Head to head: MoonshotAI: Kimi K3.1 Pro vs MiniMax: MiniMax M3

MoonshotAI: Kimi K3.1 Pro: # Flood-Risk Software Maker Raincheck Raises $8 Million *The Austin company plans to hire eight engineers and expand into Florida following a seed round led by Basin Ventures.* Raincheck, an Austin-based company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. Basin Ventures led the round, with participation from existing investor Long Mile. Raincheck has now raised $10.5 million in total. 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 provides additional detail in its [flood-risk methodology](https://example.com/raincheck-methodology). CEO Ana Flores said the software is designed to help customers determine which problems to address first: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck said the new funding will help it hire eight engineers and expand into Florida.

MiniMax: MiniMax M3: # Raincheck Raises $8 Million for Flood-Risk Planning Software *Raincheck helps cities prioritize drainage maintenance ahead of storms, with new funding supporting eight hires and expansion into Florida.* Raincheck, an Austin-based software company, announced today that it has raised $8 million in seed funding to build flood-risk planning tools for local governments. The company maps municipal drainage systems and combines those maps with rainfall forecasts. Cities use its software to decide where to deploy maintenance crews before a storm. Basin Ventures led the seed round, and existing investor Long Mile also participated. Raincheck has raised $10.5 million in total. CEO Ana Flores said: > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Raincheck’s [flood-risk forecasting methodology](https://example.com/raincheck-methodology) produces recommendations rather than directives; 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.

MoonshotAI: Kimi K3.1 Pro: # Foundry Hall to Close After Funding Loss Leaves No Path to Another Program Foundry Hall, the Milwaukee nonprofit startup accelerator that has supported 212 startups since 2018, will close October 31 after its largest corporate sponsor declined to renew its funding. The organization has enough money to operate through October, but not enough to run another full program. Six employees will lose their jobs, and the 12 startups currently using its workspace will have to find somewhere else to work. Blank River Bank supplied 38% of Foundry Hall’s annual budget. Executive director Camille Ross said the organization had hoped another sponsor would fill the gap, but the search did not produce a replacement in time. “We kept waiting for the replacement check that never came,” she said. The bank said its decision reflected a change in community investment priorities and was not related to Foundry Hall’s performance. Its withdrawal leaves the accelerator without the financial base needed to continue, despite a record that extends well beyond the companies currently in the building. Since opening, Foundry Hall has worked with startups that collectively employ 1,140 people. Those figures offer one measure of its reach. For founder Dev Malik, the measure was more personal: a place where an early-stage company could be taken seriously. “My first investor meeting happened in that building,” Malik said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” His account points to the less visible work of an accelerator: giving founders a credible setting, a routine and proximity to others trying to build businesses. The closure will not erase the companies already supported, but it will remove that setting for those who might have come next. The immediate uncertainty belongs to the 12 startups still using the workspace. They must leave while continuing to run their businesses, and the six employees facing job losses must plan their own next steps as they wind down the organization. Deputy Mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies. Those conversations could soften the disruption, but they would not preserve Foundry Hall as an institution or replace the program that is ending. For now, the accelerator will continue operating through October. Its final weeks will be shaped by two tasks at once: helping the remaining companies prepare to move and closing an organization that began with the purpose of helping others begin. The 212 startups and 1,140 jobs will remain part of its record. So will the smaller evidence of its role—a first meeting, a room to work, a founder no longer pitching from a coffee shop. After October 31, Milwaukee will still have that record, but not the place that produced it.

MiniMax: MiniMax M3: # Milwaukee's Foundry Hall to Close October 31 Foundry Hall, a nonprofit startup accelerator that has helped launch 212 companies since opening in 2018, will shut its doors on October 31. The announcement, made this week, ends a seven-year run that organizers say helped put more than a thousand people to work across the region. The accelerator's collapse traces back to a single financial withdrawal. Blank River Bank, which supplied roughly 38 percent of Foundry Hall's annual budget, declined to renew its sponsorship earlier this year. Executive director Camille Ross described the months of limbo that followed in plain terms: "We kept waiting for the replacement check that never came." Foundry Hall has enough reserves to keep operating through the end of October, but not enough to run another full program cycle. By November, six employees will be out of work and the twelve startups currently using its coworking space will need to find somewhere else to plug in their laptops. For the entrepreneurs who came through the program, the news lands in a way that statistics can't quite capture. Dev Malik, who founded a logistics software company during his time at Foundry Hall, recalled meeting his first investor on the building's second floor. Before the accelerator, he had been pitching from the back table of a coffee shop. The bank, asked about its decision, said only that it "reflects a change in our community investment priorities and is unrelated to Foundry Hall's performance." The phrasing underscored a familiar pattern in nonprofit closures: the funder walks, the mission gets buried, and the institution that built the relationships is left to explain what happened. Collectively, the 212 companies that passed through Foundry Hall now employ an estimated 1,140 people. Ross, in remarks to staff and founders, said the accelerator's legacy will outlast its lease. Deputy mayor Susan Cole's office has begun talks with building owners and local universities about temporary space for the dozen startups that will need to relocate before the end of the month. The city has not committed to long-term replacement space. For now, the Foundry Hall team is focused on a softer landing than the one they were handed. Mentorship relationships are being introduced to outside investors. Founders are being connected to peer accelerators in Chicago and Madison. The remaining staff are helping companies update their pitches in case the next meeting, like Malik's first one, happens somewhere unfamiliar. The lights at Foundry Hall will stay on through Halloween. After that, the building goes quiet, and the work that began inside it moves on without the room that helped it start.

MoonshotAI: Kimi K3.1 Pro: # QueueZero’s Proxy demo turns an audience aside into tasks for its CEO QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend a meeting, produce notes and assign follow-up work to software agents. Its launch demonstration also illustrated a central challenge for this kind of software: deciding when conversation becomes an instruction. During the 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. 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 sheet was sent; the other action item did not survive executive review. The exchange was brief, but it points to a broader issue as meeting tools move beyond transcription. A system that only produces notes can misquote or mischaracterize a discussion. One that creates work for software agents must also determine who is speaking, whether a statement is meant as a command and what that command authorizes. QueueZero said Proxy requires confirmation before an external message is sent. Internal tasks can be created automatically, as they were in the demonstration. Administrators can disable automatic task creation. Those controls establish different levels of oversight for actions that stay inside an organization’s workflow and messages that leave it. They also give administrators a way to trade some automation for tighter control. The demonstration showed why that setting may matter: a comment that a person would recognize as an aside can become structured work when software treats it literally. Automatic task creation can reduce the manual work of recording decisions and distributing follow-up. But it also makes the assistant’s interpretation of casual language part of the workflow. Evaluating Proxy will therefore involve more than judging whether its notes are accurate; buyers will also need to examine permissions, confirmation steps and the tasks it generates without approval. The episode does not by itself establish how often Proxy will make that kind of mistake, or how well its confirmation prompts will work in everyday use. Those questions will be tested more broadly when the product enters public beta next week. Proxy costs $30 per user per month. For prospective customers, the beta should provide a clearer picture of its notes, its handoffs to software agents and whether its administrative controls are sufficient to keep convenience from becoming unintended delegation.

MiniMax: MiniMax M3: # QueueZero's AI Meeting Assistant Mistakes Audience Quip for Instruction at Launch QueueZero unveiled its artificial intelligence meeting assistant, Proxy, on Wednesday, demonstrating software designed to attend meetings, generate notes, and distribute follow-up tasks to other software agents. The launch event produced an unscripted moment that quickly became the day's most discussed talking point. During a live demonstration, an audience member interrupted CEO Evan Wu's presentation with a remark that was meant as light commentary: "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, monitoring the audio feed, interpreted the statement as actionable and promptly created two tasks assigned to Wu. The first read "Stop talking," and the second read "Send pricing sheet." Wu laughed when the tasks appeared on screen in front of attendees. He completed the second by sharing the pricing information with the room and deleted the first. QueueZero said the incident illustrates a feature, not a failure. According to the company, Proxy requires explicit human confirmation before any external message, including emails or shared documents, is sent on a user's behalf. Internal task creation, however, operates automatically. Administrators can disable automatic task creation across an organization if they prefer a more conservative posture. The episode has nevertheless focused attention on a broader question facing the workplace AI market: how should software agents handle ambiguity, sarcasm, or offhand remarks in real time? As more vendors introduce autonomous assistants that can parse natural conversation and act without supervision, the risk of well-meaning comments being executed as instructions has emerged as a recurring concern for enterprise buyers evaluating the category. Proxy is priced at $30 per user per month and enters public beta next week. QueueZero has positioned the product as a productivity layer for distributed teams, with the agentic follow-up capability described as a differentiator from transcription-only rivals in a crowded market. In a statement following the event, the company noted that customers retain full control over Proxy's permissions and that its design philosophy treats internal task creation as a low-risk, high-utility capability. Whether that framing holds up in broader deployment will likely be tested as soon as the beta opens to general availability. For Wu, the takeaway was straightforward: even when an AI assistant misfires, the documentation gets done. The customer whose remark triggered the demonstration did not respond to a request for comment.

MoonshotAI: Kimi K3.1 Pro: # Arcwell’s Relay uses AI to assist support agents and flag cancellation risk Arcwell has announced Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to agents. It can also create summaries and identify customers who may cancel, extending its scope from drafting responses to flagging possible churn. The feature set reflects a broader use of AI in customer service: interpreting support activity as well as generating language. Suggested answers assist with the immediate task of handling a request. Summaries condense support activity, while cancellation warnings shift the focus from the next reply to whether a customer may leave. Relay was tested by 12 companies over six months. Arcwell says agents using it closed tickets 14% faster. The company did not release customer-retention results, so the disclosed figure addresses ticket-handling speed, not whether Relay’s warnings helped prevent cancellations. That leaves the announcement with a quantified productivity claim but no comparable result for retention. Relay costs $65 per agent each month, tying the price to the number of support staff using it. CEO Priya Shah framed the product around the difficulty of finding existing company information. “Support agents spend too much time searching for information that their companies already have,” she said. That objective depends on access to company material. Relay requires support transcripts and internal help documents, bringing together the customer conversations it examines and the resources agents use to respond. For prospective customers, adoption is therefore not only a workflow decision; it also involves granting an AI platform access to both kinds of information. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period. Those policies address separate parts of data handling: encryption and limits on model-training use, alongside two defined options for how long information is retained. They establish Arcwell’s stated safeguards for a system built to work with support conversations and internal company knowledge.

MiniMax: MiniMax M3: # Arcwell Unveils Relay, an AI Platform for Support Teams Arcwell has announced Relay, an artificial-intelligence platform that examines customer-support conversations and suggests answers to support agents. The launch comes as artificial intelligence continues to transform the customer service industry, adding another tool for companies seeking to improve customer support. For Arcwell, the announcement marks a significant step in its effort to bring AI into the customer-service workflow. Relay does more than recommend replies. It can identify customers who may cancel and create summaries of support conversations. The cancellation feature is intended to alert agents to possible churn, while summaries condense lengthy exchanges into shorter records. Together, the tools aim to give agents more context and reduce time spent searching for company information or reviewing past interactions. The platform was tested by 12 companies over six months. Arcwell said agents using Relay closed tickets 14% faster. The company did not release customer-retention results, leaving the impact of Relay’s cancellation identification on retention undisclosed. Relay costs $65 per agent each month. CEO Priya Shah said, “Support agents spend too much time searching for information that their companies already have.” Her comment underscores one of Relay’s core dependencies: the system requires access to customer-support transcripts and internal help documents. That access lets the platform draw on a company’s existing information, but it also makes data handling a central issue for prospective customers and one of the platform’s potential challenges. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or 90-day retention period. Relay extends Arcwell’s push to use artificial intelligence in customer service. The platform is designed to help companies improve support by giving agents recommended answers, risk signals and summaries. Its wider value will depend on whether the ticket-closing gains seen in testing translate into better retention, an outcome the company has not yet reported.

MoonshotAI: Kimi K3.1 Pro: 6

MiniMax: MiniMax M3: 6

MoonshotAI: Kimi K3.1 Pro: C

MiniMax: MiniMax M3: C

MoonshotAI: Kimi K3.1 Pro: # Cloudnote debugging page exposed customer prompt excerpts for four hours *Logs show excerpts from 1,842 user sessions. Cloudnote says it has found no evidence of access beyond the researcher who reported the page, but its investigation is ongoing.* Cloudnote said a debugging page containing excerpts of customer prompts was accessible without authentication for four hours on Monday, leaving a limited slice of user-generated content visible to anyone who reached the page. The company removed it 19 minutes after a researcher notified it of the exposure. Logs show excerpts from 1,842 user sessions, according to Cloudnote. Each excerpt could contain as many as 200 characters from a prompt. The figures do not establish how many individual customers were affected or whether every logged excerpt was viewed. Cloudnote has 3.2 million registered users. The company said full documents, passwords and payment information were stored elsewhere and were not present on the page. That limits the types of data at issue, though prompt text can still contain information users expected to remain private. Cloudnote said it has found no evidence that anyone other than the researcher accessed the page. Its investigation is not complete, so the finding does not rule out other access. Key questions remain about how the page became accessible, what the excerpts contained and whether customers will be notified.

MiniMax: MiniMax M3: # Cloudnote Exposed Customer AI Prompts on Unauthenticated Debugging Page *The note-taking app says no one outside the reporting researcher appears to have viewed the data, but roughly 1,842 user sessions were logged during the four-hour window.* Cloudnote, a productivity app used by 3.2 million registered users, accidentally exposed snippets of customer prompts through an internal debugging page that was accessible without authentication for approximately four hours on Monday. The page was meant for internal diagnostics and should not have been reachable from the public internet. According to company logs, the page displayed excerpts drawn from 1,842 user sessions — each containing up to 200 characters of a user's prompt. Cloudnote removed the page 19 minutes after an outside researcher flagged the exposure, the company said. The researcher alerted the company through a standard responsible-disclosure channel. The exposure matters because prompts typed into AI-powered tools often contain context users would not publish elsewhere: draft emails, internal strategy notes, medical or legal questions, and proprietary business information. Even partial excerpts can surface details users expected to remain private, particularly if the prompts referenced identifiable people, products or projects. Cloudnote said it has so far found no evidence that anyone other than the reporting researcher accessed the page during the four-hour window. The company cautioned that its investigation remains active and has not yet determined whether the researcher viewed, copied or stored any of the excerpts. The exposed page did not contain full documents, account passwords or payment information, all of which are stored on separate systems, according to the company. Cloudnote said it plans to notify affected users once it completes its review.

MoonshotAI: 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.

MiniMax: MiniMax M3: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

MoonshotAI: Kimi K3.1 Pro: # KiteBridge software failure creates $2.8 million in duplicate reimbursements A software failure at KiteBridge generated 13,400 duplicate reimbursements totaling $2.8 million on Tuesday, affecting 91 corporate customers and forcing the company to suspend automated payments. What began as a payment-processor timeout became a costly recovery effort for the expense-management software maker and the employers whose funds covered the excess payments. KiteBridge makes software that businesses use to manage expenses and reimbursements. The incident therefore went beyond a service interruption or a delayed status update: it created duplicate payments that later had to be recovered. It also exposed a gap between the payment protection KiteBridge had previously described and the behavior of its backup system when a transaction response did not arrive. The sequence began at 8:45 a.m. Tuesday, when KiteBridge deployed version 4.2. The first customer complaint arrived at 9:12 a.m. By 10:03 a.m., the company had suspended automated payments, stopping the software from issuing further automatic reimbursements while it responded to the failure. A timeout at a payment processor set off the problem. Payment processors are external services that handle transaction instructions, and a timeout means one did not respond within the expected period. In that situation, software may not know whether the earlier attempt failed, is still being processed or completed without returning a timely confirmation. Retrying can be necessary to avoid leaving a payment unfinished, but it also makes recognizing the original transaction essential. KiteBridge’s fallback system—the backup process used when the normal payment path encounters trouble—treated those retries as new reimbursement requests. A retry should represent another attempt to complete the same transaction, not authorization for an additional payment. By losing that distinction, the fallback process turned repeated attempts into fresh requests, ultimately producing the duplicates across the affected customers. The failure also contrasted with a safeguard described in KiteBridge’s release history. An April changelog said version 4.0 had introduced “exactly-once payment protection.” In payment software, “exactly once” generally means that a request should have only one financial effect, even if technical problems cause it to be submitted again. The goal is to distinguish repeated messages about one transaction from genuinely separate transactions. Tuesday’s incident demonstrated that the fallback behavior did not deliver that protection when the timeout occurred. KiteBridge’s public explanation changed as the incident unfolded. Its public incident page initially described the problem as delayed reimbursements. The page first disclosed the duplicate payments six hours after automated payments had been suspended. The distinction mattered for customers: a delay concerns when money moves, while a duplicate concerns how much was sent and how much must be recovered. By Wednesday evening, $2.3 million had been recovered, leaving $500,000 still pending. The recovery had therefore reclaimed most of the duplicate total, but not all of it. No employees had money withdrawn from their accounts; employers funded the duplicate payments. CEO Lena Ortiz apologized and said KiteBridge would publish a technical report. The incident has left the company facing questions about why its backup process failed to prevent duplication and why its public page initially described only delays.

MiniMax: MiniMax M3: # KiteBridge Software Bug Triggers $2.8 Million in Duplicate Reimbursements A routine software update from expense-management company KiteBridge spiraled into a costly malfunction Tuesday morning, generating more than 13,000 duplicate payments worth $2.8 million before the company halted its automated systems roughly an hour after deployment. The trouble began at 8:45 a.m. Tuesday, when KiteBridge pushed version 4.2 of its expense-management platform to production. Within half an hour, at 9:12 a.m., the first customer complaint arrived. The root cause, according to details that emerged later, was a payment-processor timeout that confused KiteBridge's fallback system. When transactions failed to receive confirmation from the external processor, the system treated each retry as if it were an entirely new reimbursement request rather than a duplicate attempt at the same payment. The result was a flood of 13,400 duplicate disbursements across 91 corporate clients. Because the funds originated from employer-funded reimbursement accounts—not employee bank accounts—no individual workers saw money withdrawn from their own pockets. However, corporate finance teams were left sorting out the fallout from unintended transactions totaling $2.8 million. KiteBridge suspended automated payments at 10:03 a.m., roughly 78 minutes after the first complaint and just over an hour after deployment. The company initially described the problem on its public incident page as "delayed reimbursements"—language that significantly understated the situation. The duplicate-payment nature of the failure was not publicly disclosed until six hours after automation had been paused, leaving customers to piece together what had happened on their own. By Wednesday evening, KiteBridge had recovered $2.3 million of the total. The remainder was still pending, though the company indicated recovery efforts were ongoing. ## A Safeguard That Failed The incident raises uncomfortable questions about a feature KiteBridge itself had promoted. An April changelog for version 4.0 announced the introduction of "exactly-once payment protection," a safeguard specifically intended to prevent the kind of duplicate-transaction scenario that occurred Tuesday. The fact that a subsequent update appears to have bypassed or broken that protection is likely to be the central focus of the upcoming postmortem. CEO Lena Ortiz apologized publicly and pledged to publish a full technical report. That report is expected to address how the payment-processor timeout interacted with the fallback logic, and whether existing safeguards can be hardened against similar failures in the future. ## Cleanup and Consequences For affected corporate customers, the practical impact is largely financial reconciliation rather than employee hardship. Still, finance and treasury teams at 91 companies were forced into emergency work, reviewing transactions to distinguish legitimate payments from duplicates and coordinating with their banks and KiteBridge to recover the funds. The episode underscores how a small interaction between two systems—a third-party payment processor and an internal fallback handler—can cascade into millions of dollars in unintended transactions within minutes. It also highlights the reputational cost of delayed disclosure: KiteBridge's customers learned the true scope of the problem hours after the company itself had intervened. KiteBridge has not yet announced compensation for affected corporate clients, though Ortiz's statement suggested further communication would follow the technical report. In the meantime, the incident stands as a reminder that even well-marketed reliability features can fail when adjacent systems behave unexpectedly.

MoonshotAI: Kimi K3.1 Pro: # A near-tie on quality makes cost and speed the real story Model North earned the highest average score in a five-model customer-support benchmark, but that is not the most useful finding. Model Vale finished just 0.4 points behind, a difference the researchers found was not statistically significant, while answering 18% faster and costing much less. For teams evaluating similar text-based support work, Vale is the clearest candidate for a practical quality–cost balance—not a proven universal winner. North averaged 84.2, compared with 83.8 for Vale. Because the difference was not statistically significant, the result should be read as a failure to establish a quality gap in this test, not proof that the models are identical. It does mean that treating North’s numerically higher score as a decisive advantage would overstate the evidence. The economics are less ambiguous within the benchmark’s estimates. Vale cost $0.74 per 1,000 tickets, versus $1.90 for North—about 61% less, or roughly 39% of North’s cost. At 100,000 comparable tickets, that arithmetic corresponds to $74 versus $190. These are estimated model-answering costs, not a complete deployment budget, but they are large enough to matter when no statistically significant quality advantage has been established. Vale’s speed advantage strengthens its operational case, although the summary provides no absolute response times. ## What the test measured Researchers gave each of five models the same 200 anonymized tickets and collected one proposed answer per ticket. Three experienced support managers graded the answers blindly, using a composite of factual correctness, usefulness and tone. That design is more informative than a simple price comparison: it incorporates human judgment about several qualities that matter in support and reduces the chance that model identity influenced scoring. Even so, the score is not a resolution rate. The study evaluated drafted answers, not whether customers were satisfied or problems were actually closed. A combined score also cannot show whether a model’s strength came from accuracy, usefulness or tone. And one answer per ticket does not demonstrate how consistent a model would be across repeated runs. ## Cheaper models involve quality trade-offs The remaining results form a broad quality–cost slope rather than five interchangeable choices. Ember averaged 79.1 at an estimated $0.31 per 1,000 tickets; Kite and Harbor scored 76.4 and 71.0 at $0.22 and $0.08, respectively. Those lower prices may be attractive, but the benchmark alone does not show that the quality trade-off is acceptable for any particular queue. It also does not report significance tests for every pairwise gap, so numerical differences elsewhere should not automatically become firm rankings. Category performance adds useful nuance. North was strongest on billing disputes, while Vale led on technical troubleshooting. Without category-level margins and measures of uncertainty, those findings are best used to design follow-up tests, not to justify an automatic routing policy. They also caution against describing either leader as best at support in general. ## The shared weakness matters All five models performed poorly when a ticket depended on an image attachment. For a support organization, that common weakness could outweigh modest differences in average scores on other tickets. A cheap, fast answer is not useful if it misses evidence contained in a screenshot or other image. Such tickets need a separately tested workflow, potentially including human review, rather than being approved on the strength of the overall averages. The scope is narrow in other ways. All tickets were in English and came from one software company, so the results may not transfer to other industries, customer bases or writing styles. Researchers did not test multilingual support, tool use or agentic actions. The benchmark therefore provides no reliable evidence about responding in other languages, checking live account systems, issuing refunds or completing multistep resolutions. On the evidence available, North’s leaderboard position is a weak basis for paying its higher estimated cost. Vale offers the more compelling starting point for a similar English-language, text-centric drafting workload: no statistically significant quality gap from the leader, faster answers and substantially lower cost. Before deployment, either should be validated on representative tickets, with image-dependent cases and any required tools or languages tested explicitly. The benchmark narrows the shortlist; it does not establish autonomous support readiness.

MiniMax: MiniMax M3: # Top Quality, Different Economics: A Support-Model Benchmark The clearest finding in this benchmark is a near tie at the top—and a meaningful operational difference behind it. North earned the highest average score, 84.2, but Vale’s 83.8 was not statistically distinguishable from it. Vale also answered 18% faster and cost an estimated $0.74 per 1,000 tickets, versus North’s $1.90. That combination makes Vale the more compelling efficiency candidate, even though the study does not justify crowning either model the universal winner. ## How the comparison was made Researchers gave five language models the same set of 200 anonymized customer-support tickets and collected one proposed answer per ticket. Three experienced support managers graded the answers blindly. Their score combined factual correctness, usefulness, and tone, reflecting the practical standard for a reply that must be both accurate and usable in a customer conversation. Blind grading helps limit reviewers’ expectations from biasing the comparison. Combining quality dimensions also reflects support’s reality: a polished but incorrect answer is not a success. Because each model contributed only one answer per ticket, however, the benchmark does not show how much scores might change across repeated attempts. ## Why the headline score is not the full decision North and Vale are separated by only 0.4 average-score points, and the difference was not statistically significant. That should not be rewritten as proof that the models are identical, nor should it be treated as evidence that North’s extra cost buys superior results across the board. On this evidence, both sit at the leading quality tier. The cost gap is more consequential in practical terms: Vale’s estimated cost per 1,000 tickets was roughly 61% lower than North’s. The benchmark also reports that Vale was 18% faster. For a team processing large volumes of support requests, those are advantages that can affect response workflows and operating budgets without requiring a meaningful sacrifice in judged answer quality. The reported speed advantage concerns answer production, not necessarily resolution of the underlying customer issue. There are still reasons not to read the cost figures as a complete business case. They are estimates for answer generation, not full customer-service economics: human review, escalation, integration, and maintenance can change the total expense. A cheap answer that later requires substantial correction may not be cheap after all. ## Specialization matters more than a small score gap The category results suggest a more useful deployment strategy than picking one model for every ticket. North performed best on billing disputes, while Vale performed best on technical troubleshooting. These are task-specific leads, not proof that either model dominates every issue type, and the benchmark gives no category-level margins to assess how decisive they were. Still, they support testing a routing approach: send billing questions to North and troubleshooting questions to Vale, then measure quality, speed, and cost on real traffic before making the rule permanent. The lower-cost options still need careful evaluation. The other three models averaged between 71.0 and 79.1, with estimated costs of $0.08–$0.31 per 1,000 tickets. Those averages suggest a quality trade-off, but the benchmark does not establish a universal cost threshold. Teams handling high volumes of straightforward requests might evaluate those models for narrower, lower-stakes uses, while retaining human review where accuracy matters most. ## The shared failure mode All five models performed poorly when a ticket depended on an image attachment. That is a practical boundary on the otherwise promising results. Tickets may require visual evidence—such as a screenshot, an error message, or proof of purchase—that the model cannot use reliably. These tickets need a different workflow or human escalation; the benchmark does not support trusting the models to handle that evidence unaided. ## What readers should take away The benchmark covers English-language tickets from a single software company, so it cannot establish performance across industries, languages, or customer cultures. It also did not test agentic actions, tool use, or multilingual support. An impressive drafting score says little about a model’s ability to retrieve a refund, update an account, or coordinate a multi-step resolution. The defensible conclusion is therefore narrower than a ranking. North and Vale delivered the strongest judged quality, with no statistically significant difference between them. Vale combined that quality with faster responses and markedly lower estimated costs, making it the benchmark’s clearest efficiency candidate. North remains particularly relevant for billing disputes. Teams should use the results to guide a monitored pilot, not replace judgment with a single score.

MoonshotAI: Kimi K3.1 Pro: # Startup Spotlight: LedgerLark’s Narrow Approach to Post-Trade Work ## Starting with a mismatch Mina Batra’s starting point for LedgerLark was not a broad vision of autonomous finance. It was a specific operational failure. Batra, who spent seven years in operations at a regional brokerage, began exploring the idea after seeing a six-person team spend three days resolving a trade mismatch caused by a copied account number. Her co-founder, Owen Carr, built reconciliation systems at a large bank. Their combined experience points to a focused insight: for small investment firms, the near-term opportunity may be less about replacing post-trade judgment than reducing the repetitive work of gathering evidence around discrepancies. A small error can consume skilled staff time even when its financial impact is uncertain. ## Proposing corrections, not moving money LedgerLark’s AI agents compare trade confirmations, custodian records and internal ledgers. When those records disagree, the agents propose corrections for human review and approval. The product does not execute trades or move money. That boundary matters. It leaves decision-making with the customer and gives LedgerLark a narrower task than autonomous transaction handling. It also puts weight on less visible details: whether the software reads records accurately, shows why a correction is being suggested and fits existing controls. The available information does not establish its error rate or how often reviewers reject its proposals. For a small investment firm, the potential appeal is straightforward: less time spent assembling records and more time for exceptions that require judgment. Yet the review step has to be faster than the manual process it replaces. Staff also need enough context to trust a proposal or challenge it, especially when the same discrepancy appears across different records. ## Early traction, with limits Five firms are paying customers, with annual contracts ranging from $36,000 to $110,000. That shows some willingness to pay for the approach, but five customers cannot yet demonstrate repeatable sales, durable retention or easy deployment. The wide pricing range also raises practical questions about what determines contract size and how much implementation work each account requires. The company says one customer reduced its unresolved daily exceptions from 70 to 18. If sustained, that would be a substantial operational improvement. But it is a single, company-reported example. The measurement period, the types of exceptions involved and any other process changes at the customer are not specified, so the figure should not be treated as a typical result. LedgerLark has raised a $3.5 million seed round led by Motive Row. That capital can support product development and customer work, but it does not answer the adoption questions ahead. ## Competition and caution LedgerLark faces competition from established reconciliation vendors and software built inside banks. Those alternatives may already be embedded in customers’ workflows and control frameworks. LedgerLark’s narrower scope could appeal to firms that do not want a broad systems replacement, provided integration and supervision do not offset the time saved. Financial institutions may also require lengthy security and procurement reviews. That can slow sales and create work for a young company even when a prospective customer is interested. The next useful evidence will be concrete: renewals, results across multiple customers, reviewer rejection rates and the staff time needed for deployment. LedgerLark has identified a bounded problem and attracted initial paying users. Whether its agents become dependable operational tools, rather than another system that operations teams must closely supervise, remains an open—and testable—question.

MiniMax: MiniMax M3: # Startup Spotlight: LedgerLark When a six-person operations team at a regional brokerage spent three days tracking down a single trade mismatch caused by a copied account number, Mina Batra started asking a different kind of question. The error itself was unremarkable. What struck her was the labor required to fix it. That observation has since become the founding premise of LedgerLark, a small company building AI agents for the post-trade work that happens after a trade is executed and before the books are closed. The startup is aimed squarely at small and mid-sized investment firms, a segment that often lacks the headcount to chase down reconciliation exceptions manually. ## Two Operators, One Problem Batra spent seven years in operations at a regional brokerage, working through the kinds of reconciliation puzzles that rarely make headlines but consume hours every day. Her co-founder, Owen Carr, came at the same problem from a different angle, having built reconciliation systems inside a large bank. The pair's shared conclusion, according to conversations with the company, is that the bottleneck in middle- and back-office work is not analysis but triage: figuring out which of thousands of daily exceptions actually matter. That framing shapes how LedgerLark positions its product. The agents compare trade confirmations, custodian records and internal ledgers, flag mismatches and propose corrections that a human reviewer can accept, reject or modify. The product does not execute trades or move money, a boundary the founders describe as deliberate. ## Early Traction LedgerLark is at a stage where the customer count is small enough to name out loud. Five investment firms are paying for the software, with annual contracts ranging from $36,000 to $110,000 depending on size and complexity. The company says one customer reduced its unresolved daily exceptions from roughly 70 to 18, though such figures are typically reported by vendors themselves and have not been independently verified. The startup has closed a $3.5 million seed round led by Motive Row, capital it plans to put toward engineering hires and security certifications rather than aggressive sales expansion. ## Competition and Constraints The market LedgerLark is entering is not empty. Established reconciliation vendors have spent decades embedding themselves in financial workflows, and large banks frequently operate proprietary systems built over many years. For a young company, displacing either is a long road, and incumbents are not standing still on AI features of their own. There are also structural headwinds specific to financial services. Security reviews, procurement processes and compliance checks at regulated institutions can stretch onboarding from weeks to many months. LedgerLark's decision not to handle trade execution or fund movement may simplify some of those reviews, but it also means the company must prove its value within a narrow slice of the workflow. Whether AI agents can materially change that picture remains an open question. The underlying work of reconciliation is rule-heavy and well-understood, which is both an opportunity (existing software often does it slowly) and a risk (incumbents can add similar capabilities to their own platforms). ## The Wider Lens The pitch behind companies like LedgerLark is familiar by now: that software agents will take over the routine cognitive work that fills a back-office day. Whether that thesis holds in practice, especially in environments where errors carry real financial and regulatory consequences, will depend on more than model performance. It will depend on whether the human reviewers who approve each proposed correction come to trust the system, and whether the system earns that trust consistently. For now, LedgerLark is a five-customer, seed-stage company with a focused product and a clear point of view about where the friction lives. That is a reasonable place to start.

MoonshotAI: Kimi K3.1 Pro: # HelioDock launches Night Shift for remote robot oversight, raises $24 million *The warehouse robotics company’s monitoring software will be generally available Sept. 15 for $90 per robot per month.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor robot fleets remotely. The company also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. Night Shift gives HelioDock a software product to sell alongside its robots, with a pitch centered on overseeing machines without requiring a manager to travel to the warehouse when something goes wrong. Its commercial case rests partly on company-reported results from a private beta, including a reduction in unplanned stoppages that has not been independently verified. The product also enters a market with existing fleet-management offerings, and feedback from one beta customer identified connectivity as a practical constraint. CEO Mara Velez described the problem Night Shift is intended to address in terms of the trip a manager can be forced to make because of a single machine. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” she said. Night Shift spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock said those beta customers experienced 17% fewer unplanned robot stoppages. The company is presenting that reduction as evidence of the software’s operational value, but the figure remains a company-supplied claim rather than an independently verified result. The beta provides the disclosed basis for that claim: multiple logistics operators, hundreds of robots and more than a year of private testing. It also produced a more qualified account of how the system performs outside facilities with dependable network service. Aaron Pike of Fulcrum Logistics, a beta customer, said Night Shift worked well in facilities with strong connectivity but became less useful during rural LTE outages. His experience does not undermine the basic rationale for remote oversight, but it makes the quality of the connection part of the product’s value. A system intended to help managers work from a distance becomes less useful when the link to the facility is interrupted. When connectivity fails, robots return to their locally configured behavior. Remote control becomes unavailable. That distinction separates the machines’ local fallback from the remote layer Night Shift provides: the robots retain their local instructions, while the ability to control them remotely does not continue through the outage. For operators evaluating the software, Pike’s account and the fallback behavior define two sides of the same consideration. Strong connectivity supported the more favorable experience he reported, while rural LTE outages reduced the system’s usefulness. Losing a connection does not leave robots without their locally configured behavior, but it does remove a key remote capability. Night Shift will also face competition when it reaches general availability. Relay Robotics and YardOS already offer competing fleet-management products, so HelioDock will not be introducing the category to warehouse operators. Its initial case for the software rests on remote fleet monitoring and the stoppage reduction it reports from the beta, with the 17% figure still awaiting independent verification. The $24 million Series B was announced alongside the product launch. North Coast Ventures led the round, which increases HelioDock’s total funding to $39 million as the company prepares to move Night Shift beyond the nine logistics companies that participated in its private beta. Night Shift will become generally available Sept. 15. HelioDock will charge $90 per robot per month, tying the recurring cost directly to the number of machines an operator places under the software rather than setting a single price for an entire fleet.

MiniMax: MiniMax M3: # HelioDock Unveils Night Shift Software for Remote Warehouse Robot Oversight *The platform debuts alongside a $24 million Series B as the company moves from a year-long private beta to general availability in September.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, a software platform that lets logistics operators monitor and direct robot fleets from off-site locations rather than from a control room inside the warehouse. The release is the company's new software offering built around remote supervision and arrives as warehouse automation moves from small pilots to wider deployment. Night Shift runs on HelioDock's existing robot hardware, according to the company, and does not require new on-site equipment to deploy. The debut lands alongside a $24 million Series B funding round led by North Coast Ventures, bringing HelioDock's total outside capital to $39 million. The new capital is intended to support product development, customer onboarding and an expansion of the engineering team behind Night Shift, according to the company. Night Shift spent fourteen months in a private beta with nine logistics companies running roughly 310 robots across a mix of distribution centers and fulfillment sites. HelioDock says beta customers recorded 17 percent fewer unplanned robot stoppages after adopting the software. The company has not published the methodology behind that figure, and it has not been independently verified. In an interview, HelioDock CEO Mara Velez framed the product as a response to a recurring operational headache. "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. Night Shift is designed to surface alerts, route diagnostics and intervention options on any internet-connected device, so that a single remote operator can keep tabs on robots across multiple sites and reroute individual machines when they stall. That capability drew interest from Fulcrum Logistics, one of the beta participants. Aaron Pike, who oversees automation deployment at Fulcrum, said the platform performed reliably at the company's well-connected urban and suburban facilities. The picture was different in rural locations, where Pike described Night Shift as "less useful" during LTE outages. Because the software relies on a steady link between each robot and HelioDock's cloud, interruptions in cellular coverage limited what remote operators could see and do. HelioDock acknowledges the dependency. When a robot loses its connection to Night Shift, it reverts to whatever behavior is configured locally, and remote control is unavailable until the link returns. The company describes the fallback as a safety feature, noting that robots should default to predictable routines when cut off from central oversight. For operators running sites in connectivity-poor areas, the trade-off between remote visibility and autonomous fallback will likely shape deployment decisions. The market for fleet-management software has tightened as warehouse robotics have spread. Relay Robotics and YardOS both sell competing platforms, and larger logistics firms have begun building proprietary dashboards of their own. Night Shift will need to win customers on the strength of its features, pricing and uptime as much as on HelioDock's hardware footprint. Pricing for Night Shift is set at $90 per robot per month, with general availability beginning September 15. HelioDock says the per-robot model is meant to scale costs with fleet size rather than head count, an approach the company argues ties spending directly to the value automation delivers. The September launch will be the first test of Night Shift outside a private beta, and the answers HelioDock gets from customers in connectivity-constrained sites may matter as much as the financial runway behind it.