Head to head: Anthropic: Claude Opus 5.5 vs Meta: Muse Spark 1.3

Anthropic: Claude Opus 5.5: # 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.

Meta: 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.

Anthropic: Claude Opus 5.5: # Raincheck Raises $8 Million to Help Cities Prepare for Floods *The Austin startup's software combines drainage maps with rainfall forecasts to help local governments decide where to send maintenance crews before a storm.* Raincheck, an Austin company that makes flood-risk software for local governments, has raised $8 million in seed funding, the company announced today. The software maps drainage systems and combines those maps with rainfall forecasts. Cities use it to decide where to send maintenance crews before a storm. Basin Ventures led the round, and existing investor Long Mile also participated. Raincheck has now raised $10.5 million in total. CEO Ana Flores described the problem the company is trying to solve: > "Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives." The company's forecasts are recommendations, and emergency managers retain control over deployments and public warnings. Raincheck has published 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. Raincheck's competitors include the flood-planning vendor HydroMap and municipal engineering contractors. The company says the new funding will help it hire eight engineers and expand into Florida.

Meta: Muse Spark 1.3: # Raincheck raises $8 million for flood-risk software for cities *Austin company maps drainage systems and pairs them with rainfall forecasts to help local governments prioritize repairs before storms* Austin-based Raincheck has raised $8 million to expand its flood-risk software for local governments, the company announced today. 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. The seed round was led by Basin Ventures, with participation from existing investor Long Mile. The new round brings Raincheck’s total raised to $10.5 million. 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.” More detail on its approach is available in [Raincheck’s methodology overview](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. 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.

Anthropic: Claude Opus 5.5: # Startup Spotlight: LedgerLark Brings AI Agents to the Back Office Most conversations about AI in finance focus on trading: faster signals, smarter models, better predictions. LedgerLark is looking at what happens afterward. The startup builds AI agents for post-trade operations at small investment firms. This is the work of confirming that what was traded, what the custodian recorded and what sits in the firm's own books all match. It is unglamorous work, and that may be the point. ## A Three-Day Problem Co-founder Mina Batra spent seven years in operations at a regional brokerage before starting the company. She began exploring the idea after watching a six-person team spend three days resolving a single trade mismatch. The cause was a copied account number. The incident captures a familiar tension in post-trade work. The error was small and human, but finding it meant several people combing through records from different systems that were never designed to talk to each other. Larger institutions often have dedicated reconciliation infrastructure for this. Smaller firms may rely on spreadsheets, manual checks and the institutional memory of a few experienced staff. Batra's co-founder, Owen Carr, approaches the problem from the other side. He built reconciliation systems at a large bank. That background presumably gives the pair a view of both how the problem looks on the ground and how it has been solved at scale. ## What the Product Does, and Doesn't LedgerLark's agents compare trade confirmations, custodian records and internal ledgers, flag discrepancies, and propose corrections. A human then approves or rejects each one. The product does not execute trades or move money. That boundary matters. It limits what can go wrong if the software makes a mistake, and it may make the tool easier to explain to compliance teams and regulators. It also defines the product's value proposition clearly: LedgerLark is not trying to replace operations staff' judgment, but to shorten the time it takes them to reach a decision. Whether AI agents are meaningfully better at this than conventional rules-based matching is an open question. The answer likely depends on how messy a given firm's data is and how well the agents handle edge cases that don't follow predictable patterns. ## Early Traction 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 figure has not been independently verified, and a single customer's results may not generalize. Still, it points to the kind of measurable outcome operations teams are likely to care about. The company has raised a $3.5 million seed round led by Motive Row. Five customers is an early sample. The spread in contract values suggests LedgerLark is still learning how its pricing maps to customer size and complexity. ## Competition and Friction LedgerLark is entering a field that already has occupants. Established reconciliation vendors have long sold software to financial firms, and large banks frequently build their own internal tools. Carr's own career is an example of the latter. For LedgerLark, the challenge is less about convincing firms that reconciliation matters than about showing that an agent-based approach justifies switching from what they already use, or adopting something at all. Sales cycles are another hurdle. Financial institutions may require lengthy security and procurement reviews, even for software that never touches money directly. For a seed-stage company, long reviews can slow growth and stretch runway. Focusing on smaller firms may help, if those firms have lighter processes. But smaller firms also tend to have smaller budgets. ## What to Watch The useful questions for LedgerLark over the next year are practical ones: - Can it add customers at a steady pace despite procurement delays? - Do results like the reported drop in exceptions hold up across different firms? - Can it stay differentiated as incumbent vendors add their own AI features? LedgerLark has picked a narrow, concrete problem that its founders know well. The next stage will show whether that familiarity translates into a durable business.

Meta: Muse Spark 1.3: # Startup Spotlight: LedgerLark Most retail investors never think about what happens after they click buy. For small brokerages, wealth managers and fund administrators, that after is a daily grind of matching confirmations, custodian records and internal ledgers to make sure everything lines up. LedgerLark builds AI agents focused on that post-trade operations layer for small investment firms. The idea is not to replace operations staff, according to its founders, but to reduce the time they spend hunting across spreadsheets and PDFs for small discrepancies. ## An insight from the operations desk Co-founder Mina Batra spent seven years in operations at a regional brokerage before starting the company. Co-founder Owen Carr previously built reconciliation systems at a large bank. Their backgrounds cover both sides of the problem: living with broken processes, and trying to fix them from inside a large institution. Batra traces her interest to a specific incident: a six-person team spending three days resolving a trade mismatch that turned out to stem from a copied account number. It was the sort of error that was easy to make and surprisingly hard to trace, and it raised a question she kept returning to — how much of this detective work could be assisted by software that reads across systems? ## What the product does The company's agents compare trade confirmations, custodian records and internal ledgers, flag differences and propose corrections for a human to approve. The system does not execute trades or move money, a boundary the founders present as a deliberate design choice for risk-sensitive customers. In a demonstration, that looks like side-by-side source documents with highlighted fields and a suggested fix, plus an audit trail. The open question, common to agent-based tools, is how the system behaves with messy, inconsistent data and whether reviewers can quickly catch its mistakes without adding back the time it was meant to save. ## Early customers and funding LedgerLark says it has five paying customers, with annual contracts ranging from $36,000 to $110,000. That suggests the company is finding some willingness to pay among smaller firms, though five is still a small sample for judging repeatability. As one early data point, the company says one customer reduced its unresolved daily exceptions from 70 to 18 after adopting the product. That figure has not been independently verified, and it is unclear how representative it is, how long the improvement took, or what staffing changes accompanied it. Still, daily exceptions are a metric operations managers watch closely, so it will be worth tracking whether similar results appear at other sites. The company has raised a $3.5 million seed round led by Motive Row. For a company selling into financial services, that capital may need to cover not just product development but extended pilots and compliance work. ## Competition and what to watch LedgerLark is entering a crowded area. Established reconciliation vendors already serve many firms, and some larger organizations rely on internal software built over years. Small firms may have less entrenched systems, which could make them more open to a new tool — but they also have smaller budgets and limited capacity to test new vendors. Another factor is procurement. Financial institutions may require lengthy security and procurement reviews before adopting tools that touch sensitive records. How quickly LedgerLark can get through those reviews, and whether its human-in-the-loop approach eases compliance concerns, will likely shape its pace of growth as much as model performance. For now, the question is not whether back-office errors are annoying — they are — but whether an agent that proposes fixes can earn enough trust to stay in daily use. LedgerLark has a plausible starting point and early paying users. The next evidence to look for is broader deployment and retention.

Anthropic: Claude Opus 5.5: # HelioDock Launches Night Shift Software for Remote Oversight of Warehouse Robot Fleets *The autonomous robot maker also closed a $24 million Series B, though a beta customer says the system's usefulness depends on a reliable network connection.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, software that lets operators monitor and manage robot fleets remotely. The company also announced a $24 million Series B funding round led by North Coast Ventures. The launch targets a familiar problem in automated warehouses: a single robot stalls, and someone has to show up to fix it, often outside regular hours. HelioDock says Night Shift reduced unplanned robot stoppages by 17% during a 14-month private beta, a figure that has not been independently verified. Its value also appears tied to network conditions. One beta customer said the system worked well in well-connected facilities but became less useful during connectivity outages at rural sites. Chief Executive Mara Velez framed the product around the people who get called in when automation stalls. "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. ## A 14-month beta across nine companies HelioDock developed Night Shift through a private beta involving nine logistics companies and 310 robots. That works out to an average of roughly 34 robots per participant, though the company did not break down how the machines were distributed across customers. The 17% reduction in unplanned stoppages is the company's headline performance claim. HelioDock attributes it to beta customers' experience with the software. No outside party has confirmed the number. ## Connectivity sets the limits Aaron Pike of Fulcrum Logistics, one of the beta customers, offered a more qualified assessment. He said Night Shift performed well in facilities with strong connectivity but was less useful during LTE outages at rural locations. That limitation follows from how the system is designed. When a connection fails, HelioDock's robots return to their locally configured behavior, and remote control becomes unavailable. In practice, the robots do not simply stop when the network drops, but operators lose the ability to intervene from afar until the connection returns. For logistics companies whose facilities sit in areas with dependable networks, the trade-off may matter little. For operators with sites that rely on rural LTE, Pike's experience suggests that remote oversight could be least available in the places where a long drive to the warehouse is most likely. ## Pricing and availability Night Shift will cost $90 per robot per month and will become generally available on September 15. At that rate, the software would cost about $1,080 per robot annually. A fleet the size of the full beta, 310 robots, would cost $27,900 a month, or about $334,800 a year. Customers weighing the price will likely compare that expense against the cost of downtime and after-hours staff visits, the scenario Velez described. ## New funding The Series B brings HelioDock's total funding to $39 million, meaning the company had raised $15 million before this round. North Coast Ventures led the new financing. HelioDock announced the raise alongside the product launch but did not specify in its announcement how it plans to allocate the money. ## A competitive market HelioDock enters a field where warehouse operators already have options. Relay Robotics and YardOS both offer competing fleet-management products, so Night Shift will need to win customers who are evaluating alternatives rather than buying remote oversight for the first time. The company's pitch rests on two claims buyers will be able to test once the product reaches general availability: that it meaningfully cuts unplanned stoppages, and that it saves managers from overnight trips to the warehouse. Pike's comments suggest that the answer may vary by site, depending largely on how reliably each facility stays online.

Meta: Muse Spark 1.3: # HelioDock Introduces Night Shift to Let Operators Watch Warehouse Robots From Afar *The autonomous warehouse robot maker says 14 months of beta testing reduced unplanned stoppages, as it adds $24 million in new funding and sets September availability.* HelioDock, which makes autonomous warehouse robots, on Thursday launched Night Shift, software that allows operators to monitor fleets of its robots remotely instead of responding to problems on site in the middle of the night. The release pairs a new paid offering with fresh capital. Night Shift will be generally available September 15 at $90 per robot per month. HelioDock also announced a $24 million Series B led by North Coast Ventures, bringing its total funding to $39 million. Chief executive Mara Velez framed the product around overnight disruptions. “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 spent 14 months in private beta with nine logistics companies and 310 robots, according to HelioDock. The company says beta customers experienced 17% fewer unplanned robot stoppages during that period. That figure has not been independently verified. For operators, the appeal described by the company centers on watching a fleet from elsewhere when an issue arises outside normal hours. Rather than traveling to a warehouse to check on a stopped machine, a manager could assess the fleet remotely. The scenario Velez described — a single confused robot triggering a late-night trip — illustrates the kind of interruption Night Shift is meant to handle remotely. HelioDock presents remote monitoring as a way to evaluate such stoppages without an on-site visit. The beta scale gives HelioDock nine operators’ experience to cite at launch. With 310 robots involved, the test covered multiple fleets rather than a single warehouse deployment. HelioDock says those customers experienced the 17% reduction in unplanned stoppages. HelioDock provided the 17% figure as a company-supplied result from the beta. The company also noted the private beta context, and the figure has not been independently verified. Aaron Pike of Fulcrum Logistics, a beta customer, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His account suggests network reliability shapes how consistently remote monitoring can be used from site to site. HelioDock said that when connectivity fails, robots return to their locally configured behavior. At those times, remote control becomes unavailable. An operator who loses a connection cannot intervene from afar until service is restored and must rely on whatever behavior was configured locally in advance. Night Shift will cost $90 per robot per month. For an operator, the monthly charge scales with fleet size, so a deployment with more robots would carry a higher recurring cost than a smaller one. General availability is set for September 15, following the 14-month private beta. The Thursday announcement identifies September 15 as the date for general availability. That date marks the end of the private beta period and the start of sales at $90 per robot per month. The $24 million Series B led by North Coast Ventures brings HelioDock’s total funding to $39 million, the company said. The financing was disclosed alongside the Night Shift launch. HelioDock is entering an area where competing fleet-management products are already available from Relay Robotics and YardOS. HelioDock’s September 15 date gives operators a timeline for when Night Shift will move beyond the nine-company beta to broader availability at the stated per-robot price.

Anthropic: Claude Opus 5.5: **6** 1. `#Flood software moves upstream` has no space after `#`, so it breaks the `# Headline` format. 2. `The company is expanding its pilot.` is the dek, but it is not italicized. 3. `##Details` has no space after `##`, so it breaks the `## Heading` format. 4. `Read [the announcement](www.example.com/news).` uses a URL that is not absolute HTTPS. 5. `* The first pilot begins Monday.` starts the list item with an asterisk instead of a hyphen. 6. `<div>The second pilot begins Friday.</div>` is raw HTML.

Meta: Muse Spark 1.3: 6

Anthropic: Claude Opus 5.5: # Arcwell's Relay Uses AI to Help Support Agents Find Answers Faster Arcwell is betting that customer support agents already have most of the answers they need. The problem, it says, is finding them. The company has launched Relay, an artificial intelligence platform that analyzes customer support conversations and suggests responses to agents. Relay can also generate summaries and flag customers who may be at risk of canceling, putting it in direct competition with a growing number of AI tools aimed at support teams. "Support agents spend too much time searching for information that their companies already have," said Priya Shah, Arcwell's chief executive. Arcwell says 12 companies tested Relay over six months and that agents using it closed tickets 14% faster. That number comes from the company itself. Arcwell did not release customer-retention results from the trial, leaving one of Relay's most commercially appealing features without public evidence. Faster ticket handling is useful, but the ability to identify customers who may leave could matter more to businesses trying to reduce churn. Relay costs $65 per agent each month, so a team of 100 agents would pay $6,500 a month. To work, Relay needs access to companies' support transcripts and internal help documents. That requirement may give some buyers pause, since transcripts can contain customer details and internal documents may include information companies prefer to keep private. Arcwell says customer data is encrypted and is not used to train shared models. Companies can also choose whether to keep data for 30 days or 90 days. Those safeguards address some obvious privacy concerns, but they also highlight the trade-off Relay asks customers to accept. To get faster answers, companies must give the software access to the records and documents their support teams rely on. For Arcwell, the pitch is simple: help agents spend less time searching and more time helping customers. Whether Relay can also reduce customer losses remains an open question until the company shares retention results.

Meta: Muse Spark 1.3: # Arcwell Launches Relay, an AI Assistant for Support Agents Arcwell has launched Relay, a platform that uses artificial intelligence to examine customer support conversations and suggest answers to agents. The tool is designed to work alongside agents rather than replace them. As an exchange unfolds, Relay reads the dialogue and surfaces a proposed reply drawn from material the company already holds. It can also identify customers who may cancel, highlighting conversations that may need a more experienced hand, and create summaries when an interaction ends, reducing the afterwork of documenting what was decided. 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 it remains unclear whether flagging at-risk customers translated into fewer cancellations. Relay costs $65 per agent each month. The pricing follows the familiar per-agent model for support software, tying cost directly to team size. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. The comment points to the problem Relay is meant to solve: answers are often buried in help articles and prior cases, and retrieving them mid-conversation slows resolutions and frays patience on both sides. Operating that way requires broad access. Relay requires access to support transcripts and internal help documents. Arcwell says customer data is encrypted and is not used to train shared models, a distinction intended to reassure buyers concerned about sensitive dialogue being repurposed elsewhere. Companies can choose a 30-day or 90-day retention period, which determines how long transcripts and related material remain in the system. For Arcwell, Relay is a bid to improve customer service using artificial intelligence not by adding another dashboard, but by shortening the distance between a customer’s question and an answer the company has already written. Whether that speed holds up outside the pilot will depend on how teams integrate the suggestions into daily work.

Anthropic: Claude Opus 5.5: # KiteBridge Software Update Sent $2.8 Million in Duplicate Reimbursements A software update at expense-management company KiteBridge sent $2.8 million in duplicate reimbursements to employees at 91 corporate customers on Tuesday. The failure is the kind the company had said, months earlier, its system was built to prevent. The update, version 4.2, generated 13,400 duplicate payments before KiteBridge halted automated payments. As of Wednesday evening, the company had recovered $2.3 million. The remaining $500,000 was still pending. ## How the duplicates happened KiteBridge's software lets companies reimburse employees for business expenses such as travel, meals, and supplies. To move that money, the software sends payment instructions to an outside payment processor. On Tuesday, some of those instructions timed out, meaning the processor did not respond within the expected window. Payment systems commonly handle timeouts by retrying the request. In this case, KiteBridge's fallback system, the backup logic meant to handle such interruptions, treated each retry as a brand-new reimbursement request rather than a second attempt at the original one. Employees who were owed one reimbursement could receive two or more. The failure is notable because of an April changelog, the running list of updates the company publishes with each release. That changelog said version 4.0 had introduced "exactly-once payment protection." The term describes safeguards designed to ensure that a payment goes through once and only once, even when a request is retried after a network hiccup. KiteBridge has not yet explained why that protection did not stop the duplicates in version 4.2. ## A morning of compounding payments KiteBridge deployed version 4.2 at 8:45 a.m. Tuesday. The first customer complaint arrived less than half an hour later. Automated payments kept running for nearly another hour, until the company suspended them at 10:03 a.m. In all, the faulty system ran for about 78 minutes. The duplicate payments came from employers, not employees. KiteBridge said no workers had money withdrawn from their own accounts. The 91 affected companies funded the extra reimbursements, and they bore the cost while recovery efforts continued. ## Slow disclosure KiteBridge's public account of the incident lagged behind events. The company's incident page, where it posts updates on outages and service problems, initially described the issue as delayed reimbursements. That description was accurate in one narrow sense, since payments stopped once automation was suspended. But it left out the core problem. The page did not disclose the duplicate payments until six hours after automation was halted, around 4 p.m. Tuesday. During that time, customers relying on the status page would not have learned that money had already gone out in error. ## What comes next KiteBridge CEO Lena Ortiz apologized for the incident and said the company would publish a technical report. She did not say when it would be released. Several questions remain open: why the exactly-once protection described in April failed, what changed in version 4.2, why payments continued for nearly an hour after the first complaint, and why the duplicates were not disclosed sooner. It is also unclear how and when KiteBridge expects to recover the outstanding $500,000, and who will absorb any amount that cannot be retrieved.

Meta: Muse Spark 1.3: # Duplicate Reimbursements Cost KiteBridge Customers $2.8 Million After Failed Update **91 employers hit; $2.3 million recovered** Employers using expense-management software from KiteBridge woke up Tuesday to a costly surprise: their workers had been paid twice, with a software bug turning routine reimbursement requests into $2.8 million in duplicate payments funded by the companies themselves. In total, the error generated 13,400 extra payments affecting 91 corporate customers, leaving finance teams scrambling to trace money already in employees’ accounts. ## A safeguard billed as foolproof KiteBridge sells software that handles the plumbing of corporate expenses: employees submit receipts, managers approve them, and the system sends money through a payment processor to reimburse staff. In April, the company’s changelog for version 4.0 said it had introduced “exactly-once payment protection,” a safeguard designed to ensure that even if a payment attempt had to be retried, it could only ever result in a single transfer. That protection failed after the company deployed version 4.2 at 8:45 a.m. on Tuesday. According to the company’s preliminary account, a timeout at its outside payment processor set off the failure. When KiteBridge did not get a timely confirmation that a payment had gone through, its fallback system tried again — but instead of recognizing the retry as a continuation of the same request, it treated each retry as a brand-new reimbursement. ## How retries became new payments In practice, online payments rarely travel in a straight line. If a processor is slow to respond, the sending software cannot tell whether the money is still on its way or never left at all, so it retries to avoid leaving an employee unpaid. To prevent double-paying, systems typically attach a fingerprint, often called an idempotency key, to each payment. The safeguard KiteBridge had advertised is supposed to work that way: duplicates with the same fingerprint are ignored. On Tuesday, that logic did not hold, and retries created fresh payments rather than being filtered out. The first sign of trouble came from customers, with the first complaint arriving at 9:12 a.m. as finance teams noticed employees receiving second deposits for expenses they had already been reimbursed for. KiteBridge suspended automated payments later that morning at 10:03 a.m., stopping further duplicates but leaving thousands of erroneous transfers already in motion. ## Response outpaced explanation Public communication lagged behind the operational response. KiteBridge’s public incident page initially described the problem as delayed reimbursements, suggesting workers would be paid late rather than twice. The page did not disclose that duplicate payments had been issued until six hours after automation was suspended, by which time employers had spent the day reconciling unexpected payouts without an explanation for the extra money. ## Recovery and reckoning By Wednesday evening, KiteBridge said $2.3 million had been recovered, with the remainder still pending in the banking system. The company emphasized that no employees had money withdrawn from their accounts. Instead, the duplicates were paid out of employer funds, meaning companies — not workers — bore the shortfall while reversals were processed. Chief executive Lena Ortiz has apologized for the incident and said the company will publish a technical report explaining how a system billed as exactly-once could double-pay at scale. KiteBridge says automated payments remain paused while it completes recovery of the outstanding funds and reviews its retry logic. The episode underscores a risk in financial software: a routine update combined with a network hiccup can defeat a single safeguard — and leave dozens of businesses to clean up the consequences.

Anthropic: Claude Opus 5.5: # Cloudnote Page Exposed Prompt Excerpts From 1,842 User Sessions *A debugging page was reachable without a password for four hours Monday. The company says it has found no sign that anyone other than the researcher who reported it viewed the data, but its investigation is not finished.* An internal debugging page at Cloudnote that displayed fragments of customer prompts was accessible to anyone, without authentication, for four hours on Monday. The company removed the page 19 minutes after a security researcher notified it of the problem. Cloudnote's logs show the page held excerpts from 1,842 user sessions, each containing up to 200 characters of a prompt. Full documents, passwords and payment information were stored separately and did not appear on the page. Cloudnote has 3.2 million registered users. The exposure was limited in size and duration, but short excerpts can still be sensitive. Depending on what someone typed, 200 characters can be enough to include a name, a medical concern or confidential work details. Cloudnote says it has found no evidence that anyone besides the researcher accessed the page. It has not completed its investigation, however, and several questions remain unresolved. It is not yet known whether anyone else viewed the page, how many individual users the 1,842 sessions represent, or how a debugging tool came to be publicly reachable in the first place.

Meta: Muse Spark 1.3: # Cloudnote Left Customer Prompt Excerpts Exposed for Four Hours *Excerpts from 1,842 sessions were visible without a login; the company says it sees no sign of wider access but its investigation continues* Cloudnote left a debugging page that displayed excerpts of customer prompts accessible on the open internet without a password for four hours on Monday, exposing snippets from user sessions before taking the page down 19 minutes after being notified by a researcher. Logs reviewed by the company show the page contained excerpts of up to 200 characters from 1,842 user sessions, a small fraction of Cloudnote’s 3.2 million registered users but still a sizable cache of potentially sensitive material that did not require authentication to view. The incident matters because prompts entered into note-taking and AI tools can include private plans, business details and other personal information, even in a brief snippet. Cloudnote said full documents, passwords and payment information are stored in separate systems and were not present on the exposed page, limiting but not eliminating the potential harm. The company says it has found no evidence so far that anyone other than the researcher who reported the issue accessed the page, but it said its investigation is not complete. It is still unclear whether the excerpts were indexed, copied or shared during the exposure window, and Cloudnote has not said when affected users will be notified.

Anthropic: Claude Opus 5.5: C

Meta: Muse Spark 1.3: C

Anthropic: Claude Opus 5.5: # Foundry Hall, Milwaukee's Startup Launchpad, Will Close Oct. 31 Foundry Hall, the nonprofit accelerator that has helped launch more than 200 Milwaukee startups since 2018, will close its doors October 31 after its largest corporate sponsor declined to renew its support. Blank River Bank had supplied 38% of the accelerator's annual budget. Without that money, Foundry Hall can cover its costs through the end of October but cannot fund another full program, according to executive director Camille Ross. The closing will eliminate six staff positions. It will also displace the 12 startups now working out of the building, which must find somewhere else to go in a matter of weeks. Ross said the organization spent months trying to close the gap, hoping another sponsor would step in before the money ran out. "We kept waiting for the replacement check that never came," she said. A spokesperson for Blank River Bank said the decision reflected a shift in the bank's community investment priorities and had nothing to do with how Foundry Hall performed. By most measures, the accelerator performed well. The 212 startups it has supported now employ a combined 1,140 people. Those jobs represent payrolls, leases and tax revenue that did not exist in the city before Foundry Hall opened, and the companies behind them are scattered across Milwaukee's neighborhoods, many still growing. For many founders, though, the building's value was harder to capture in a spreadsheet. It offered a desk, a mailing address and a room where an early idea could be treated as a real business. Dev Malik, one of the founders who came through the program, remembers what came before. "My first investor meeting happened in that building," Malik said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." That kind of starting point is what the current tenants stand to lose. For the 12 companies still inside, the questions are now practical ones: where to set up next, what it will cost and whether they can find a place that offers more than four walls. City officials are trying to soften the landing. Deputy Mayor Susan Cole said the city has begun talks with building owners and local universities about finding space for the companies that must leave. No agreements have been announced. What the city is not yet discussing is a replacement for the accelerator itself, the programming, the introductions and the steady supply of first chances that made the building matter to people like Malik. Until the end of the month, the lights stay on. Staff will keep working with the companies that remain, and the conference rooms will still fill for meetings. Then, on October 31, the place that served as a first office for more than 200 young businesses will go quiet.

Meta: Muse Spark 1.3: # Foundry Hall to Close in October **MILWAUKEE —** Foundry Hall, the nonprofit startup accelerator that became a fixture of the city's early-stage business scene, will close on Oct. 31, ending a seven-year run that helped launch more than 200 companies. The accelerator, which opened in 2018, has supported 212 startups that collectively employ 1,140 people. Its downtown workspace, with its shared desks, glass-walled meeting rooms and calendar of mentor sessions, became a first address for founders who otherwise had nowhere to go. Former participants often describe it less as an office than as a proving ground, where a pitch could be tested over lunch and revised by evening. Over time, its demo days drew investors from across the Midwest, giving Milwaukee founders a stage they had long lacked. That work will end this fall after its largest corporate backer withdrew. Blank River Bank, which provided 38% of Foundry Hall's annual budget, declined to renew its sponsorship. The bank said the move stemmed from shifting community investment priorities, not from dissatisfaction with the accelerator's performance. Executive director Camille Ross said staff spent months searching for replacement funding. “We kept waiting for the replacement check that never came.” While Foundry Hall has enough cash to operate through October, Ross said it cannot finance another full program cycle. The closure will eliminate six staff positions and displace 12 startups now working out of its building. The small staff carried much of that weight. Coaches reviewed business plans late into the night, made introductions investors trusted, and calmed founders through missed deadlines and failed prototypes. Their departure, Ross acknowledged, will leave a gap not easily measured in jobs or dollars. For many alumni, the news landed as a quiet loss rather than a sudden shock. Founder Dev Malik recalled holding his first investor meeting inside Foundry Hall after months of pitching from a coffee shop, a step he described as giving his company legitimacy. In its last month, the hallways remain busy, if quieter. Whiteboards still carry half-erased revenue projections. Boxes have begun to appear beside desks. No one, staff say, is quite ready to wipe the boards clean. City officials said they are trying to soften the disruption for those current tenants. Deputy Mayor Susan Cole said the city has begun talking with building owners and local universities about finding interim space. Ross said details of the final weeks, including any closing events, will be announced later. For now, she said, the focus is on helping founders and employees make what comes next a little less abrupt.

Anthropic: Claude Opus 5.5: # QueueZero's AI Meeting Assistant Assigns Tasks to Its Own CEO During Launch **Proxy, which attends meetings and delegates follow-up work to software agents, enters public beta next week at $30 per user per month.** QueueZero launched Proxy, an AI meeting assistant, on Wednesday. Its demonstration offered an unscripted look at how the software handles the informal language common in meetings. Proxy is designed to attend meetings, produce notes and assign follow-up work to software agents. During the launch demonstration, a customer in the audience remarked, "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. ## A Question of Interpretation The exchange lasted only moments, but it pointed to a central challenge for AI tools that act on what they hear. Meeting conversation is full of asides, jokes and offhand comments. A system that converts speech into assigned work must decide which statements are actual requests, who made them and who is responsible for carrying them out. In this case, a single sentence contained both a plausible business request and a figure of speech. Proxy treated them the same way, producing a legitimate action item alongside one that was never meant literally. The pricing sheet, at least, was delivered promptly. ## Safeguards and Controls QueueZero says Proxy requires human confirmation before it sends any external message. Internal tasks, however, can be created automatically. Under that design, a misread comment could produce an internal assignment without anyone approving it. It could not, according to the company, produce an outgoing message without approval. Administrators can disable automatic task creation. That option may matter to organizations that want Proxy's notes and suggestions but prefer that a person decide what becomes assigned work. The tasks created during the demonstration were visible and easy to correct. Wu reviewed both and removed the one that did not apply. The incident nonetheless showed why the default settings may shape how such tools perform in practice, where automatically generated tasks may receive less scrutiny than they did on a launch stage. ## Pricing and Availability Proxy costs $30 per user per month. It enters public beta next week. The beta will likely show how often Proxy turns everyday conversation into work that no one intended to assign. It may also show how many organizations keep automatic task creation turned on once they find out.

Meta: Muse Spark 1.3: # QueueZero Launches AI Meeting Assistant Proxy **SAN FRANCISCO —** QueueZero on Wednesday launched Proxy, an AI meeting assistant designed to attend virtual meetings, produce notes and delegate follow-up work to software agents. The product, demonstrated live onstage at the company's developer event, is intended to reduce manual coordination by listening to conversations, summarizing decisions and creating tasks without human intervention. During the demonstration, the system misinterpreted an audience comment as a directive. 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 transcribed the comment in real time and generated two tasks assigned to QueueZero CEO Evan Wu: “Stop talking” and “Send pricing sheet.” The tasks appeared on screen alongside meeting notes visible to attendees. Wu laughed, completed the second task and deleted the first. He then sent the pricing information to attendees and continued the presentation, using the moment to illustrate how Proxy handles spoken requests. QueueZero said the incident reflects intended behavior for internal workflows. According to the company, internal tasks can be created automatically from meeting discussion without additional approval. For communications outside an organization, additional safeguards apply. QueueZero says Proxy requires confirmation before an external message is sent, meaning drafts involving customers, vendors or other third parties must be reviewed by a user. Administrators can disable automatic task creation entirely, the company said, allowing IT teams to require manual review before any task is generated or assigned. The setting is available at both the organization and team level. Proxy costs $30 per user per month. It enters public beta next week, when existing QueueZero business customers will be able to enable the assistant for video calls and in-person meetings with transcription. Analysts said meeting assistants have become a competitive category as vendors seek to connect transcription with workplace automation. QueueZero is positioning Proxy as distinct by linking notes directly to agent-based task execution rather than stopping at summaries. The company said further controls, including audit logs and permission-based task assignment, will be available during the beta period. The company did not disclose additional enterprise pricing details. QueueZero, founded in 2021, currently provides scheduling and transcription tools to more than 2,000 business customers worldwide. The company said Proxy integrates with its existing calendar and task management products.

Anthropic: Claude Opus 5.5: # The Top Two Support Models Tied on Quality. One Cost 61% Less. When teams pick a language model to draft customer-support replies, the instinct is to choose whichever one tops the leaderboard. A new benchmark of five models suggests that instinct can be expensive. The highest-scoring model, North, finished essentially level with the runner-up, Vale. But Vale cost less than half as much to run and answered faster. ## How the test worked Researchers gave five language models the same 200 anonymized customer-support tickets. Each model produced one proposed answer per ticket. Three experienced support managers then graded every answer without knowing which model wrote it. Blind grading matters here, because it reduces the chance that reputation or familiarity with a particular model's style influenced the scores. Each answer received a combined score reflecting three qualities: - **Factual correctness:** whether the answer was accurate. - **Usefulness:** whether it actually helped resolve the customer's problem. - **Tone:** whether it read like an appropriate support reply. The researchers also estimated what each model would cost to run across 1,000 tickets. ## The headline: a tie at the top, with a large price gap North earned the highest average score at 84.2. Vale followed at 83.8. The researchers found that this 0.4-point difference was **not statistically significant**. With 200 tickets and three graders, the gap is small enough that it could plausibly reflect noise in the sample rather than a real difference in quality. In practical terms, the benchmark cannot say that North writes better answers than Vale. The cost difference is much clearer: | Model | Avg. score | Est. cost per 1,000 tickets | |---|---|---| | North | 84.2 | $1.90 | | Vale | 83.8 | $0.74 | | Ember | 79.1 | $0.31 | | Kite | 76.4 | $0.22 | | Harbor | 71.0 | $0.08 | At $0.74 per 1,000 tickets, Vale costs about 39% of North's $1.90, a reduction of roughly 61%. Vale also responded 18% faster than North. For a support team, faster drafts can mean shorter queues and less time agents spend waiting. For organizations choosing a general-purpose drafting model, Vale looks like the stronger value. It offers comparable measured quality at a substantially lower price and with better speed. ## Where the two leaders differed The overall tie hides some differences in strengths. North performed best on **billing disputes**, and Vale performed best on **technical troubleshooting**. That pattern could matter to a company whose ticket volume leans heavily one way. A business dominated by billing complaints might reasonably test North more closely, while a software company fielding mostly bug reports might favor Vale. Readers should still treat these category results with caution. Splitting 200 tickets into categories leaves fewer examples in each one, and the results summary does not say whether these category-level differences were statistically significant. They are best read as leads to investigate, not settled conclusions. ## The cheaper tiers Below the top two, scores drop more noticeably. Ember (79.1) and Kite (76.4) sit several points behind the leaders, and Harbor trails at 71.0. These models are far cheaper. Harbor costs just $0.08 per 1,000 tickets. Whether that trade-off is worthwhile depends on how answers are used. If human agents review and edit every draft, a lower-scoring but very cheap model might still save time. If drafts go to customers with little oversight, the quality gap carries more risk. ## A shared blind spot: image attachments One weakness appeared across the board. **All five models performed poorly when a ticket depended on an image attachment**, such as a screenshot of an error message. In real support queues, screenshots are common. Any team deploying these models should plan to route image-dependent tickets to humans, or test that scenario specifically, regardless of which model it chooses. ## What this benchmark doesn't tell you The results come with important boundaries: - **One company, one language.** All tickets were in English and came from a single software company. Performance could differ for other industries, products or customer bases. - **No multilingual testing.** Teams supporting customers in other languages will need their own evaluation. - **No agentic actions or tool use.** The models only drafted replies. They did not issue refunds, look up accounts or take actions in other systems, which many modern support deployments require. - **One answer per ticket.** Models can vary from run to run, and a single sample per ticket does not capture that variation. ## The takeaway The most useful lesson from this benchmark is not which model ranked first. It is that ranking first was not meaningfully better than ranking second, while costing about 2.6 times as much. Before paying a premium for the top of a leaderboard, teams should check whether the lead is real. Then they should run a test on their own tickets, with special attention to the image-dependent cases that every model here struggled to handle.

Meta: Muse Spark 1.3: # The Tie That Matters: In Support AI, Cheap and Fast Just Tied for First Customer-support leaders shopping for AI usually start with a simple question: which model writes the best answer? A new benchmark suggests a better question: which model writes a good enough answer for far less money, and faster? Researchers tested five language models — dubbed North, Vale, Ember, Kite and Harbor — on 200 anonymized customer-support tickets drawn from one software company. Each model produced one proposed answer per ticket. Three experienced support managers graded every answer blind, without knowing which system wrote it. Their scores combined three things customers actually care about: factual correctness, usefulness, and tone. ## A dead heat, by design On overall quality, the race at the top was effectively a tie. North averaged 84.2, Vale averaged 83.8. But the researchers found that difference was not statistically significant, which means it could easily be chance. For anyone deploying one of these systems, the takeaway is blunt: on this kind of English-language support work, judges could not distinguish North from Vale. That parity changes how to read everything else. If two systems deliver the same quality, cost and speed stop being footnotes and become the decision. ## When equal quality, price wins Here Vale wins clearly. Its estimated cost was $0.74 per 1,000 tickets, compared with $1.90 for North — less than 40% of the price. It was also 18% faster to answer. At support scale, where a midsize team can burn through hundreds of thousands of tickets a year, paying more than twice as much for no measurable gain in quality is difficult to justify, and waiting longer for each draft adds queue time no manager wants. The rest of the field illustrates the real price of going cheaper still. The three lower-cost models all landed noticeably lower on quality, dropping into the 70s while the leaders sat in the mid-80s. Their running costs were strikingly low — Harbor, the cheapest, came in at just cents per thousand — but the score gap was large enough to suggest customers would notice in the form of less accurate, less complete, or less polished responses. In other words, Vale looks like a sweet spot in this test, not proof that cheapest is always best. There were hints of specialization beneath the averages. North looked strongest on billing disputes, where precision and careful phrasing matter. Vale looked strongest on technical troubleshooting, where step-by-step usefulness counts. Those patterns are interesting for teams whose queues skew one way or the other, but they come from slices of an already modest 200-ticket sample, so they should guide pilot testing, not purchase orders. More telling was a weakness all five models shared: when a ticket depended on an image attachment — a screenshot of an error, a photo of a broken layout — performance fell apart. Text-only drafts will always struggle there, and human judges noticed. Any team whose customers regularly send screenshots should treat these scores as an upper bound, not a promise. ## What the test can't tell you That caveat points to the study's broader limits. It covered English tickets from a single software company, scored as one-shot drafts. It did not test what modern support systems increasingly do: take agentic actions, call tools, look up accounts, issue refunds, or work across languages. It also relied on manager judgment, not on what happened after the answer was sent — did the customer reply, reopen, churn? Even with those limits, the practical message holds. Don't chase a 0.4-point lead. Test the two leaders in your own queue, measure resolution and handling time, and weigh cost and latency as first-class metrics. In this benchmark, the smart money isn't on the model that technically scored highest. It's on the one that tied it for quality while costing less than half as much and answering nearly a fifth faster.