Head to head: Anthropic: Claude Fable 5 vs OpenAI: GPT-5.6 Luna

Anthropic: Claude Fable 5: # Raincheck Raises $8 Million to Help Cities Get Ahead of Flooding *The Austin startup's software tells local governments which drains, pumps and repairs to prioritize before a storm hits.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million in seed funding. The company maps municipal drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. The round was led by Basin Ventures, with participation from existing investor Long Mile. It brings Raincheck's total funding to $10.5 million. CEO Ana Flores said the product is aimed at a gap between knowing where flooding happens and knowing what to do about it: > Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives. The company has published details of [how its forecasting model works](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. It declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations only; emergency managers retain control over deployments and public warnings. Raincheck competes with other flood-planning vendors, including HydroMap, as well as municipal engineering contractors. The company says the new funding will help it hire eight engineers and expand into Florida.

OpenAI: GPT-5.6 Luna: # Raincheck Raises $8 Million to Expand Flood-Risk Software for Cities *The Austin company maps drainage systems and combines them with rainfall forecasts to help local governments prepare for storms.* Raincheck, an Austin-based company that makes flood-risk software for local governments, has raised $8 million in a seed round led by Basin Ventures. Existing investor Long Mile also participated, bringing the company’s total funding to $10.5 million. Raincheck maps drainage systems and combines those maps with rainfall forecasts. Cities use the software to decide where to send maintenance crews before a storm. “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,” CEO Ana Flores said. The company provides more detail about its approach in its [methodology documentation](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. ## Competition Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Anthropic: Claude Fable 5: # Arcwell Bets on AI Assistant to Speed Up Customer Support Arcwell has launched Relay, an AI-powered platform that reads customer support conversations and drafts suggested responses for agents — the company's entry into an increasingly crowded race to automate the help desk. Beyond recommending replies, Relay flags customers who show signs of canceling and generates summaries of support interactions, giving agents a running digest of long or complicated cases. "Support agents spend too much time searching for information that their companies already have," said CEO Priya Shah, framing the product as a fix for institutional knowledge that sits buried in transcripts and documentation. Arcwell has some early numbers to back the pitch. Twelve companies piloted Relay over six months, and the company says agents using the tool closed tickets 14% faster. What Arcwell hasn't shared is whether the platform's churn-detection feature actually kept customers from leaving — no retention results were released, leaving one of Relay's headline capabilities unproven for now. The pricing is straightforward: $65 per agent per month. The harder question for prospective customers may be data access. Relay works by ingesting support transcripts and internal help documents — precisely the kind of material companies are most reluctant to hand to a third party. Arcwell says customer data is encrypted and, notably, is not used to train shared models, an assurance aimed at businesses wary of their support logs improving a competitor's suggestions. Customers can also choose how long their data sticks around, with retention periods of either 30 or 90 days. Those guardrails reflect a broader tension in the market. AI tools for customer service are only as useful as the data they can see, but every additional integration widens a company's exposure. Relay's value proposition depends on buyers deciding the trade is worth it. For Arcwell, the launch is a test of whether measurable efficiency gains — a 14% improvement is meaningful at scale — can overcome that hesitation. The six-month pilot suggests the speed benefits are real. Whether Relay can deliver on the retention promise, and whether Arcwell will eventually publish numbers to prove it, remains to be seen.

OpenAI: GPT-5.6 Luna: # Arcwell Launches Relay AI Platform for Customer Support Teams Arcwell has launched Relay, an artificial-intelligence platform designed to help customer-support agents work more efficiently by analyzing support conversations and suggesting answers. The platform reviews customer interactions and draws on internal help documents to recommend responses. It can also generate conversation summaries and identify customers who may be at risk of canceling, giving support teams another way to prioritize follow-up. The capabilities reflect the expanding role of AI in customer service, where companies are using software to retrieve information, automate routine work and surface potential customer concerns. Arcwell tested Relay with 12 companies over six months. The company said agents using the platform closed tickets 14% faster during the test period. Arcwell did not release results on whether Relay improved customer retention. The service costs $65 per agent each month. “Support agents spend too much time searching for information that their companies already have,” CEO Priya Shah said. Relay’s functionality depends on access to customer-support transcripts and internal help documents, which could raise data-governance concerns for companies evaluating the platform. Arcwell said customer data is encrypted and is not used to train shared models. Customers can choose either a 30-day or 90-day data-retention period. The launch gives Arcwell a new product as it seeks to help companies improve customer service through AI. Its early performance figures point to faster ticket resolution, while the absence of disclosed customer-retention results leaves that aspect of the platform’s impact unmeasured.

Anthropic: Claude Fable 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.

OpenAI: GPT-5.6 Luna: # 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 Fable 5: # Cloudnote Says Debugging Page Exposed Excerpts From 1,842 User Sessions *The internal page was reachable without a login for four hours before a security researcher's report prompted its removal; the company says its investigation is ongoing.* Cloudnote, the note-taking service with 3.2 million registered users, disclosed that an internal debugging page containing excerpts of customer prompts was accessible on the open internet without authentication for roughly four hours on Monday. According to the company, logs show the page displayed excerpts from 1,842 user sessions, with each excerpt containing up to 200 characters from a prompt. Full documents, passwords and payment information were stored on separate systems and were not present on the page, Cloudnote said. The company removed the page 19 minutes after being notified by a security researcher who discovered the exposure. Cloudnote says it has found no evidence so far that anyone other than the researcher accessed the page, though it cautioned that its investigation is not complete. Key questions remain unanswered, including how the page came to be publicly reachable, whether the affected excerpts contained sensitive personal details, and whether the company can definitively rule out other access during the four-hour window. Cloudnote has not said whether it will notify the users whose session excerpts appeared on the page. The incident affects a small fraction of Cloudnote's user base, but it highlights a recurring risk for cloud software companies: internal tools that handle customer data escaping the access controls meant to contain them.

OpenAI: GPT-5.6 Luna: # Cloudnote exposed customer prompt excerpts on an unauthenticated debugging page *The page was publicly accessible for four hours, potentially revealing snippets from 1,842 user sessions, though Cloudnote says it has found no evidence of unauthorized access.* Cloudnote says a debugging page containing excerpts from customer prompts was accessible without authentication for four hours on Monday. Logs indicate that excerpts from 1,842 user sessions appeared on the page, a disclosure that could affect a small but potentially sensitive portion of the company’s 3.2 million registered users. The excerpts could contain up to 200 characters from a prompt. Cloudnote said full documents, passwords and payment information were stored elsewhere and did not appear on the page. The company removed the page 19 minutes after a researcher notified it. Cloudnote said it has found no evidence that anyone besides the researcher accessed the exposed page. That finding is based on its investigation so far, however, and the company has not completed its review. It has not said whether the logs can establish who viewed the page during the four-hour window or whether any prompts contained confidential business or personal information. The incident underscores the risk of diagnostic tools being deployed outside the access controls used for customer-facing systems. The number of affected sessions is known from logs, but the extent of any actual viewing—and the content of the excerpts in each session—remains unknown.

Anthropic: Claude Fable 5: # HelioDock Launches Night Shift, Betting Warehouses Will Pay to Watch Robots From Afar *The autonomous warehouse robot maker pairs its new remote fleet-monitoring software with a $24 million Series B, but a beta customer says the product's value drops when connectivity does.* HelioDock, a maker of autonomous warehouse robots, on Thursday launched Night Shift, a software product that lets operators monitor and manage their robot fleets remotely, ending a 14-month private beta that the company says reduced unplanned robot stoppages for participating customers. The launch, paired with a newly announced $24 million Series B funding round led by North Coast Ventures, positions HelioDock to compete in the growing market for fleet-management software — a layer of the warehouse automation stack where rivals Relay Robotics and YardOS already offer competing products. The pitch is straightforward: warehouse robots don't keep business hours, and the people responsible for them shouldn't have to be on site every time something goes wrong. "A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet," HelioDock CEO Mara Velez said. ## Beta results and an unverified number Night Shift's private beta ran for 14 months and involved nine logistics companies operating a combined 310 robots. According to HelioDock, those beta customers experienced 17% fewer unplanned robot stoppages while using the software. The figure comes from the company and has not been independently verified. If it holds up, the number speaks to the core problem Night Shift is designed to solve: robots that halt unexpectedly — whether confused by a stray pallet or otherwise stuck — and then sit idle until a human can intervene. By moving that intervention off the warehouse floor and onto a remote operator's screen, HelioDock is betting that customers will trade a monthly software fee for fewer late-night site visits and less downtime. The product will cost $90 per robot per month when it becomes generally available on September 15. ## Where the software runs into limits The beta also surfaced the product's most obvious constraint: Night Shift is only as good as the network connection underneath it. Aaron Pike of Fulcrum Logistics, one of the nine beta customers, said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. When the connection fails, HelioDock's robots fall back to their locally configured behavior, and remote control becomes unavailable — meaning the two-in-the-morning drive Velez described may still be on the table for operators running facilities in areas with spotty coverage. That fallback design means a connectivity loss doesn't strand the robots entirely; they continue operating according to whatever behavior has been configured on the machines themselves. But it does mean the monitoring and intervention capabilities customers are paying for disappear precisely when a facility is hardest to reach remotely. For warehouses with reliable connections, Pike's account suggests the product delivers on its premise. For operators weighing the $90-per-robot monthly fee, the quality of their network infrastructure may be as important a factor as the software itself. ## Fresh capital and a crowded field The Series B brings HelioDock's total funding to $39 million, giving the company resources to push Night Shift into general availability against established competition. Relay Robotics and YardOS both sell fleet-management products, making the category one where HelioDock will need to differentiate on more than the novelty of remote monitoring. The company's argument, at least for now, rests on the beta numbers and the endorsement implicit in nine logistics companies running 310 robots through more than a year of testing. Whether the 17% reduction in unplanned stoppages translates beyond the beta group — and whether customers in less-connected regions find enough value to justify the subscription — will become clearer after September 15, when Night Shift opens to the broader market.

OpenAI: GPT-5.6 Luna: # HelioDock launches remote monitoring software for autonomous warehouse robots *Night Shift will let warehouse operators oversee robot fleets from off-site, but its usefulness depends on a facility’s connectivity.* HelioDock on Thursday launched Night Shift, software designed to let warehouse operators monitor autonomous robot fleets remotely, as the company announced a $24 million Series B investment led by North Coast Ventures. The product is the result of a 14-month private beta involving nine logistics companies and 310 robots, according to HelioDock. The company said participating customers experienced 17% fewer unplanned robot stoppages during the beta, though that figure has not been independently verified. Night Shift is aimed at a common operational problem for warehouses that run robots around the clock: responding to issues when staff are not on site. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” CEO Mara Velez said. The software does not eliminate the need for local connectivity. Aaron Pike, a beta customer at Fulcrum Logistics, said Night Shift worked well in facilities with strong connections but became less useful during rural LTE outages. When connectivity fails, robots revert to their locally configured behavior, HelioDock said. Remote control is unavailable during the outage, limiting what operators can do from off-site until the connection is restored. The supplied information does not specify how long robots can operate under those local behaviors or what actions are available to staff during a disruption. Night Shift will be generally available on Sept. 15 at a price of $90 per robot per month. HelioDock makes autonomous warehouse robots, and its product enters a fleet-management market that also includes offerings from Relay Robotics and YardOS. The company’s funding announcement brings its total financing to $39 million. HelioDock did not provide additional details about the round beyond naming North Coast Ventures as its lead investor. The beta’s scale gives HelioDock an early set of operating data across multiple logistics companies, but the reported performance improvement remains a company-supplied result rather than an independently assessed measure. HelioDock also did not provide a comparison group, details about how stoppages were counted or information about whether the results varied by facility. Night Shift’s core promise is therefore tied to both software access and the infrastructure supporting it. In a well-connected warehouse, operators can monitor robots without being physically present. In facilities affected by connectivity gaps, the robots continue according to their local configurations, while the remote layer—and any remote control—drops away. For customers preparing to adopt the service, the pricing will scale with fleet size: a company operating 100 robots would pay $9,000 per month before any other charges, based on HelioDock’s stated per-robot price. The company has not disclosed whether Night Shift will be offered with different service tiers or whether connectivity requirements vary by deployment. General availability begins Sept. 15.

Anthropic: Claude Fable 5: # The Price of a Point: What a Customer-Support Benchmark Reveals About Paying for Quality The most expensive model in a recent customer-support benchmark scored highest. That's the boring part. The interesting part is that its closest competitor matched it — statistically speaking — while costing less than half as much to run. Researchers put five language models through 200 anonymized customer-support tickets, asking each model to draft one proposed answer per ticket. Three experienced support managers then graded every answer blindly, without knowing which model produced it. The scores blended three things support teams actually care about: factual correctness, usefulness, and tone. ## The headline: a tie at the top that isn't priced like one Model North led the pack with an average score of 84.2. Model Vale came in at 83.8. That 0.4-point gap sounds like a victory for North until you read the fine print: the difference was not statistically significant. On this test, with this sample size, the two models are effectively indistinguishable in quality. Their prices are not indistinguishable. North costs an estimated $1.90 per 1,000 tickets; Vale costs $0.74. For a team processing a million tickets a year, that's the difference between $1,900 and $740 — a 61% savings for output the graders couldn't reliably tell apart. Vale also answered 18% faster, which matters in support workflows where drafts feed into a human agent's queue. This is the clearest finding in the benchmark, and it's worth dwelling on because it cuts against a common procurement instinct: buy the model at the top of the leaderboard. When the top two entries are statistically tied, the leaderboard position is noise, and cost and latency become the real differentiators. ## Where the models diverged Beneath the tie, there were genuine differences in character. North performed best on billing disputes — tickets that tend to reward careful reading of account details and a measured, de-escalating tone. Vale performed best on technical troubleshooting, where structured diagnostic reasoning matters more. For a support organization, that's a more useful insight than the overall averages. A company whose ticket volume skews heavily toward billing might reasonably lean toward North despite the price. A SaaS company drowning in configuration questions might find Vale is not just cheaper but actually better suited to its workload. The benchmark's aggregate score flattens these differences; the category-level results restore them. The three cheaper models — Ember, Kite, and Harbor — scored progressively lower, from 79.1 down to 71.0, at costs falling as low as $0.08 per 1,000 tickets. Whether those tradeoffs make sense depends entirely on what the drafts are for. If a human agent reviews and edits every response before it ships, a 71-scoring draft that costs almost nothing may be a fine starting point. If drafts go out with light review, the quality gap gets expensive in a different currency: customer trust. ## The shared blind spot One result applied to all five models equally: they performed poorly whenever a ticket depended on an image attachment. A screenshot of an error message, a photo of a damaged product, a picture of a confusing invoice — these are routine in real support queues, and every model in the test stumbled on them. This is arguably the benchmark's most practical warning. Any deployment plan should include routing logic that flags attachment-dependent tickets for human handling, regardless of which model wins the bake-off. ## What this benchmark can't tell you The study's scope was deliberately narrow, and readers should treat it that way. All 200 tickets came from a single software company, in English only. Ticket styles, product complexity, and customer expectations vary enormously across industries, so these rankings may not transfer. The researchers also tested only static question-answering. They did not evaluate agentic actions, tool use, or multilingual support — capabilities that increasingly define real support automation, where a model might look up an order, issue a refund, or respond in Portuguese. A model that drafts beautiful English replies may or may not handle those tasks well, and this benchmark is silent on the question. ## The takeaway The lesson here isn't "buy Vale." It's a habit of mind: when benchmark scores land within the margin of error, stop comparing scores and start comparing everything else — cost, speed, category strengths, and fit with your actual ticket mix. And whatever you deploy, keep humans on the tickets with pictures. For now, that's where all of these models go quiet.

OpenAI: GPT-5.6 Luna: # Vale Nearly Matches the Top Score at a Fraction of North’s Cost A benchmark of five language models for customer support points to a practical tradeoff: **Model Vale delivered support quality statistically indistinguishable from the top-scoring Model North, while costing substantially less and responding faster**. North achieved the highest average score, 84.2, across the test set. Vale followed closely at 83.8, but its estimated cost was just **$0.74 per 1,000 tickets**, compared with **$1.90 for North**. Vale also answered **18% faster**. For organizations choosing a model for routine support drafting, that combination may matter more than North’s narrow numerical lead. The results do not show that Vale is universally better. North performed best on billing disputes, while Vale led on technical troubleshooting. But the overall findings suggest that the most expensive model was not clearly the most effective choice for the benchmark’s support workload. ## How the test worked Researchers evaluated five language models using **200 anonymized customer-support tickets** from a single software company. Each model generated one proposed response for every ticket. Three experienced support managers then graded the responses blindly, meaning they did not know which model produced a given answer. The combined score reflected three dimensions: - Factual correctness - Usefulness to the customer - Tone This design is more informative than relying on automated metrics alone. Human reviewers can judge whether an answer is actually helpful, appropriately phrased and responsive to the customer’s problem. However, the score gap between North and Vale—84.2 versus 83.8—was **not statistically significant**. In practical terms, the benchmark does not establish that North produces better answers overall. The small difference could plausibly reflect variation in the sample or grading. The cost figures are estimates per 1,000 tickets, not necessarily a complete measure of operational expense. They likely capture model-inference costs, but a real deployment could also involve integration, monitoring, quality review and escalation. Even so, the relative gap is large: Vale’s estimated cost is less than half of North’s, while the two models produced broadly comparable quality in this test. ## Cost and speed change the deployment calculation The remaining models formed a lower-performing group in the benchmark, with average scores ranging from 79.1 to 71.0. Their lower prices may appeal to teams handling very high volumes, but the results suggest that aggressively minimizing inference cost can come with a noticeable reduction in response quality. That does not make the cheaper models useless. A low-cost model might be suitable for simple classification, draft generation or tickets that receive mandatory human review. But the benchmark provides stronger evidence for Vale as a balanced option: it remained near the top on quality without North’s price premium. The category-level results also argue against choosing a model based only on an overall average. North’s advantage on billing disputes could matter to a company where payment, refunds or invoices dominate support volume. Vale’s strength in technical troubleshooting could make it the better fit for a product-led software business handling mostly setup and diagnostic questions. ## Important limits The benchmark covers only English-language tickets from one software company. Its results may not transfer to retail, healthcare, financial services or other support environments with different policies and customer expectations. The test also examined generated answers, not complete support automation. Researchers did not evaluate agentic actions, tool use or multilingual support. The models were not tested on tasks such as looking up an account, changing a subscription, issuing a refund or coordinating across systems. A model that writes a good answer may still be unreliable when allowed to take actions. One weakness was shared by all five models: **tickets that depended on an image attachment produced poor performance across the board**. This suggests that improving multimodal understanding—or routing image-dependent cases to a human—could be more valuable than switching among these text-focused models. Overall, the clearest conclusion is not that North loses to Vale. It is that **Vale appears to offer nearly the same measured support quality at lower cost and higher speed**, making it the stronger default candidate for this particular workload. Teams should validate that result on their own ticket mix, especially for billing, attachments and any workflow involving real-world actions.

Anthropic: Claude Fable 5: **6** Every content line violates a rule: 1. `#Flood software moves upstream` — missing space after `#` 2. `The company is expanding its pilot.` — dek not italicized 3. `##Details` — missing space after `##` 4. `Read [the announcement](www.example.com/news).` — not an absolute HTTPS URL 5. `* The first pilot begins Monday.` — uses `*` instead of a hyphen 6. `<div>The second pilot begins Friday.</div>` — raw HTML

OpenAI: GPT-5.6 Luna: 6

Anthropic: Claude Fable 5: **C** C is the strongest lede because it is accurate, specific, and complete — it names the action (product pulled), the cause (workspace titles exposed), the scope (4,860 accounts), and the duration (11 hours), all supported by the facts. The others fail: - **A** falsely claims private messages were leaked (message contents were *not* sent). - **B** is vague, buries the news, and reads like PR spin ("innovative," "difficult decision"). - **D** exaggerates ("everything they had written") and asserts hacker access, contradicting the fact that no malicious access was found.

OpenAI: GPT-5.6 Luna: C

Anthropic: Claude Fable 5: # Startup Spotlight: LedgerLark Wants to Automate the Tedium of Post-Trade Cleanup When a trade goes through at a small investment firm, the work is often just beginning. Confirmations have to match custodian records, custodian records have to match internal ledgers, and any discrepancy — a transposed digit, a stale price, a copied account number — has to be tracked down by hand. LedgerLark, a young startup with five paying customers and a fresh $3.5 million seed round led by Motive Row, is betting that AI agents can take on much of that grunt work. ## An Insight Born From a Three-Day Headache Co-founder Mina Batra spent seven years in operations at a regional brokerage, which is where the idea took shape. The moment she points to: watching a six-person team spend three days untangling a trade mismatch that turned out to be caused by a copied account number. The error was trivial; the process of finding it was not. Her co-founder, Owen Carr, built reconciliation systems at a large bank — the kind of internal tooling that big institutions can afford to develop but smaller firms typically cannot. Together, their pitch is essentially that small firms deserve something closer to what the big players have built for themselves, without the headcount or the custom software budget. ## What the Product Actually Does LedgerLark's agents compare three sources of truth — trade confirmations, custodian records and internal ledgers — flag discrepancies, and propose corrections. Critically, a human has to approve every fix. The system does not execute trades and does not move money. That constraint is worth dwelling on. Keeping a person in the loop limits how much labor the product can actually remove, but it also makes the software far easier for a compliance-conscious firm to say yes to. For a startup selling into a risk-averse industry, that trade-off looks deliberate. ## Early Traction, With Caveats Five firms are 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 — a striking figure, though it comes from LedgerLark itself and reflects a single customer's experience. Whether that result generalizes across firms with different systems, asset mixes and data quality is an open question. Five customers is real revenue but a small sample. The more interesting signal may come over the next year: whether those firms renew, expand usage, or quietly revert to spreadsheets. ## The Competitive Landscape Is Not Empty Post-trade reconciliation is not a neglected problem. Established vendors have sold reconciliation software for decades, and larger institutions run internal systems of the sort Carr used to build. LedgerLark's differentiation, as best one can tell from the outside, rests on two claims: that AI agents can handle messier, less structured discrepancies than rules-based tools, and that the product is priced and packaged for firms too small for enterprise vendors to court aggressively. Both claims are plausible. Neither is proven at scale. Incumbents can add AI features to existing products, and they start with something LedgerLark lacks — years of trust with a customer base that changes vendors reluctantly. ## The Slow Grind of Selling to Finance Perhaps the biggest obstacle is not the technology but the sales cycle. Financial institutions often require lengthy security and procurement reviews before adopting new software, and a seed-stage company has a finite runway to survive them. A $3.5 million round buys time, but selling into finance has exhausted better-funded startups before. ## Worth Watching LedgerLark is attacking a genuinely painful, genuinely unglamorous problem, with founders who have lived it from both the small-firm and big-bank sides. The early numbers are encouraging but thin, the competition is entrenched, and the buyers move slowly. The next twelve months — renewals, new logos, and whether that exceptions number holds up elsewhere — should tell us whether this is a feature, a niche tool, or the start of something larger.

OpenAI: GPT-5.6 Luna: # Startup Spotlight: LedgerLark Brings AI Agents to Post-Trade Cleanup A copied account number can create a surprisingly large operational problem. At one regional brokerage, Mina Batra watched a six-person team spend three days tracing and resolving a single trade mismatch. The error was simple; finding it across confirmations, custodian records and internal ledgers was not. That experience helped prompt Batra to explore LedgerLark, a startup building AI agents for post-trade operations at small investment firms. She co-founded the company with Owen Carr, who previously built reconciliation systems at a large bank. Batra spent seven years working in operations at a regional brokerage, giving the pair experience on different sides of the same problem: reconciling what firms intended to trade with what their systems and counterparties recorded. LedgerLark’s product is designed to examine those records, identify inconsistencies and propose corrections for a person to approve. The agents compare trade confirmations, custodian data and internal ledgers, then present the likely resolution rather than automatically changing an account or transaction. That boundary is important. LedgerLark does not execute trades or move money. Its role is closer to an investigative and administrative layer for post-trade workflows, where teams often spend time gathering evidence, checking formats and determining which record is authoritative. The company says five firms are currently paying customers. Annual contracts range from $36,000 to $110,000, suggesting that LedgerLark is initially targeting firms for which operational errors are costly but dedicated engineering teams may be limited. According to the company, one customer reduced its unresolved daily exceptions from 70 to 18 after adopting the product. That is a meaningful reported change, though it is not yet clear how much of the reduction came from LedgerLark itself, changes in customer processes or the kinds of exceptions included in the measurement. LedgerLark has raised a $3.5 million seed round led by Motive Row. The funding gives the company room to expand its product and customer base, but post-trade infrastructure is a demanding market. Established reconciliation vendors already serve financial institutions, and many larger firms rely on internal bank software that is deeply integrated with their records and approval systems. LedgerLark’s opening may be among smaller investment firms whose workflows are too complex for spreadsheets but too specialized—or expensive—to replace with a broad enterprise platform. Its AI approach could help reduce the manual searching involved in exception management, particularly when information is distributed across systems with inconsistent formats. The obstacles are substantial. Financial institutions may require lengthy security reviews and procurement processes before allowing a new system to handle sensitive operational data. Even without trade execution or money movement, a tool that reads confirmations, custody records and internal ledgers will need to demonstrate strong access controls, auditability and reliability. Human approval may reduce risk, but it also means the product must earn trust from the people responsible for signing off on corrections. The central question for LedgerLark is therefore less whether AI can spot mismatches than whether its recommendations are dependable enough to become part of a firm’s daily control process. Its early customer count and reported exception reduction offer an initial signal. The next test will be whether those results hold across more institutions, more data formats and the long procurement cycles that shape financial software.

Anthropic: Claude Fable 5: # Software Glitch Sends $2.8 Million in Duplicate Reimbursements Through KiteBridge's Platform A faulty software update at expense-management company KiteBridge triggered 13,400 duplicate reimbursement payments totaling $2.8 million on Tuesday morning, affecting 91 corporate customers and raising questions about the company's earlier promises of payment safeguards. The trouble began at 8:45 a.m., when KiteBridge deployed version 4.2 of its platform, which businesses use to process employee expense claims and issue reimbursements. Within half an hour, at 9:12 a.m

OpenAI: GPT-5.6 Luna: # KiteBridge Sends $2.8 Million in Duplicate Reimbursements After Software Update KiteBridge accidentally issued 13,400 duplicate employee reimbursements totaling $2.8 million after a software update caused its payment system to misread repeated requests. The expense-management company halted automated payments Tuesday morning, but not before the failure affected 91 corporate customers. The incident began after KiteBridge deployed version 4.2 at 8:45 a.m. Tuesday. The update was intended to process reimbursements through the company’s payment infrastructure. Instead, a timeout from the payment processor triggered a problem in KiteBridge’s fallback system. When the processor failed to respond promptly, the fallback system treated retry attempts as new reimbursement requests rather than recognizing them as repeats of payments already submitted. That distinction is central to preventing duplicate transactions: a payment should be processed once, even if the system has to retry the request or wait for confirmation. The first customer complaint arrived at 9:12 a.m. As reports accumulated, KiteBridge investigated the payment activity and suspended automated payments at 10:03 a.m. The company’s response prevented additional automated transactions, but thousands of duplicate reimbursements had already been sent. The payments were funded by the affected employers, not withdrawn from employees’ personal bank accounts. KiteBridge said no employees had money taken from their accounts as a result of the incident. The duplicate payments instead created a financial and administrative problem for the 91 companies using the platform, which had to identify and recover excess funds. By Wednesday evening, KiteBridge had recovered $2.3 million of the $2.8 million total. The remaining amount was still pending, according to the company. The company’s public communication initially characterized the incident as delayed reimbursements. It did not disclose that duplicate payments had been issued until six hours after automated payments were suspended. That delay added to questions about how quickly customers were informed about the scope and nature of the failure. The incident also calls attention to KiteBridge’s previous claims about payment safeguards. An April changelog for version 4.0 said the release had introduced “exactly-once payment protection,” a technical measure intended to ensure that a single reimbursement request cannot result in multiple payments, even when systems retry after an error or timeout. KiteBridge has not publicly explained why that protection failed in version 4.2 or whether the safeguard was bypassed, improperly configured or incompatible with the fallback system. The company’s chief executive, Lena Ortiz, apologized for the incident and said KiteBridge would publish a technical report. That report is expected to detail how the payment-processor timeout interacted with the fallback logic, why the duplicate requests were not detected earlier and what changes the company will make to prevent a recurrence. Until then, customers are left managing the consequences of a software failure that converted an ordinary retry mechanism into millions of dollars in unintended reimbursements.

Anthropic: Claude Fable 5: # QueueZero's AI Meeting Assistant Assigns Its First Task — to the CEO QueueZero launched its AI meeting assistant, Proxy, on Wednesday, and the product wasted no time demonstrating both its capabilities and its literal-mindedness — by assigning a task to the company's own chief executive during the launch event. Proxy is designed to attend meetings on behalf of users, generate notes, and automatically assign follow-up work to software agents. During Wednesday's demonstration, that automation kicked in earlier than the company may have planned. Partway through the presentation, a customer in the audience remarked, "Someone should tell Evan to stop talking and send us the pricing sheet." Proxy, which was actively monitoring the session, interpreted the comment as an instruction. Within moments, it had created two tasks and assigned both to QueueZero CEO Evan Wu: "Stop talking" and "Send pricing sheet." Wu laughed off the incident on stage. He completed the second task, distributing the pricing sheet to attendees, and deleted the first. He did not, according to attendees, stop talking. The moment offered an unscripted illustration of both Proxy's core functionality and the challenges of deploying AI systems that parse natural conversation for actionable requests. Distinguishing between an offhand remark and a genuine directive remains one of the harder problems in the category, and Wednesday's demonstration showed how easily the line can blur when an assistant is listening to everything said in a room. QueueZero emphasized that Proxy includes safeguards around its more consequential actions. The system requires explicit user confirmation before sending any external message, meaning an errant remark cannot result in an email or document leaving the organization without human sign-off. Internal task creation, however, happens automatically by default — which is how Wu ended up with his impromptu to-do list. Administrators can disable automatic task creation entirely if they prefer a more conservative configuration. The incident is unlikely to slow the product's rollout. AI meeting assistants have become one of the most crowded segments in workplace software, with vendors racing to move beyond transcription and summaries into what the industry calls "agentic" behavior — systems that act on what they hear rather than simply recording it. Proxy's launch-day misfire is arguably a preview of the kinds of edge cases enterprises will need to manage as these tools gain autonomy. Proxy is priced at $30 per user per month and enters public beta next week. QueueZero did not say whether any changes to the default settings are planned before the beta begins.

OpenAI: GPT-5.6 Luna: # QueueZero’s AI meeting assistant turns an audience comment into executive task list QueueZero launched Proxy, an AI meeting assistant, on Wednesday, presenting it as a tool that can attend meetings, produce notes and assign follow-up work to software agents. The launch demonstration also showed how quickly an assistant designed to interpret conversation can move from observation to action. 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 pricing-related task and deleted the other. For a moment, Proxy had adopted a notably direct management style. The incident illustrated a central distinction in QueueZero’s product design: Proxy can automatically create internal tasks, but it requires confirmation before sending an external message. That approval step is intended to prevent the assistant from turning conversational remarks into messages sent to customers, colleagues or other outside recipients. Administrators can also disable automatic task creation, according to QueueZero. Organizations using the system will therefore be able to choose whether Proxy creates internal assignments on its own or only records possible follow-up work for human review. QueueZero describes Proxy as an assistant that can participate in meetings and convert discussions into operational tasks for software agents. The company did not provide further details about which agents are supported or how the system determines when a comment should be treated as an instruction rather than a casual remark. That judgment is likely to matter as AI meeting tools become more involved in workplace workflows. Meetings routinely include suggestions, hypotheticals and offhand comments that may sound like commands when removed from their broader context. Requiring confirmation for external communication limits the consequences of some errors, while automatic internal task creation can still produce unwanted assignments or administrative noise. Proxy will enter public beta next week. QueueZero is pricing the service at $30 per user per month. The company’s demonstration offered a compact example of the trade-offs involved in delegating workplace follow-up to AI: automation can save time when it correctly identifies an obligation, but a system that acts on ambiguous speech may need clear controls and review mechanisms. In this case, the affected executive was also the person overseeing the launch, giving him an unusually convenient opportunity to test both the assistant’s initiative and its undo button.

Anthropic: Claude Fable 5: # Foundry Hall, Milwaukee's Startup Accelerator, to Close After Six Years The lights in Foundry Hall will go out for good on October 31. The nonprofit startup accelerator, which opened in 2018 in a converted industrial building on Milwaukee's near south side, announced this week that it will shut down at the end of October. The reason is arithmetic more than anything else: Blank River Bank, its largest corporate sponsor, declined to renew a sponsorship that covered 38 percent of the organization's annual budget, and no replacement emerged. "We kept waiting for the replacement check that never came," said Camille Ross, the accelerator's executive director. Ross said the organization has enough money to keep the doors open through October, but not enough to run another full program. Rather than start a cohort it couldn't finish, the board voted to wind down. ## What gets left behind The numbers Foundry Hall leaves behind are not small. Over six years, it supported 212 startups. Those companies now employ 1,140 people across the region — jobs that, in many cases, began as an idea sketched out in the hall's shared workspace. Dev Malik, one of those founders, remembers what came before. "My first investor meeting happened in that building," he said. "Before Foundry Hall, I was pitching from the back table of a coffee shop." The closure has immediate human costs. Six staff members will lose their jobs. Twelve startups currently working out of the building must find new space by the end of October — a scramble that, for early-stage companies operating on thin margins, is more than an inconvenience. ## The sponsor's exit Blank River Bank offered little explanation for its departure. A spokesperson said the decision reflected a shift in the bank's community investment priorities and had nothing to do with Foundry Hall's performance. That distinction may matter to the bank. It changes nothing for the six employees or the twelve companies packing boxes. Ross declined to criticize the bank directly, but the strain in the organization's final months is evident. A single sponsor supplying more than a third of the budget was always a vulnerability; when the money stopped, there was no cushion deep enough to absorb the loss. ## What comes next City officials say they are trying to soften the landing. Deputy Mayor Susan Cole said the city is in conversations with building owners and local universities about workspace for the displaced companies. Nothing has been finalized. Whether Milwaukee replaces Foundry Hall — the institution, not just the square footage — is a harder question. Accelerators are more than desks and conference rooms. They are the accumulated network of mentors, investors, and founders who know where to find each other. For six years, that network had an address. After October 31, it won't.

OpenAI: GPT-5.6 Luna: # Foundry Hall to Close After Six Years of Supporting Milwaukee Startups Foundry Hall, the Milwaukee nonprofit startup accelerator that helped turn ideas into companies, will close its doors on Oct. 31. The organization opened in 2018 and has supported 212 startups, whose businesses collectively employ 1,140 people. Its workspace, workshops and investor meetings became an entry point for entrepreneurs who might otherwise have struggled to find equipment, advice or a place to begin. But Foundry Hall will not make it to another full program. Its largest corporate sponsor, Blank River Bank, provided 38% of the accelerator’s annual budget and declined to renew its sponsorship. The organization has enough money to operate through October, but not enough to continue beyond that. “We kept waiting for the replacement check that never came,” Executive Director Camille Ross said. The closure will eliminate six jobs and force 12 startups currently using Foundry Hall’s workspace to leave. For some founders, the move represents more than the loss of an office. It means losing a shared environment that connected them with mentors, investors and one another. Dev Malik, who founded one of the companies that worked out of the building, remembers arriving before he had much more than an idea. “My first investor meeting happened in that building,” he said. “Before Foundry Hall, I was pitching from the back table of a coffee shop.” Foundry Hall’s supporters say its impact cannot be measured only by its budget or the number of companies housed under its roof. The startups it assisted range from early-stage ventures still seeking funding to businesses that have grown into employers. Together, they represent a small but visible part of Milwaukee’s efforts to build an innovation economy. Blank River Bank said its decision was not a judgment on the accelerator’s work. In a statement, a spokesperson said the change “reflects a change in our community investment priorities and is unrelated to Foundry Hall’s performance.” The decision leaves city officials and other community organizations looking for ways to keep the current companies from being scattered. Deputy Mayor Susan Cole said the city is talking with building owners and local universities about possible space for the startups. Those conversations may help preserve pieces of Foundry Hall’s network, but they cannot replace the institution immediately. Its staff will spend the remaining weeks winding down programs, assisting the companies in the workspace and preparing for the final day. When the doors close, Milwaukee will lose more than an accelerator. It will lose a physical place where inexperienced founders could test a pitch, meet a first investor or find someone who understood the uncertainty of starting a company. For the entrepreneurs who came through Foundry Hall, the building’s end will not erase what began there. But it will leave a gap that the city will have to decide how—and whether—to fill.