Head to head: DeepSeek: DeepSeek V4 Pro 0813 vs muse-spark-1.3

DeepSeek: DeepSeek V4 Pro 0813: # QueueZero’s Proxy AI Turns Demo Quip Into CEO Task SAN FRANCISCO — QueueZero on Wednesday launched Proxy, an AI meeting assistant that can attend meetings, produce notes and assign follow-up work to software agents. The launch demonstration quickly showed how automatic task creation can turn an offhand remark into an executive assignment. During the event, 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 and created two tasks assigned to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. The episode reflected Proxy’s default behavior: internal tasks can be created automatically from meeting dialogue without a user opening the application. QueueZero said Proxy requires confirmation before any external message is sent. Internal tasks, such as assignments to project management tools or follow-up reminders, can be created automatically, and administrators can disable automatic task creation if a team prefers a review-first workflow, according to the company. “Proxy is designed to close the gap between a meeting decision and the work that follows it,” Wu said. “Today’s demo showed that context and control settings are just as important as automation.” Proxy costs $30 per user per month and enters public beta next week. QueueZero said early users include professional services firms and product teams that want to automate note-taking, status updates and client follow-up. The company framed the launch as part of a broader push toward agentic workplace tools that can act on a user’s behalf without direct supervision. The launch comes as a growing number of enterprise software vendors are adding meeting-summarization and task-automation features, trying to reduce busywork but also confronting questions about how much autonomy AI agents should have. QueueZero’s approach is tiered: high automation for internal task creation, with a mandatory confirmation layer for outward-facing actions. The incident drew a brief laugh from the room, but it also served as a live stress test of the product’s task-creation logic. QueueZero emphasized that external communications remain protected by a confirmation step. The pricing sheet was sent. The software’s interpretation was technically accurate, if operationally blunt.

muse-spark-1.3: # QueueZero Launches Proxy AI Meeting Assistant **SAN FRANCISCO —** QueueZero launched Proxy, its AI meeting assistant, on Wednesday, pitching the tool as a way to move meetings from transcription to execution by attending calls, producing notes and assigning follow-up work to software agents. CEO Evan Wu demonstrated Proxy during a live product event. As Wu described the system’s ability to listen for action items, 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. The system created two tasks assigned to Wu: “Stop talking” and “Send pricing sheet.” The tasks appeared on screen in the demonstration workspace. Wu laughed, completed the second task and deleted the first. QueueZero said the exchange was unplanned and used it to outline how Proxy handles permissions. According to the company, Proxy requires confirmation before an external message is sent. Internal tasks can be created automatically. Administrators can disable automatic task creation, QueueZero said, requiring Proxy to submit proposed tasks for human review before they are added to queues or assigned to agents. The launch places QueueZero among a growing group of enterprise software vendors competing to automate post-meeting work. Unlike basic transcription tools, Proxy is designed to identify owners for next steps and route those steps to connected software agents for scheduling, documentation and customer follow-up. QueueZero said Proxy is intended for sales, support and product teams that spend significant time converting conversations into tickets and summaries. The company said early pilot customers used Proxy to draft recaps, update customer records and prepare follow-up emails, with employees retaining final approval over outbound communication. Analysts said the category faces questions around accuracy, permissions and user trust, particularly as assistants gain the ability to act across calendars, inboxes and internal task systems without step-by-step direction. Vendors have responded by adding review steps and audit logs for agent actions. QueueZero did not disclose pilot results or customer names on Wednesday. The company said pricing and packaging could change before general release. Proxy costs $30 per user per month. It enters public beta next week, the company said. QueueZero said general availability will follow after feedback from the beta period, including testing of task controls and agent integrations.

DeepSeek: DeepSeek V4 Pro 0813: C

muse-spark-1.3: C

DeepSeek: DeepSeek V4 Pro 0813: # Foundry Hall to close after main sponsor exits Foundry Hall, the Milwaukee nonprofit startup accelerator, will close Oct. 31 after losing its largest corporate sponsor and failing to close the resulting budget gap. Opened in 2018, Foundry Hall supported 212 startups that now employ 1,140 people. But its operating model leaned heavily on one donor: Blank River Bank supplied 38 percent of the annual budget. When the bank declined to renew its sponsorship, the organization had enough money to operate through October but not enough to run another full program. Executive director Camille Ross described a long, ultimately fruitless wait. “We kept waiting for the replacement check that never came,” she said. The closure means six employees will lose their jobs, and twelve startups currently working in the space must find somewhere else to go. For some founders, the loss is personal. Dev Malik remembered what the space gave him. “My first investor meeting happened in that building. Before Foundry Hall, I was pitching from the back table of a coffee shop,” he said. Blank River Bank said the decision reflects a change in its community investment priorities, not a judgment on Foundry Hall’s performance. The bank’s statement was cordial but did not suggest an immediate reversal was possible. City officials are looking for ways to keep the displaced startups together, or at least housed. Deputy mayor Susan Cole said the city is speaking with building owners and local universities about space for the current companies. The response suggests that the companies themselves are still seen as viable; it is the institution around them that proved fragile. Foundry Hall’s closing is not a story of scandal or sudden failure. It is a story about concentration. A nonprofit built on one major corporate relationship can survive as long as that relationship holds. When it ends, the countdown begins. The programs, the desks, the introductions, the noise of early-stage work — all of it depends on a renewal decision made in a bank boardroom. Ross did not frame the ending as anger. She framed it as a deadline that could not be postponed. The startups will scatter. The employees will move on. The numbers — 212 companies, 1,140 jobs — will remain as evidence of what the accelerator produced. But the building will close with the same suddenness that founders know too well: everything can change between one quarter and the next, and the replacement check does not always arrive.

muse-spark-1.3: # Foundry Hall to Close Oct. 31 After Seven Years in Milwaukee Foundry Hall, the nonprofit startup accelerator that became a fixture of Milwaukee’s startup scene, will close on Oct. 31, ending a seven-year run that helped launch more than 200 young companies. The organization opened in 2018 to give Wisconsin founders a place to work, meet investors, and learn how to grow. Since then, it has supported 212 startups that now collectively employ 1,140 people, a footprint that extended well beyond its downtown building. That building will soon go quiet. Six employees will lose their jobs, and 12 startups currently working out of its co-working space must find somewhere else to go by the end of the month. Executive Director Camille Ross said the closure came down to money. Blank River Bank, Foundry Hall’s largest corporate sponsor, supplied 38% of its annual budget and declined to renew its support for next year. With enough cash to operate through October but not to fund another full cohort, the board voted to shut down. “We kept waiting for the replacement check that never came.” The bank attributed the decision to a shift in its community investment priorities rather than to concerns about Foundry Hall’s work. For founders who came up through the program, the loss is personal as well as practical. Dev Malik, who built a logistics software company after joining Foundry Hall’s second cohort, still remembers his early days there. “My first investor meeting happened in that building. Before Foundry Hall, I was pitching from the back table of a coffee shop.” Malik said mentors reviewed his pitch deck, introduced him to his first major customer, and taught him how to hire. His company now employs 34 people in Milwaukee, jobs he traces directly to those months downtown. Stories like his are why city officials say they are trying to soften the landing. Deputy Mayor Susan Cole’s office is talking with building owners and local universities about temporary space for the displaced companies. Still, those efforts would preserve desks, not the program itself. Foundry Hall’s mentorship network, investor demo days, and workshops have no successor lined up. Ross said staff spent the summer approaching other funders, but none could replace the bank’s share on short notice. In its final weeks, the accelerator is focused on hand-offs — connecting founders to lawyers, accountants, and potential landlords, and archiving curriculum materials for community use. Ross has avoided framing the closure as a failure. Attendance remained strong this year, and several alumni companies continue to raise funding. The problem, she suggested, was structural: a nonprofit dependent on a single large donor leaves little room when priorities change. On an afternoon, the hum remained. Whiteboards were still covered in projections. Only the calendar on the wall, ending in October, suggested an ending. For Milwaukee, which has worked to keep young companies from leaving for Chicago or the coasts, Foundry Hall’s closing removes one of the few places built specifically to make them stay.

DeepSeek: DeepSeek V4 Pro 0813: Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

muse-spark-1.3: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

DeepSeek: DeepSeek V4 Pro 0813: # HelioDock Launches Night Shift for Remote Warehouse Robot Oversight *The software, tested across 310 robots in a 14-month beta, arrives with $24 million in new funding and a $90-per-robot monthly price.* HelioDock, a maker of autonomous warehouse robots, on Thursday introduced Night Shift, a fleet-management application that lets operators monitor and control robots from outside the facility. The launch was paired with a $24 million Series B led by North Coast Ventures, bringing the company’s total funding to $39 million. Night Shift targets a familiar operational headache: a manager being called to a warehouse after hours because a robot has stopped or misread a pallet. Instead of requiring an on-site response, the software provides remote visibility into a fleet and allows operators to intervene from elsewhere. HelioDock says the product spent 14 months in a private beta involving nine logistics companies and 310 robots. The nine logistics companies in the beta operated a combined 310 robots, a scale that exposed the software to real-world variability in network conditions and warehouse layouts. The company reports that beta customers saw 17% fewer unplanned robot stoppages. That figure has not been independently verified, and HelioDock did not release the underlying data behind the measurement. The beta group was small enough that the result should be read as an early signal rather than a proven industry benchmark. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” said CEO Mara Velez. Her framing points to the human cost of robot errors: a single stopped machine can pull a manager out of bed for a problem that may take minutes to fix once someone is on site. Beta feedback points to both the promise and the limits of remote management. Aaron Pike of Fulcrum Logistics said Night Shift worked well in facilities with strong connectivity but became less useful during rural LTE outages. When connectivity fails, robots return to their locally configured behavior, and remote control becomes unavailable. In those moments, the software cannot substitute for a person on site. The rural LTE outages described by Pike are a reminder that remote management depends on infrastructure outside the warehouse operator’s control. Pike’s account suggests that the software’s usefulness is tied to network reliability. That constraint is not unique to HelioDock, but it shapes how Night Shift should be evaluated. The product can reduce the number of late-night trips to a facility, yet it does not remove the need for local fallback procedures. Facilities with unstable network connections may see fewer benefits than those with reliable infrastructure. That means Night Shift is best understood as a tool for reducing, not eliminating, on-site responses. Night Shift will cost $90 per robot per month and will become generally available September 15. The launch puts HelioDock in direct competition with existing fleet-management products from Relay Robotics and YardOS. Those competitors already offer remote monitoring tools, which means HelioDock is entering a market where buyers have alternatives to compare. At $90 per robot per month, a fleet of 100 robots would cost $9,000 per month, a recurring expense that operators will weigh against the cost of on-site troubleshooting. The product will be sold as a subscription.

muse-spark-1.3: # HelioDock Launches Night Shift to Let Operators Watch Warehouse Robots Remotely *After a 14-month beta with nine logistics operators, the warehouse robotics company is selling remote fleet monitoring alongside a $24 million funding round* HelioDock on Thursday launched Night Shift, software that lets operators monitor fleets of its autonomous warehouse robots from outside the facility, removing the need for a manager to travel to the warehouse each time a machine stops. The release shifts HelioDock from selling autonomous warehouse robots alone to selling remote oversight of those machines. The company paired the launch with a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million, at a time when competing fleet-management products from Relay Robotics and YardOS are also available to warehouse operators. Night Shift arrives after 14 months in private beta. HelioDock says nine logistics companies tested the software across 310 robots, giving the company more than a year of feedback before opening the product to general availability. That test population provides the basis for the performance claims HelioDock is making at launch. HelioDock says those beta customers experienced 17% fewer unplanned robot stoppages while using Night Shift. The figure has not been independently verified. No independent verification was provided. The company has not disclosed how it defined an unplanned stoppage across different warehouses or whether the comparison reflects performance by the same fleets before Night Shift was installed. Chief executive Mara Velez described the problem in terms of overnight callouts. “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. The comment identifies the case Night Shift is intended to address: a single robot halted by an unexpected obstacle, with work waiting until a person arrives to assess the scene. By moving supervision off site, Night Shift allows that assessment to begin without a trip to the facility. HelioDock has not said how many robots a single remote operator is expected to watch at once or what specific interventions remote monitoring permits beyond observation of fleet activity. Results in the beta depended in part on connectivity. Aaron Pike of Fulcrum Logistics, one of the beta customers, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His experience points to uneven performance between well-connected sites and locations where the wireless link was less reliable. HelioDock said robots return to their locally configured behavior when connectivity fails. Remote control becomes unavailable during those periods. The company has not detailed how an operator is notified when the link is lost or how quickly a robot makes the transition to local behavior. That fallback sets the practical limit of remote supervision. While the connection holds, an operator outside the warehouse can remain involved in fleet activity. Once the connection breaks, oversight pauses and each machine proceeds under its own onboard settings until service returns. For a buyer, the quality of in-building and regional coverage therefore shapes the usefulness of Night Shift as much as the behavior of the robots themselves during a shift. HelioDock will charge $90 per robot per month for Night Shift, with general availability starting September 15. At that rate, a 20-robot deployment would cost $1,800 per month for the software, while a 100-robot fleet would cost $9,000 per month. A 50-robot fleet would therefore cost $4,500 per month. The September date will move Night Shift from a limited test involving nine companies and 310 robots to open sale. Whether wider deployment reproduces beta outcomes will turn in part on network conditions across different facilities and on whether the 17% reduction in unplanned stoppages reported by HelioDock holds outside the original test group.

DeepSeek: DeepSeek V4 Pro 0813: # LedgerLark’s small-bore attack on the post-trade exception pile *Startup Spotlight* When Mina Batra describes the problem behind LedgerLark, she does not start with market structure or AI. She starts with a copied account number. At a regional brokerage where she worked, a six-person operations team spent three days resolving a trade mismatch. The error was simple. Finding it was not. Batra spent seven years in operations at that brokerage. Owen Carr, her co-founder, built reconciliation systems at a large bank. Together they are building AI agents for post-trade operations at small investment firms. The product compares confirmations, custodian records and internal ledgers, then proposes corrections for human approval. It does not execute trades or move money. That constraint is central to the pitch: LedgerLark wants to be a control tool, not a trading system. ## The product The insight is not that reconciliation is new. It is that smaller firms often run the same operational risk as larger institutions with fewer tools. A mid-sized shop may still reconcile across spreadsheets, email and portal downloads. When something breaks, the investigation is manual and slow. LedgerLark is trying to give those firms an AI agent that reads the documents, spots mismatches and drafts a suggested fix, leaving the final decision to a person. ## Early traction Early traction is real but limited. LedgerLark has five paying customers, with annual contracts ranging from $36,000 to $110,000. The company says one customer reduced its unresolved daily exceptions from 70 to 18. That is a promising operational number, but it is a single data point from a small base, and early adopters may be unusually willing to adjust workflows. The company has raised a $3.5 million seed round led by Motive Row. The round should give it room to build integrations and hire, but it does not remove the harder question: whether small investment firms will buy a standalone exception tool rather than waiting for their existing providers to add similar features. ## Competition and uncertainty Competition is the main source of uncertainty. Established reconciliation vendors and internal bank software already cover parts of this workflow. Those incumbents have years of security documentation, audit trails and relationship history. LedgerLark’s response is focus. It is not trying to serve the largest banks; it is aiming at firms that may be underserved by enterprise platforms. That positioning helps, but it also means the startup needs to prove it can integrate with enough custodians, portfolio systems and internal formats to be useful without becoming a heavy implementation project. Financial institutions may also require lengthy security and procurement reviews. AI-based suggestions add another layer of scrutiny. Buyers will ask how the models handle ambiguous data, whether suggestions are explainable, and what happens when the system is wrong. The human-approval step is a reasonable safeguard, but it still needs to be tested across many trading days, instruments and failure modes. What makes LedgerLark worth watching is not the claim that AI can fix operations. It is the narrower question of whether a focused agent can turn a three-day mismatch hunt into a same-day decision. If the company can show that across more customers, and survive the procurement gauntlet, it may have a small but durable place in the back office. If not, it will remain an interesting experiment in applying language-model automation to a conservative corner of finance.

muse-spark-1.3: # Startup Spotlight: LedgerLark Wants to Clean Up Post-Trade Operations For many small investment firms, the trade is the easy part. What happens afterward — matching confirmations, custodian records, and internal ledgers — can absorb hours of careful, repetitive work. LedgerLark is building AI agents to handle that reconciliation work. The startup sells to small investment firms where operations teams are thin and mismatches can linger for days. ## An operations problem Co-founder Mina Batra spent seven years in operations at a regional brokerage before starting LedgerLark. Co-founder Owen Carr previously built reconciliation systems at a large bank. Batra began exploring the idea after watching a six-person team spend three days resolving a trade mismatch caused by a copied account number, according to the company. The incident was not unusual, in the company's account, but illustrative: a small clerical error triggering emails, phone calls, and manual checks across systems that do not quite align. That experience shapes LedgerLark's pitch. Rather than replace operations staff, the company argues that much of exception handling follows recognizable patterns that software can surface and draft for review. ## What the product does LedgerLark's agents compare three sources — trade confirmations, custodian records, and a firm's internal ledger — and flag inconsistencies. When the system spots a likely cause, such as a mismatched account number, quantity, or settlement date, it proposes a correction for a human to approve or reject. The product does not execute trades or move money. That limitation is intentional, according to the company, both to reduce risk and to fit within the permissions firms are willing to grant a young vendor. Humans remain in the loop for every proposed fix. Operations staff can see the underlying records the agent used and accept, edit, or dismiss the suggestion. The company positions this as assisted review rather than automation, though how much time that review actually saves across different firm types is still an open question. ## Early customers and numbers LedgerLark says five firms are now paying customers, with annual contracts ranging from $36,000 to $110,000. The range suggests some variation in firm size or scope of deployment, though the company did not disclose additional details or name the customers. As an example of impact, the company says one customer reduced its unresolved daily exceptions from 70 to 18 after adopting LedgerLark. That figure has not been independently verified, and it represents a single deployment rather than typical performance. The startup has raised a $3.5 million seed round led by Motive Row. ## Competition and open questions LedgerLark is entering a crowded and cautious market. Competitors include established reconciliation vendors with long-standing banking relationships, as well as internally built software that larger institutions maintain and smaller firms adapt informally. Even where the technology works, distribution may be slow. Financial institutions often require lengthy security and procurement reviews before allowing a startup to touch operational data, particularly systems adjacent to ledgers and custodians. For a seed-stage company with five customers, those cycles could shape growth more than product features. Another question is how LedgerLark performs when exceptions are ambiguous rather than clerical. A copied account number is relatively straightforward. Mismatches tied to corporate actions, fee schedules, or counterparty delays can require judgment and context the system may not have. For now, LedgerLark offers a narrow, pragmatic thesis: small firms face the same post-trade complexity as larger ones, without the staff to absorb it. Whether supervised agents can reliably earn trust in that workflow — and clear procurement — will be the test to watch.

DeepSeek: DeepSeek V4 Pro 0813: 6

muse-spark-1.3: 6

DeepSeek: DeepSeek V4 Pro 0813: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin startup maps drainage systems and rainfall forecasts to help local governments prioritize storm maintenance.* Raincheck, an Austin-based flood-risk software company 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, bringing the company’s total raised to $10.5 million. The company maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. Raincheck 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.” More detail is available in the company’s [flood-risk methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana, though 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. 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.

muse-spark-1.3: # Raincheck raises $8 million to help cities prioritize flood work before storms *The Austin startup maps drainage systems against rainfall forecasts to guide pre-storm maintenance, with backing led by Basin Ventures* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding. The round was led by Basin Ventures, with participation from existing investor Long Mile. The raise brings Raincheck's total funding to $10.5 million. The company maps drainage systems and combines those maps with rainfall forecasts to help cities decide where to send maintenance crews before a storm. More detail is available in [an overview of 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 contracts have not been publicly announced. CEO Ana Flores said the challenge for customers is prioritization: > “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.” Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

DeepSeek: DeepSeek V4 Pro 0813: **The Cost of a Point: What Our Customer-Support AI Benchmark Actually Proved** We started with a simple question: which language model writes the best customer-support answers? After putting five leading models through a blind evaluation on 200 real tickets, we ended up with a much more interesting conclusion. The biggest difference between the top models isn’t how well they answer customers—it’s how much you pay them to do it. In our test, the two highest-scoring models produced answers that human managers could not tell apart in quality. Yet one of them costs roughly 2.5 times more than the other. That finding, more than any single score, should shape how support teams think about AI adoption in 2025. ### The headline result The top performer, which we’ll call **Model North**, earned an average score of **84.2 out of 100**, as judged by three experienced support managers. Close behind was **Model Vale** at **83.8**. Statistically, that gap is meaningless. The difference of 0.4 points is well within the margin of error for a test this size. When the managers reviewed answers blindly, they effectively rated North and Vale as peers in terms of factual correctness, usefulness, and tone. The difference shows up in the price tag. North costs an estimated **$1.90 per 1,000 tickets**. Vale costs **$0.74**. That means a team handling 100,000 tickets per month would spend **$190,000** with North versus **$74,000** with Vale—a six-figure annualized gap for a quality difference no one could detect. And Vale answered **18% faster** on top of that. This is not a call to crown Vale as the overall winner. It’s a reminder that once models reach a certain competence threshold, the economic questions matter more than leaderboard rankings. ### Where the models actually diverged The rest of the field was more clearly separated. **Model Ember** landed at **79.1**, **Model Kite** at **76.4**, and **Model Harbor** at **71.0**. Each step down brought a meaningful quality drop—answers became slightly less complete, tone less natural, and a few more facts went slightly sideways. But here’s the nuance: the “best” model depended on the ticket type. **North performed best on billing disputes**, where precision and careful wording matter enormously. A wrong sentence about a charge can escalate a customer’s frustration. **Vale performed best on technical troubleshooting**, where clarity and step-by-step structure dominate. So even among the top two, the right choice may depend on what your customers actually ask about. We can’t tell you which model to buy without knowing your ticket mix. That’s not a hedging answer—it’s the only honest one. ### The shared failure point If there was one universal finding, it was this: **every model struggled when a ticket depended on an image attachment.** Screenshots of error messages, photos of broken hardware, scanned invoices—when the critical information lived in an image, scores dropped across the board. None of the five models reliably extracted and acted on visual context. For support teams that handle meaningful volumes of image-based tickets, this remains a hard limitation, not a minor inconvenience. ### What this test did and didn’t measure The test used **200 anonymized tickets from a single software company**, all in English. Three support managers graded answers blindly, combining factual correctness, usefulness, and tone into a single score. That design is strong for controlled comparison but narrow in scope. We did not test: - **Agentic actions**—whether models can actually update a record, issue a refund, or change a subscription, rather than just draft a reply. - **Tool use**—calling internal APIs or knowledge bases in real time. - **Multilingual support**—all tickets were English. - **Long-tail edge cases**—200 tickets is enough to find broad patterns, not rare failure modes. The cost estimates are just that: estimates based on current pricing and typical token usage. They will shift as providers update pricing. ### The practical takeaway If your team is choosing a model for customer support, don’t fixate on the top score in a benchmark. The top models in this test were statistically indistinguishable in quality. Choose based on your actual ticket mix, your volume, your speed requirements, and your budget. And keep a human in the loop for anything involving a screenshot. The models aren’t ready for that yet.

muse-spark-1.3: # The Sweet Spot in Support AI Is Cheaper Than You Think *In a blind test of 200 real tickets, the priciest model failed to pull ahead, and that changes the buying math.* For support leaders shopping for AI help, the usual assumption is simple: better answers cost more money. A new benchmark of five language models suggests that logic no longer holds at the top end. The headline finding is not that one model crushed the rest. It is that two models effectively tied for best quality, yet one costs less than half as much to run and answers noticeably faster. For anyone deploying at scale, that tie is the story. ### How the test was run Researchers collected 200 anonymized customer-support tickets from a software company and asked each of the five models to draft one proposed answer per ticket. Three experienced support managers then graded those answers blindly, without knowing which model wrote what. That blind grading matters. It prevents brand bias and forces judges to focus on the text in front of them. Scores combined three things customers actually care about: factual correctness, usefulness for resolving the issue, and tone. This was not a test of autonomous agents that can issue refunds, query databases, or follow up across channels. It was narrower and, for many teams, more relevant: can the model write a good first-draft reply a human agent could send with minimal edits? ### A statistical tie at the top Model North earned the highest average score at 84.2 out of 100. Model Vale was right behind at 83.8. That 0.4-point gap sounds like a win for North until you look at the statistics: the researchers say the difference was not statistically significant. In plain terms, with this sample size and judging variation, we cannot confidently say North is actually better. They are, for practical purposes, tied. What is significantly different is the bill. North cost an estimated $1.90 per 1,000 tickets. Vale cost about $0.74 for the same volume, a saving of more than 60%. Vale was also 18% faster in its responses. Across tens of thousands of tickets per month, the tie stops looking boring. A support operation handling 100,000 tickets a month would spend $190 with North’s pricing versus $74 with Vale’s for equivalent average quality, plus get quicker drafts for agents waiting on suggestions. There were hints of specialization inside that tie. North performed best on billing disputes, where precise policy language and careful handling matter. Vale led on technical troubleshooting, where step-by-step reasoning and diagnostic thoroughness pay off. Neither advantage was large enough to break the overall deadlock, but a company drowning in one type of ticket might reasonably lean that way. ### What about the cheaper options? The other three models show why chasing the lowest price alone still has limits. Ember and Kite landed in the upper 70s on average, with Harbor trailing around 71. Their estimated costs were dramatically lower, from $0.31 down to just $0.08 per 1,000 tickets. For simple FAQs or internal triage, that tradeoff might be acceptable. But a 5- to 13-point drop on a composite of correctness, usefulness, and tone is not noise. In support, that gap is the difference between a draft an agent lightly edits and one they have to rewrite or fact-check line by line, erasing any savings in inference cost. One weakness cut across all price tiers: tickets that depended on an image attachment. Screenshots of error messages, annotated photos, blurry invoices, all five models performed poorly when the crucial context lived in an image. Teams where visual evidence is common should not assume text-only performance will transfer. ### Limits and takeaway Context matters. All tickets were in English and came from a single software company, so results may not generalize to other industries, ticket complexity, or languages. The study did not test tool use, agentic actions like looking up an order, or multilingual support, all areas where rankings could shift. Even with those caveats, the practical advice is clear. Do not buy on headline score alone, and do not assume a tiny lead justifies a large premium. Pilot the top-value contenders on your own tickets, measure edit rates and resolution times, and pay attention to your mix of billing versus technical issues. In this test, excellence did not require the premium pick. It required recognizing a tie when you see one, and pocketing the difference.

DeepSeek: DeepSeek V4 Pro 0813: KiteBridge, an expense-management software provider, triggered $2.8 million in duplicate reimbursement payments Tuesday after a payment-processor timeout disabled a safeguard that was supposed to prevent exactly that kind of failure. The incident affected 91 corporate customers and left employees confused, employers covering duplicate payouts, and the company scrambling to recover funds while facing questions about the accuracy of its early public statements. The trouble began at 8:45 a.m., when KiteBridge deployed version 4.2 of its platform. The release was intended as a routine update, but within half an hour it collided with a problem in the payment-processing chain. When the company’s payment processor timed out, KiteBridge’s fallback system interpreted the stalled transactions as failures. Instead of holding them for verification, the system treated each retry as a new, legitimate reimbursement request. The result was a cascade of duplicates. Between the first customer complaint at 9:12 a.m. and the company’s decision to suspend automated payments at 10:03 a.m., the software generated roughly 13,400 duplicate reimbursements worth $2.8 million. The affected customers were corporate clients, not individual employees, and no employee had money pulled back from a personal account. Employers funded the duplicate payments, which meant companies were temporarily out of pocket while KiteBridge attempted to recover the funds. The failure was especially awkward because of what KiteBridge had promised. An April changelog for version 4.0 said the platform had introduced “exactly-once payment protection,” a phrase common in financial software that describes a system’s ability to ensure a transaction is processed once and only once. Tuesday’s incident showed that protection could be bypassed when a fallback path misjudged the state of a payment. KiteBridge’s public communications added to the problem. Its incident page initially described the issue as delayed reimbursements, a characterization that understated the scale and nature of the failure. It was not until six hours after automated payments were suspended that the company first disclosed the duplicate payments. By then, finance teams at affected businesses were already discovering the unusual payout activity on their own. By Wednesday evening, KiteBridge had recovered $2.3 million of the $2.8 million in duplicate payments. The remaining amount was still pending, and the company said it was working with clients and banking partners to complete the recovery. CEO Lena Ortiz apologized for the incident and said KiteBridge would publish a full technical report. The company has not yet said whether employees or finance teams should expect any changes to how reimbursements are reviewed before future releases. The incident highlights a familiar risk in financial automation: fallback systems designed to keep payments moving can become sources of error when they misread a processing failure. For KiteBridge, the challenge is no longer just recovering the outstanding funds. It is also rebuilding confidence among customers who were told the software would never let the same payment go through twice.

muse-spark-1.3: # KiteBridge Update Sends $2.8 Million in Duplicate Reimbursements **Ninety-one employers were left chasing double payments after a routine release broke a safeguard meant to prevent exactly this failure.** Dollars arrived twice in employee bank accounts on Tuesday after expense-management software maker KiteBridge deployed a flawed update that generated 13,400 duplicate reimbursements totaling $2.8 million. The incident began at 8:45 a.m. Tuesday when KiteBridge rolled out version 4.2 of its platform, which companies use to process and pay employee expense reports. At 9:12 a.m., less than half an hour after the deployment, the first customer complaint arrived, reporting that reimbursements appeared to be duplicating. The cause, according to the company, was a breakdown in how its system handles uncertainty. When KiteBridge sends a payment instruction to its outside payment processor, it expects a confirmation back. On Tuesday, those confirmations timed out, leaving the system unsure whether the money had moved. Its fallback system tried again — but instead of recognizing the retries as copies of the same request, it treated each one as a new, separate reimbursement. In practical terms, that meant an employee owed $200 for travel might receive $400, with the employer unknowingly funding both payments. KiteBridge emphasized that no money was withdrawn from employee accounts; the extra cash flowed outward from corporate accounts to workers who received more than they were owed. At 10:03 a.m., a little more than an hour after deployment, KiteBridge suspended all automated payments to stop further duplicates. By then, the scale was already substantial. The company's public messaging initially understated the problem. Its incident page first described the issue as delayed reimbursements, a framing that suggested payments were slow rather than doubled. Only six hours after automation was suspended did the page disclose that duplicate payments had been issued. That sequence has drawn attention because KiteBridge had previously promised protection against this exact scenario. An April changelog for version 4.0 said the platform had introduced "exactly-once payment protection," a technical term for a common payments safeguard that ensures a transaction is executed a single time even if computers retry, crash, or lose contact with each other. The episode highlights the fragile reliance modern payroll systems place on brief digital handshakes between companies, where a missed confirmation can cascade into thousands of errors in minutes. Recovery is now underway. By Wednesday evening, $2.3 million of the duplicate funds had been recovered, with the rest still pending as banks process reversals and finance teams reconcile their books. For affected finance departments, the error created days of extra work, matching duplicate transfers to original claims and asking banks to pull back funds already deposited in employee accounts. While most of the money has been returned, the outstanding balance underscores how difficult automated payments can be to unwind once they enter the banking system. Chief Executive Lena Ortiz has apologized to affected customers and said KiteBridge will publish a technical report detailing the failure and its fixes.

DeepSeek: DeepSeek V4 Pro 0813: # Arcwell Launches Relay to Speed Customer Support With AI Arcwell has launched Relay, an artificial-intelligence platform that examines customer support conversations and suggests answers to agents. The software also flags customers who may cancel and creates summaries of support cases, bundling three common support workflows into a single tool. The company said 12 organizations tested Relay over six months. During that pilot, agents using the platform closed tickets 14% faster than they did without it, according to Arcwell. The company did not release customer-retention results, so it remains unclear how well the churn-risk detection feature performs in production. Relay costs $65 per agent per month. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said in a statement. That premise is central to Relay’s design: the platform scans support transcripts and internal help documents, then surfaces relevant guidance while an agent is still working with a customer. Because Relay requires access to support transcripts and internal help documents, it enters a space where enterprise security and procurement teams tend to ask hard questions. Arcwell said customer data is encrypted and is not used to train shared models. Companies can choose between 30-day and 90-day retention periods, giving them some control over how long conversational data stays in the system. Arcwell is positioning Relay as an assistant rather than a replacement for human agents. The pilot points to a concrete efficiency benefit—14% faster ticket closures—while the lack of retention data means the platform’s broader effect on customer loyalty has yet to be demonstrated. For support leaders evaluating the tool, the data protections and retention options may matter as much as the speed gains, particularly in industries where chat transcripts contain sensitive customer information. That balance of speed and governance is likely to shape how quickly enterprises adopt the platform in the coming months.

muse-spark-1.3: # Arcwell Launches Relay, an AI Copilot for Support Agents Arcwell has launched Relay, an artificial intelligence platform that examines customer support conversations and suggests answers for agents to send. The idea, according to CEO Priya Shah, is to reduce the time agents spend hunting for material their employer has already produced. “Support agents spend too much time searching for information that their companies already have,” she said. In practice, that means Relay sits alongside the live chat or ticket queue, reads the exchange as it develops, and surfaces a proposed response drawn from prior transcripts and company documentation. The agent remains the sender, but starts from a draft rather than a blank box. Alongside suggested replies, Relay identifies customers who may be about to cancel and creates summaries of support interactions, reducing the need for agents to write up cases by hand after a difficult call. Arcwell said the system was tested by 12 companies over six months. Agents using Relay closed tickets 14% faster, the company said. What the pilot does not yet show is whether early warnings about cancellations changed outcomes. The company did not release customer-retention results, leaving prospective buyers without a direct measure of whether flagged accounts were ultimately saved. Relay costs $65 per agent each month, with companies paying for each support seat that uses the system. To operate, Relay requires access to a company’s support transcripts and internal help documents, the same records that make its suggestions relevant. That access will likely shape how quickly compliance and security teams approve a deployment. Arcwell said customer data is encrypted and is not used to train shared models. Buyers can choose a 30-day or 90-day retention period for that data, a control that determines how long transcripts and documents remain on Relay’s servers. How support organizations balance those speed gains against data access and unproven retention benefits will determine how broadly Relay is adopted. Arcwell’s bet is that faster resolutions and lighter after-call work will outweigh lingering questions about privacy and proof.

DeepSeek: DeepSeek V4 Pro 0813: # Cloudnote Left Debug Page With Customer Prompt Excerpts Open Without Authentication *The company says the page was accessible for four hours and logged excerpts from 1,842 user sessions before a researcher’s warning led to its removal.* Cloudnote confirmed Monday that an internal debugging page containing short excerpts of customer prompts was publicly accessible without authentication for approximately four hours. The company removed the page 19 minutes after a security researcher reported it, according to a statement. The exposed data consisted of logs from 1,842 user sessions, with each excerpt limited to as many as 200 characters from a prompt. Cloudnote said full documents, passwords, and payment information were stored in separate systems and were not present on the page. The service has about 3.2 million registered users, meaning the logged sessions represent a small fraction of its accounts, though the company did not say how many unique users were affected. The incident matters because prompts may contain personal, professional, or otherwise sensitive information, and even brief fragments can reveal private details or unfinished work. Cloudnote has not completed its investigation. It said it has found no evidence that anyone besides the researcher accessed the page, but that does not yet rule out other access. The company has not described what logging or access records are available and has not said whether it will notify affected users.

muse-spark-1.3: # Cloudnote Left Prompt Excerpts Exposed for Four Hours *Debugging page showed up to 200 characters from 1,842 user sessions; company says it found no evidence of wider access but investigation is ongoing.* Cloudnote left a debugging page containing excerpts of customer prompts accessible on the public internet without requiring a login for four hours on Monday, the company said. The company said its logs showed excerpts linked to 1,842 user sessions were present on the page, with each entry showing up to 200 characters from a prompt. The exposure affected a small fraction of Cloudnote's 3.2 million registered users, and the company said full documents, passwords and payment information are stored in separate systems and were not included on the page. Cloudnote said it removed the page 19 minutes after an outside security researcher notified it of the exposure on Monday. The takedown ended the four-hour period of exposure. The company said it has found no evidence so far that anyone other than the researcher accessed the page, but it said its investigation is not complete and it is still working to determine whether the page was viewed or copied by others. Even short excerpts of up to 200 characters can reveal sensitive personal or business information depending on what users typed, which makes the question of additional access central to the potential impact for affected users.