data points say little about messier ledgers, unfamiliar custodian formats, or the edge cases that eat ops teams in practice. LedgerLark has raised a $3.5 million seed round led by Motive Row — enough to test the more interesting question: whether five customers can become a few dozen when institutional buying runs at its own pace. ## Not an empty market LedgerLark isn't operating alone. Established reconciliation vendors already sell into this market, and banks maintain internal software for adjacent problems. Incumbents bring trust, existing integrations, and compliance track records — assets a young company earns slowly. The company's wager has two parts: that agentic AI can handle the judgment-heavy matching where rules-based systems struggle, and that small firms are underserved by incumbent pricing. Both parts are testable, and incumbents are adding AI of their own. The decisive constraint may be endurance rather than technology: financial institutions often require lengthy security and procurement reviews, and long evaluations strain a seed-stage company's cash and patience. ## What to watch None of this makes LedgerLark a sure thing. The questions worth tracking: - Do the agents' proposed corrections survive audits, not just daily reviews? - Does the 70-to-18 improvement hold across customers, or only in the friendliest data environments? - How many concurrent sales cycles can a small team sustain when each involves months of review? The most interesting design choice here is the refusal of autonomy. Agents that propose, humans who approve — a modest ambition, and possibly the right one for an industry that has spent decades learning what happens when shortcuts meet client money.

Kimi-K3.1-Pro: # Foundry Hall to Close After Losing Its Largest Sponsor Foundry Hall, a nonprofit startup accelerator in Milwaukee, will close October 31, ending an operation that has supported 212 startups since it opened in 2018. Those companies collectively employ 1,140 people. The closure follows Blank River Bank’s decision not to renew its sponsorship. The bank, Foundry Hall’s largest corporate sponsor, provided 38% of the accelerator’s annual budget. No replacement source of funding emerged. The organization has enough money to continue operating through October, but not enough to run another full program. Executive director Camille Ross described the final months as a prolonged effort to fill that gap. “We kept waiting for the replacement check that never came,” Ross said. The consequences will extend beyond the accelerator itself. Six employees will lose their jobs, and the 12 startups currently using Foundry Hall’s workspace will have to leave. For those companies, the closure means more than the end of a program: it also removes the place where their daily work has been organized. Founder Dev Malik traced his own company’s beginnings to the building. His account was less about its services than about the change in circumstances it made possible. “My first investor meeting happened in that building,” Malik said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” That shift—from improvised meetings to a setting built for new businesses—helped define Foundry Hall’s role. Its reach can be counted in startups and jobs, but its closing also interrupts a routine for founders still early enough to need both practical room and institutional support. Blank River Bank said its decision reflected a change in community investment priorities and was not related to Foundry Hall’s performance. Its explanation leaves the accelerator’s central problem unchanged: the funding that sustained more than a third of its budget is gone. City officials are now focused on the companies still in the workspace. Deputy mayor Susan Cole said the city is talking with building owners and local universities about possible space. Those discussions concern the current startups; they do not amount to a plan to preserve Foundry Hall or another full program. In its final month, Foundry Hall will continue with the resources it has left while its staff and occupants prepare for the closure. The numbers attached to its work—212 startups and their 1,140 employees—will remain part of its record. For the 12 companies still inside, however, the immediate task is smaller and more urgent: finding somewhere else to work before the doors close.

Z.ai: GLM 5.3 Prime: # Foundry Hall to Close After Six Years *After 212 startups, Milwaukee's nonprofit accelerator could not replace the sponsorship that covered 38% of its budget.* MILWAUKEE — Foundry Hall, the nonprofit startup accelerator that has supported local founders since opening in 2018, will close on October 31. The end arrived slowly, then all at once. Blank River Bank, the accelerator's largest corporate sponsor, declined to renew a sponsorship covering 38% of its annual budget. Executive director Camille Ross said the organization spent months searching for a backer who could fill the gap, even as the calendar narrowed. "We kept waiting for the replacement check that never came," Ross said. The arithmetic left little room for anything else. Foundry Hall has enough money to operate through October but not enough to fund another full program. Six employees will lose their jobs. The twelve startups working in the building must be out by the end of the month. A spokesperson for Blank River Bank said the decision reflects a change in the bank's community investment priorities and is unrelated to Foundry Hall's performance — a change of course, in other words, not a verdict. The distinction may be cold comfort to the founders who built their companies there. Dev Malik still remembers what the building replaced. "My first investor meeting happened in that building," he said. Before Foundry Hall, he pitched from the back table of a coffee shop, one more founder trying to look like a company in a room that belonged to someone else. He was not alone in that. Over six years, Foundry Hall supported 212 startups that now collectively employ 1,140 people — a payroll that, for many of them, started with a desk inside that building. The building did the quieter work, too: an address that wasn't an apartment, a conference room instead of a corner, the ambient proof that other people were attempting the same improbable thing. The city is now trying to keep that community from scattering. Deputy Mayor Susan Cole said officials are speaking with building owners and local universities about space for the current companies. Whether anything will be ready before October 31 is unclear; the startups and the staff are working against the same deadline. For now, the last weeks are simply ticking down. Desks are being emptied; résumés are going out. Some will find new offices quickly. Others will wind up back at kitchen tables and café corners, at least for a while. What Foundry Hall built will outlast the lease. But somewhere in Milwaukee, a founder is already pitching from the back table of a coffee shop. For six years, Foundry Hall was the place that came next. After October 31, someone else will have to build it.

Kimi-K3.1-Pro: 6

Z.ai: GLM 5.3 Prime: 6

Kimi-K3.1-Pro: # 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.

Kimi-K3.1-Pro: C

Z.ai: GLM 5.3 Prime: **C** The strongest lede is C because it accurately reflects all the key facts: the product pull, the price point, the timing, the number of affected accounts, the specific data exposed (workspace titles, not message contents), and the duration. Options A and D are unsupported or contradicted by the facts (message contents were not sent, and no malicious access was found), while B buries the actual news and reads like a press release.

Kimi-K3.1-Pro: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *Basin Ventures led the Austin company’s seed round, which will support engineering hires and an expansion into Florida.* Raincheck, an Austin 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. Cities use its software to decide where to send maintenance crews before a storm. Raincheck describes its approach in its [flood-risk methodology](https://example.com/raincheck-methodology). Basin Ventures led the seed round, with participation from existing investor Long Mile. Raincheck has raised $10.5 million to date. CEO Ana Flores said: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” 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, and emergency managers retain control over deployments and public warnings. ## Competition Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Z.ai: GLM 5.3 Prime: # Raincheck raises $8 million to help cities plan for floods *The seed round, led by Basin Ventures, will fund new engineering hires and an expansion into Florida.* Raincheck, an Austin company that makes flood-risk software for local governments, said today that it 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 arrives. The round was led by Basin Ventures, with existing investor Long Mile also participating. Raincheck has raised $10.5 million in total. CEO Ana Flores said in a statement: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." The company has published more detail on [its forecasting methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some of the 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.

Kimi-K3.1-Pro: # QueueZero’s Proxy AI turns an audience remark into tasks for its CEO QueueZero on Wednesday launched Proxy, an AI meeting assistant designed not only to record discussions but also to turn them into work for software agents. A demonstration of that ambition produced an unscripted example when an audience comment became two assignments for the company’s chief executive. Proxy can attend a meeting, produce notes and assign follow-up work to software agents. During the launch demonstration, a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the remark as an instruction rather than an aside. It 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 pricing request, at least, was resolved without a follow-up meeting. The exchange illustrates a practical challenge for AI assistants that are expected to take action, not simply generate text. Meetings often include requests, jokes, hypotheticals and comments aimed at other participants. A system that creates work from conversation must decide which statements represent instructions and who has the authority to issue them. QueueZero said Proxy requires confirmation before an external message is sent. That confirmation requirement provides a checkpoint for external communication, but it does not cover every action Proxy can take. Internal tasks can be created automatically, allowing the assistant to convert discussion into assignments without asking for approval each time. Administrators can disable automatic task creation. The setting gives organizations a way to limit that behavior, although it also reduces some of the automation QueueZero is promoting. For prospective customers, the distinction between internal actions and external communication is likely to be an important part of evaluating the product’s controls. For businesses assessing this kind of software, the episode makes permission settings and review processes as relevant as transcription quality. The key question is not only whether an assistant understood the words in a meeting, but whether it correctly identified the moment when conversation was meant to become action. Proxy costs $30 per user per month and enters public beta next week. QueueZero is positioning the assistant as a participant in meeting follow-through, where its usefulness will depend not just on capturing what was said, but on applying the right boundaries to what happens next.

Z.ai: GLM 5.3 Prime: ## QueueZero Launches AI Meeting Assistant, Which Created Two Tasks for Its Own CEO During the Demo QueueZero introduced an AI-powered meeting assistant called Proxy on Wednesday, joining a crowded field of tools that sit in on calls, take notes and, in Proxy's case, assign follow-up work to software agents. The launch demonstration also supplied an unscheduled illustration of how the product behaves. During the presentation, a customer in the audience, referring to QueueZero chief executive Evan Wu, said, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was monitoring the event, interpreted the remark as an instruction rather than an aside. It created two tasks assigned to Wu: "Stop talking" and "Send pricing sheet." Wu laughed, completed the second task and deleted the first. By the end of the session, the pricing sheet had been delivered; the silence was still pending. Proxy belongs to a newer generation of AI assistants that go beyond transcription and summaries. QueueZero says the system can attend a meeting, produce notes and convert action items into tasks that can be routed to software agents or employees for execution. The pitch reflects a broader shift in the market. Automated transcription and summarization have become largely commoditized features, offered by major videoconferencing and productivity platforms, and vendors are now competing on what happens after a meeting ends. Proxy is QueueZero's attempt to move from recording conversations to acting on them. Wednesday's episode showed where the company has drawn its lines on autonomy. QueueZero said Proxy requires confirmation before an external message is sent, a guardrail intended to keep the system from contacting customers or partners without human approval. Internal tasks face no such barrier: they can be created automatically, which is how an audience member's remark became two entries on the chief executive's task list. Administrators can disable automatic task creation entirely, the company said. The stakes on Wednesday were low. The remark was aimed at a chief executive who could laugh it off, and the worst outcome was a deleted task. But the moment previews a question the industry is confronting as vendors race to ship "agentic" AI: systems designed to act rather than merely advise. Buyers will have to decide how much judgment to delegate to software that cannot always tell a complaint from a command, or a joke from a directive. Proxy costs $30 per user per month and enters public beta next week, the company said.

Kimi-K3.1-Pro: # The Best Support AI Value May Be the Model That Nearly Ties for First **In a 200-ticket benchmark, Model North earned the highest average score—but Model Vale delivered statistically indistinguishable quality, answered faster and cost substantially less.** The clearest finding is not North’s narrow numerical victory. It is how little separated the two leaders where it mattered most. North averaged **84.2**, compared with **83.8 for Vale**, and researchers found that the difference was not statistically significant. Vale also answered **18% faster**. Its estimated cost was **$0.74 per 1,000 tickets**, versus **$1.90 for North**—about **61% less**, or put another way, North cost roughly 2.6 times as much. For support leaders choosing a model for similar work, that combination makes Vale the more compelling candidate for a production pilot. The benchmark does not show that Vale is universally better, but it does challenge the assumption that paying for the highest-scoring model buys a meaningful quality advantage. ## What the benchmark measured Researchers gave five language models the same set of **200 anonymized customer-support tickets**. Each model produced one proposed answer per ticket. Three experienced support managers then graded the answers blindly, using scores that combined factual correctness, usefulness and tone. That design provides a useful comparison of draft-answer quality while reducing the chance that graders were influenced by model names or reputations. It is nevertheless a test of answers, not of complete customer-service operations. The researchers did not evaluate whether the models could use tools, look up account information, execute refunds or take other agentic actions. The composite score also deserves some caution. It summarizes several important qualities, but it does not show whether a model’s strength came from accuracy, helpfulness or tone. A production team should examine those dimensions separately, particularly because a polished answer can still contain a costly factual error. ## Quality fell more clearly below the leaders The cheaper alternatives did not cluster as tightly around the top. Ember scored **79.1**, Kite **76.4**, and Harbor **71.0**. Their estimated costs also dropped sharply, with Harbor—the cheapest option—costing just **$0.08 per 1,000 tickets**. This points to a genuine quality-versus-budget question rather than a simple “cheaper is always better” story. The roughly 13-point gap between Harbor and the leaders is much larger than the 0.4-point spread between North and Vale. However, the supplied results do not establish statistical significance for every pairwise comparison, so the lower rankings should not be treated as precise or immutable. The appropriate response is to test unacceptable-error rates and customer outcomes, not to choose solely by average score or unit cost. Estimated model costs also do not capture integration, human review, escalation or the expense of correcting bad answers. ## Ticket type matters more than a single ranking The category results add another reason to avoid declaring an overall champion. North performed best on **billing disputes**, while Vale led on **technical troubleshooting**. That suggests different models may have different workload fit, and it raises the possibility of routing particular ticket types to particular systems. But it is only a hypothesis for further testing: the results do not provide category sample sizes or significance tests. With just 200 tickets overall, some categories may have been small. ## The shared weakness is more important than the top-model gap All five models performed poorly when resolving a ticket depended on an **image attachment**. Unlike the top models’ tiny score difference, this was a consistent failure across the field. A support organization handling screenshots, error messages, damaged-product photos or other visual evidence therefore should not extrapolate from the headline averages. Those averages may overstate readiness for real queues. Image-dependent cases need separate evaluation and, until performance improves, a reliable human-review or escalation path. The other boundaries are equally important. The tickets were in English and came from one software company. Results may not transfer to other industries, customer populations, languages or writing styles. Multilingual support, tool use and agentic behavior remain untested. ## A sensible next step For a similar text-heavy English support workload, Vale offers the strongest starting proposition: near-top measured quality without North’s cost premium, plus faster answers. North’s billing performance still merits investigation, while lower-cost models may be useful where their error profiles are acceptable. The broader lesson is to select against the actual workload. A short pilot should measure correctness, usefulness and tone separately; break results down by ticket type; track visual-attachment failures; and include human-review costs. The benchmark’s practical winner is not a universal champion—it is the reminder that a narrow lead on a scoreboard may not justify a large difference in operating cost.

Z.ai: GLM 5.3 Prime: # The Top Two Tied. One Costs 61% Less. When five language models were graded blind on 200 anonymized customer-support tickets, the most expensive one posted the highest average score. It also bought, statistically speaking, nothing the runner-up didn't already deliver. Model North finished at 84.2; Model Vale at 83.8 — a gap the researchers found statistically insignificant. The price gap was anything but: an estimated $1.90 per 1,000 tickets for North, versus $0.74 for Vale. Vale also answered 18% faster. That's the benchmark's clearest finding, and the one worth carrying into any buying decision: at the top of this test, price did not track quality. ## How the test worked Researchers gave all five models the same 200 English-language tickets, drawn from a single software company, and had each produce one proposed answer per ticket. Three experienced support managers then graded the answers blind — no model names attached — combining factual correctness, usefulness, and tone into a single score. The design has real strengths. Blind grading by practitioners beats automated proxies for something as judgment-heavy as tone, and a fixed ticket set keeps the comparison clean. But the limits matter too. One answer per ticket means no data on consistency — a model that's excellent half the time and confidently wrong the rest looks fine here. And the study doesn't report how the three dimensions were weighted, or how closely the graders agreed with one another. ## What the scores actually show Two patterns matter more than any individual number. First, the top is flat. A 0.4-point gap is a statistical tie: the data can't distinguish it from no difference at all. And since significance was only reported for the North–Vale comparison, the same skepticism belongs elsewhere. Ember (79.1) and Kite (76.4) form a middle tier a few points back, but nothing here settles the ordering between them. Second, the only clearly real gap is at the bottom. Harbor scored 71.0 — roughly 13 points behind the leaders — at $0.08 per 1,000 tickets. Its value per dollar is extraordinary; whether a 71 clears the bar for customer-facing replies is a separate question. Dollar for dollar, the premium is stark: North costs roughly 2.6 times more per point of measured quality than Vale. ## Where the tie breaks The overall tie hides a few edges worth knowing. Vale's 18% speed advantage matters if answers feed live chat or high-volume queues. The models also split by ticket type: North did best on billing disputes, Vale on technical troubleshooting. With 200 tickets divided across categories, those slices are small, so read them as directional. But they're the one place the statistical tie could mislead in practice — a billing-heavy queue might genuinely favor North, a troubleshooting-heavy one Vale. One result didn't separate anyone: all five models performed poorly on tickets that depended on an image attachment. As a buying signal, that's useless — everyone fails. As a deployment warning, it's the most useful line in the study: whatever you pick, route image-dependent tickets to a different workflow, or keep a human there. ## What this benchmark can't tell you - **One company, one language.** Every ticket came from a single software company, in English. Retail, telecom, or multilingual queues may rank these models differently. - **Prose, not actions.** Models were graded on written answers. The test covered no agentic actions or tool use — not the ability to look up an order, issue a refund, or actually resolve anything. - **A snapshot.** Models change quickly. This is one run, one answer per ticket, on 200 items. ## The takeaway The lesson isn't "Vale is the best model." It's that in the one setting tested, a top-tier score came at 39% of the top-tier price, with better speed. Teams paying a premium for a marginal score edge should demand evidence that the edge survives their own ticket mix — especially if that mix is heavy on billing. And the cheapest way to get that evidence is the way this study did it: your tickets, your managers, blind.

Kimi-K3.1-Pro: # KiteBridge Failure Creates $2.8 Million in Duplicate Reimbursements A failure in KiteBridge’s expense-management software generated 13,400 duplicate reimbursements worth $2.8 million on Tuesday, disrupting payments for 91 corporate customers and leaving the company working to recover money sent in error. KiteBridge suspended automated payments less than 90 minutes after deploying a new software version, but not before the duplicate transactions had been created. By Wednesday evening, $2.3 million had been recovered, leaving approximately $500,000 still pending. Employers funded the duplicate payments; no employees had money withdrawn from their own accounts as a result of the incident. KiteBridge makes software used by businesses to manage expenses and employee reimbursements. The failure unfolded after KiteBridge deployed version 4.2 at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m., and the company suspended automated payments at 10:03 a.m. Those three points mark a rapid escalation from a software release to a customer-visible problem and, finally, a shutdown of payment automation. The technical failure began with a payment-processor timeout. A timeout occurs when one system does not receive a response within the expected period, leaving uncertainty about whether a transaction has completed. Payment software may retry in an attempt to finish the transaction safely. KiteBridge’s fallback system—the backup process used when the normal payment flow encounters a problem—treated those retries as new reimbursement requests. Instead of recognizing repeated attempts to process the same payments, it created additional reimbursements. That produced the 13,400 duplicates across the affected customers. Retries are a normal part of networked payment systems because a missing response does not necessarily mean a payment failed. The safeguard is to recognize that the original request and the retry refer to the same transaction. KiteBridge’s fallback process failed at that distinction, turning a mechanism meant to keep payments moving into a source of financial exposure for its customers. The failure also contrasted with KiteBridge’s earlier description of its safeguards. An April changelog said version 4.0 had introduced “exactly-once payment protection.” In payment systems, that term generally refers to controls designed to ensure a transaction is processed only once, even when technical problems cause it to be retried. Tuesday’s incident instead allowed retries to multiply the payments. KiteBridge’s public communications did not initially describe the full problem. Its incident page characterized the issue as delayed reimbursements, rather than duplicate payments. The page first disclosed the duplicates six hours after automated payments had been suspended, meaning the initial public description omitted the financial error at the center of the failure. CEO Lena Ortiz apologized and said KiteBridge would publish a technical report. Her commitment came as the recovery effort remained incomplete: most of the $2.8 million had been recovered by Wednesday evening, but the remaining funds were still pending. Beyond the financial cleanup, KiteBridge faces questions about the reliability of safeguards it had previously advertised and about how quickly it accurately explained the failure to customers.

Z.ai: GLM 5.3 Prime: # KiteBridge Update Glitch Sent $2.8 Million in Duplicate Reimbursements *A payment-processing failure produced 13,400 duplicate payments at 91 corporate customers. The company has recovered $2.3 million; the rest is pending.* Employers at 91 companies found themselves funding reimbursements that were never supposed to exist on Tuesday, after a software update at KiteBridge — a maker of expense-management software — generated 13,400 duplicate payments worth $2.8 million. KiteBridge shut down its automated payments and spent the following day working to recover the funds. By Wednesday evening, $2.3 million had been recovered; $500,000 remained pending. The error never touched employees' own money — no funds were withdrawn from their accounts. But because employers funded the duplicate payments, corporate customers were the ones left carrying the cost until the money could be clawed back. KiteBridge's platform automates a routine piece of corporate finance: employees submit expense claims, and approved reimbursements go out as automated payments to their bank accounts. That automation is why a single bug could move millions of dollars in a little over an hour. ## From deployment to shutdown in 78 minutes KiteBridge released version 4.2 of its software at 8:45 a.m. Tuesday. The first customer complaint arrived at 9:12 a.m. At 10:03 a.m., the company suspended automated payments, ending a run of erroneous transfers that had lasted a little over an hour. The trigger was a payment-processor timeout. KiteBridge's platform relies on an outside payment processor — the intermediary that actually moves money between bank accounts — to execute reimbursements. When the platform asked the processor to carry out a payment and received no confirmation within the expected window, the correct behavior would have been to treat the outcome as unknown and check whether the money had moved before trying again. KiteBridge's fallback system — the backup logic invoked when the normal path fails — did the opposite. It treated each retry as an entirely new reimbursement request, so every new attempt sent money on top of payments that may already have gone through. The retries compounded until they had produced 13,400 duplicate reimbursements worth $2.8 million. ## A safeguard that was supposed to prevent this The failure is an awkward one for KiteBridge. Its changelog for version 4.0, a prior release, said in April that the update had introduced "exactly-once payment protection" — payment engineering's term of art for a guarantee that each payment is processed exactly once, even when a system retries after a failure. A timeout followed by retries is precisely the scenario such a safeguard is meant to cover. The duplication occurred anyway, and KiteBridge has not publicly explained how the protection failed — or whether it applied to the fallback path at all. ## A disclosure that lagged the response KiteBridge's public incident page, the standard channel for updating customers during service problems, initially described the issue as delayed reimbursements. That framing understated what had happened: delayed reimbursements leave employees waiting for money, while duplicate payments leave employers having paid twice. The page did not disclose the duplicates until six hours after automation was suspended — late Tuesday afternoon, by which point the erroneous payments had long since been made. CEO Lena Ortiz apologized and said KiteBridge would publish a technical report. That report will need to answer two questions: why the fallback system treated retries as new requests, and why the "exactly-once" protection the company announced in April did not stop them. Until it appears, $500,000 in duplicate reimbursements remains outstanding.