Today's 12 Stories — Wednesday, November 11, 2026

AI Models Story 1 of 12

Anthropic Ships Claude Fable 5.2 and Mythos 5.2, and Splits Access by Risk

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, upgrades to its Claude Fable and Claude Mythos models. Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens. Anthropic also offers an optional fast mode that generates tokens up to 2.5 times faster, at $16 per million input tokens and $80 per million output tokens. For a buyer, that makes speed a line item you can switch on for specific workloads rather than a property of the model you are stuck with.

The performance claims lean on coding and security. Anthropic said Fable 5.2 scores 49.1 percent on the FrontierCode 2.0 benchmark at Xhigh effort, compared with 32.4 percent for Fable 5.1 and 45.5 percent for Claude Opus 5.5. On security work, Anthropic said Mythos 5.2 scores 89.6 percent on its CVE Rediscovery evaluation, up from 78.3 percent for Mythos 5.1. Those are vendor reported numbers from a vendor chosen set of tests, and they deserve the same skepticism as any launch day scorecard. They are still a useful signal about where Anthropic is putting its engineering effort.

The more interesting decision is how Anthropic is distributing the capability. The company said Mythos 5.2 access goes through its Life Sciences and Cyber Verification Programs, while Fable 5.2 carries cyber and biology safeguards and is available to all users. In plain terms, the most sensitive capability is reserved for organizations that have been vetted, and everyone else gets a version with guardrails. That is a template other frontier labs are likely to copy, and it has a practical consequence for enterprises: if your security or life sciences team needs the unrestricted version, plan for an application and verification process rather than a purchase order.

Anthropic also described what the model can do over long stretches of autonomous work. The company said an experiment with more than 1,000 Claude Fable 5.2 subagents over eight days produced a prototype operating system that boots on a laptop. It said a team of 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in 30 hours, compared with 11 days for a first effort with an unreleased Claude model. These results describe large scale agent coordination, not a typical business workload, but they show the direction of travel: the unit of work is shifting from a single prompt to a managed team of agents running for days.

For executives, the takeaways are procedural. Run your own evaluation on your own tasks before moving production traffic. Test the effort settings, because Anthropic's own results show the same model landing at very different scores depending on that dial. And decide now who inside your organization owns the access question for restricted models, because the verification programs will not move at procurement speed.

AI Models Story 2 of 12

Google Makes Gemini 4 Carbon Broadly Available at a Fraction of Fable's Price

Google said on November 10 that Gemini 4 Carbon is now available to paid API users and to Google AI Ultra subscribers, after an earlier limited release to cybersecurity defenders. Google said the introductory price is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 from November 30. The timing matters, because Anthropic priced Claude Fable 5.2 at $8 and $40 a day earlier. By simple arithmetic, Fable's list price is four times Carbon's introductory rate on both input and output, and still double it once Carbon's higher price takes effect.

Google's headline claim is breadth. The company said Gemini 4 Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench. Those are agent and coding style benchmarks run by outside parties, which makes them more relevant to enterprise buyers than pure knowledge tests. As always, a first place ranking on a leaderboard can shift within weeks as rivals update, so the durable question is whether the model holds up on your own workflows.

The second theme is safety and reliability, and Google put numbers on it. The company said Gemini 4 Carbon passed 92 percent of a 375 task agentic honesty benchmark, compared with 70 percent for Gemini 4 Argon. Google also said Carbon blocked 99 percent of attack trajectories from an adaptive red teaming agent, with a per attempt breach rate of 0.8 percent versus 3.1 percent for Argon. Those are Google's own measurements, so treat them as claims to be tested rather than guarantees. Still, the direction is notable. Vendors are now competing not only on how capable a model is but on how reliably it behaves when it is given tools and autonomy, which is the question that decides whether agents get approved for real work.

For a chief information officer, the practical picture is a two vendor frontier market with converging capability and diverging price. The introductory pricing window is a reason to pilot now, since you can learn how the model behaves before the higher rate applies. It is also a reminder to build your applications so the model underneath can be swapped. Pricing that doubles on a fixed date, and rankings that change weekly, punish teams that hard wire a single provider.

The competitive dynamic is worth watching. Anthropic and Google released flagship updates within a day of each other. Expect pricing and benchmark claims to keep moving in both directions through the end of the year.

Enterprise AI Story 3 of 12

IBM and Microsoft Pair 6,000 Consulting Agents With Microsoft's Stack

IBM and Microsoft announced a new agentic platform called Imagine Enterprise on November 10. The platform pairs more than 6,000 IBM Consulting AI agents and assets with Microsoft technologies including Microsoft Foundry, Copilot Studio and Fabric. The pitch is a pre built fleet of industry specific agents that can be deployed on top of the Microsoft environment many large companies already run.

The industry coverage is staged. IBM and Microsoft said Imagine Enterprise launches with frameworks for financial services, insurance and retail, with manufacturing, energy and public sector planned for later. That sequencing tells you where the two companies see the fastest return and the most regulated demand. Financial services and insurance combine large volumes of document and process work with strong compliance requirements, which is exactly where a governed agent platform has the most to offer and the most to prove.

IBM named SEI as a client exploring the platform. A named customer that is still exploring is a normal stage for a launch announcement, and executives should read it that way: this is an early commitment, not proof of production results. The announcement is also a services led offering. The agents come from IBM Consulting, which means the product is tied to a consulting relationship rather than sold as a standalone tool. That has advantages, because buyers get implementation help and accountability. It also means the economics resemble a services engagement, with the customization, change management and lock in that implies.

For buyers, the strategic question is whether a packaged agent catalog beats building on a general purpose platform. Packaged agents promise speed, since a company does not start from a blank page. The tradeoff is fit. Industry frameworks encode someone else's idea of your process, and the work of adapting them can erase the head start. A sensible approach is to treat the catalog as a source of tested patterns, pick one workflow where the framework closely matches your own, and measure cycle time and error rates against your current baseline before expanding.

The deal also reflects a wider pattern in enterprise AI. Platform vendors and consultancies are combining into joint offerings because customers want one accountable party for strategy, build and governance. Ask any vendor in this category for pricing, an availability timeline and a clear account of who is responsible when an agent makes a mistake. Those answers matter more than the size of the catalog.

AI Safety Story 4 of 12

Fireworks AI Breach Pushes Customers Away and Exposes Notion AI Chats

A security incident at AI inference provider Fireworks AI is rippling through its customers. Notion said that some unattributable Notion AI chat transcripts were exposed in an incident at its vendor Fireworks AI, and that fewer than 1 percent of Notion workspaces may be affected. Notion said Fireworks notified it of the incident on November 6 and confirmed the new dataset on November 8, and that Notion had disconnected from Fireworks as of November 3.

Forbes reported on November 9 that two Fireworks AI customers said they had moved business off the company's systems after a series of hacks. The reporting describes a company under pressure to explain what was accessed and how it is responding. Fireworks sits in the middle of the AI supply chain: companies send prompts and data to providers like it so that open weight models can run quickly and cheaply, and the data in transit is often sensitive.

The Notion disclosure is a useful case study in how third party AI risk actually looks. The exposed chats were described as unattributable, meaning they could not be tied to a specific author or workspace, which limits the harm to individuals. But the incident shows how much information flows to vendors that customers of the end product never see. A Notion user typing into an AI feature has no direct relationship with Fireworks and no way to know the vendor was involved until a notice arrives.

The timeline also shows how slowly these things surface. Notion disconnected from Fireworks on November 3, but heard about the problem from Fireworks on November 6 and had the dataset confirmed on November 8. Anyone relying on a vendor's notification to start their own response is already days behind.

For executives, the lesson is about the supply chain behind every AI feature. Inventory the inference providers, model hosts and tooling vendors that sit behind the AI products you buy and build. Ask them for incident notification commitments in contract language, including how quickly they must tell you, what they must tell you, and whether they must identify their own subprocessors. Decide ahead of time which data is allowed to go to a third party model host at all. Finally, give employees clear guidance about what not to paste into AI tools, because credentials and customer data entered into a chat window can end up in places no one planned for.

The reputational stakes are plain. Two customers leaving over security concerns is the kind of evidence that moves procurement committees, and it will make security posture a more visible part of how inference providers compete.

AI Safety Story 5 of 12

VulnCheck Audits Anthropic's Glasswing: Hundreds of CVEs, Few Exploited

Seven months after Anthropic launched Project Glasswing, its effort to find software vulnerabilities with AI, security firm VulnCheck has published an independent accounting of the results. The picture is mixed. VulnCheck said it has tracked 442 CVEs attributable to Glasswing, about 0.5 percent of the 93,450 CVEs published so far in 2026. VulnCheck also said 2 of those 442 appear in its Known Exploited Vulnerabilities catalog, meaning attackers are known to be using them.

The headline for executives is the gap between volume and impact. VulnCheck said Anthropic's public Glasswing ledger listed 37,966 findings as of November 2, of which 7,349 were public and 217 were fixed. Anthropic's ledger claims a 94.2 percent true positive rate as of the same date, according to VulnCheck. Set those numbers next to each other and you can see the pressure point. A system that generates tens of thousands of findings can overwhelm the people who have to confirm and fix them, and fixes lag far behind discovery.

The sharpest data point involves curl, the widely used open source data transfer tool. VulnCheck said that of 34 Anthropic reports to the curl project that had been published, 19 were judged not a vulnerability and 6 were confirmed low severity. That is a stark contrast to the ledger's claimed accuracy, and it illustrates a long standing tension in automated security research. A tool can be accurate on paper while still creating work for volunteer maintainers who must read each report and reject the ones that do not hold up.

It is worth being fair to both sides. VulnCheck is an outside analyst with its own methodology, and Anthropic's ledger reflects Anthropic's own review. The two measure different things. One counts reports that survive scrutiny by maintainers, the other counts findings by the vendor's own standard. Neither is necessarily wrong, and the disagreement itself is the finding.

For chief information security officers, the practical guidance is to judge AI security tools by outcomes that matter to you rather than by volume. Ask vendors how many findings were independently confirmed, how many were fixed, and how many were later withdrawn. Measure the triage burden these tools create for your own team, and budget for the human review that sits behind every automated finding. The promise of AI assisted vulnerability discovery is real, but the bottleneck has moved from finding bugs to verifying and fixing them, and that is where the next round of investment needs to go.

AI Research Story 6 of 12

Stanford's AI100 Study Publishes Its Third Report on the Scaling Era

The One Hundred Year Study on AI at Stanford released its third report in November, titled "AI Goes Mainstream: The Scaling Era and the Rise of Generative AI." The report runs 127 pages, according to the summary published by Stanford and reposted by infoDOCKET on November 10. The 2026 study panel was led by Michael Wooldridge of the University of Oxford.

AI100 is a different kind of document from the usual stream of vendor announcements and benchmark leaderboards. It is a long running, academic effort designed to take stock of the field at regular intervals, written by a panel of researchers rather than by companies with products to sell. That independence is the reason executives should pay attention to it. When the same technology is described every week by the firms that profit from it, a periodic review by outside experts is a useful corrective.

The title itself is informative. By framing the period as "the scaling era," the panel is describing a stretch in which progress has come largely from training larger models on more data with more computing power, and by describing AI as having "gone mainstream," it is acknowledging that generative systems have moved from research curiosity to everyday business tool. Both phrases carry a message for boards. The first suggests that the dominant strategy of recent years has a defined character and may not be the only path forward. The second suggests that the question is no longer whether to adopt AI but how to govern the adoption that is already underway across your workforce.

Readers should note what this report is and is not. It is a synthesis, not a buying guide. It will not tell a company which model to purchase or which vendor to trust. What it can offer is context: a structured, expert view of where the technology stands and which questions the research community considers open.

A practical way to use it is to hand the executive summary to your AI steering committee and ask for a short written response on three points. Which of the report's themes matter most to your industry. Where your current plans assume that recent trends will simply continue. And which risks you are managing by policy that you have not actually tested. A 127 page report will not answer those questions, but it can make sure your team is asking them with the same shared vocabulary as the researchers shaping the field.

AI Research Story 7 of 12

Nvidia's GR00T 2 Reaches 71 Percent Success on 50 Robot Tasks From One Demonstration

Nvidia used the Conference on Robot Learning in Austin, Texas to introduce GR00T 2, its robot model. Nvidia said GR00T 2 averages 71 percent success across 50 tasks with one demonstration per task, rising to 95 percent with five demonstrations. The company also said GR00T 2 completed 40 of 50 evaluated tasks at least once without task specific demonstrations, and that it was trained on 100,000 hours of Nvidia open source human egocentric data, meaning video recorded from a person's point of view.

The significance is in the training approach. Teaching a robot a new job has traditionally required collecting many demonstrations for that exact job, which is slow and expensive. A model that can pick up a task from one example, and sometimes from none, moves robotics closer to the economics of software, where a new capability is a configuration change rather than a data collection project. The reliance on human first person video is the key idea. People are producing far more footage of hands doing real work than any lab could capture with robots, and a model that learns from it widens the supply of training material.

Nvidia paired the model with tooling. The company said Isaac Lab 3.0 with the Newton physics engine achieved up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX on a vision based dexterous manipulation benchmark. Faster simulation matters because robots increasingly practice in virtual environments before they touch the physical world, and every improvement in throughput shortens the cycle between idea and test. Nvidia also said two of its papers, GRAIL and RoboTTT, are among six Best Paper finalists at the conference, a signal of research credibility.

Executives should keep the claims in proportion. These are Nvidia's own benchmark results, measured on a defined set of 50 tasks, and a 71 percent average success rate still means roughly three attempts in ten fail. A warehouse or factory needs reliability far higher than that for tasks that run thousands of times a day. The gap between a research benchmark and a production line is where most robotics projects stall.

Even so, the direction is clear, and the strategic implication is for operations leaders. If robots can learn from a handful of demonstrations, the cost of piloting automation in a new process falls sharply, and the bottleneck shifts from engineering to deciding where the work is worth automating. Companies with manufacturing, logistics or inspection operations should identify two or three repetitive physical workflows now, so that they are ready to run controlled pilots as the technology matures and the software costs continue to drop.

Enterprise AI Story 8 of 12

Google Launches Gemini for Healthcare With Seven Launch Customers

Google launched Gemini for Healthcare on November 10, with Northwell Health, Highmark Health, HCA Healthcare, Optum, Blue Cross Blue Shield of Massachusetts, Hartford HealthCare and Stanford School of Medicine as launch customers, according to Fortune. The list spans hospital systems, insurers and an academic medical school, which tells you the product is aimed at the administrative seams between providers and payers as much as at clinical care.

Fortune reported that Gemini for Healthcare can read Epic records but cannot change them, and that clinicians make all updates. That design choice is deliberate and important. Electronic health records are among the most sensitive and most regulated data a company can touch, and a read only connection sharply limits what can go wrong. It also keeps accountability with the licensed professionals who sign off on changes. For a regulated industry, a tool that suggests rather than writes is far easier to approve.

The business case rests on administrative cost. Fortune quoted Northwell Health CEO John D'Angelo saying about 400 employees work solely on fighting insurance denials and that Northwell overturns over 90 percent of them. A success rate that high suggests many denials are errors or paperwork failures rather than genuine disputes, which is exactly the kind of repetitive, rules heavy work that software can accelerate. Fortune also cited an American Medical Association figure that each physician completes about 40 prior authorizations a week, taking about 13 hours of physician and staff time. Prior authorization is the process by which a doctor must get an insurer's approval before certain treatments, and it is a notorious source of delay and burnout.

The same technology that helps a hospital prepare an appeal also helps an insurer review one faster. That is the quiet tension in the launch. Providers and payers are both customers, and each can use the tool to strengthen its side of a dispute. Whether the net effect is lower costs for patients or simply a faster arms race depends on how organizations choose to use the time savings.

For healthcare executives, the practical steps are to start with a single high volume administrative workflow, such as prior authorization or denial appeals, and to measure turnaround time and error rates before and after. Insist on audit trails, clear human sign off and a documented policy on what patient data the system may see. For leaders outside healthcare, the pattern is worth studying: AI is being adopted first where there is a large, measurable administrative burden and where a read only deployment keeps risk contained.

Industry Dynamics Story 9 of 12

Nvidia Reportedly Tightens Overseas Cloud Rentals to Chinese Customers

Nvidia has reportedly told cloud partners worldwide that at least 51 percent of their services must go to non Chinese firms, according to the Financial Times, as relayed by Tom's Hardware. Nvidia has neither confirmed nor denied the rule. If accurate, it targets a loophole in export controls, in which Chinese companies that cannot buy the most advanced Nvidia chips directly rent that computing power from data centers outside China.

The economics of the workaround are visible in the price of hardware. Tom's Hardware reported that a Nvidia B300 rack selling for about $600,000 internationally reportedly fetched $2.8 million in China last month. A markup of that size, more than four times the international price, shows how strong the demand is and how large the incentive to find a path around the controls. It also suggests that supply restrictions are working in one sense, by making advanced chips scarce and expensive in China, while generating a gray market that rewards intermediaries.

The significance for executives is that the rules governing AI computing are being enforced through commercial contracts as well as government regulation. A quota imposed by the dominant chip supplier on its own customers is a form of private policy, and it can change the availability and pricing of cloud capacity for companies that have nothing to do with China. Cloud providers that must reserve a share of capacity for non Chinese clients will plan their data center footprint and customer mix differently, which can ripple into contract terms for everyone.

Several cautions apply. The report rests on verbal instructions that Nvidia has not publicly confirmed, so the details could differ from what is described, and the details of enforcement are unclear. Readers should treat the 51 percent figure as reported rather than established.

The broader lesson is that AI infrastructure has become a geopolitical asset. Companies that depend on cloud computing for AI workloads should track export control developments and ask their providers how compliance obligations might affect capacity, pricing and service levels. Build contingency plans that allow workloads to move between regions and providers, and avoid assuming that capacity available today will be available on the same terms next year. In a market where the most important component is scarce, policy decisions made far from your data center can show up quickly on your invoice.

AI Models Story 10 of 12

Meituan Open Sources LongCat 2.5, a 1.6 Trillion Parameter Model Under the MIT License

Chinese technology company Meituan released the open weights of LongCat-2.5 on Hugging Face under the MIT License on November 9, according to VentureBeat. An MIT license is among the most permissive in software, allowing companies to use, modify and build on the model with few restrictions. Meituan's model card describes LongCat-2.5 as a mixture of experts model with 1.6 trillion total parameters and about 48 billion active per token. In a mixture of experts design, only a fraction of the model runs for any given piece of text, which keeps the cost of running it far below what the total parameter count would suggest.

The benchmark results show a large jump from the previous generation. VentureBeat reported that LongCat-2.5 scores 88.4 on Terminal Bench v2.1 versus 70.8 for LongCat-2.0, and 64.6 on DeepSWE v1.1 versus 9.7. Both are tests of agentic software work, where a model has to operate a command line or resolve real coding tasks. A move from 9.7 to 64.6 on one of them is the kind of gain that can reflect a change in training method rather than incremental tuning. As with any vendor reported result, outside testing will determine how well the numbers hold up.

Price is the other headline. According to VentureBeat, Meituan's hosted API for LongCat-2.5 is priced on a limited time basis at $0.30 per million uncached input tokens and $1.20 per million output tokens. Compare that with the $8 and $40 Anthropic charges for Claude Fable 5.2 and the gap is more than 25 times on input and more than 30 times on output. A comparison like that does not mean the models are equivalent, and the LongCat price is described as limited time, but it does show how much pressure open weight models from Chinese developers are placing on the pricing of frontier systems.

For executives, the decision is more nuanced than cost. Open weights let a company run the model on its own infrastructure, which can address data residency and privacy concerns. They also bring questions that procurement and legal teams must answer. Where does the model come from, what are the regulatory and reputational implications of using a model developed in China, and does your industry have rules about it? Those questions apply even when the license is permissive.

A pragmatic approach is to evaluate open weight models as a negotiating lever and a resilience option. Testing one on a non sensitive workload gives you a real cost and quality comparison to bring to your main vendor, and it builds the internal skills to switch if prices or terms change.

Funding & Investment Story 11 of 12

Tendrils Raises $30 Million From Benchmark to Redesign the CPU for Parallel Work

Chip startup Tendrils announced a $30 million round led by Benchmark, with Visionaries Club, the UK Sovereign AI Fund, Conviction and SV Angels among the participants. Tendrils said Jack Altman is its newest board member and that it is building a processor with thousands of cores working in parallel, plus an open source language called Vine.

The pitch addresses a mismatch at the heart of modern computing. Conventional processors have a modest number of powerful cores designed to run one instruction stream quickly, while the work AI demands is massively parallel. Graphics processors have filled that gap, but they are specialized and programming them requires specialized skill. Tendrils is betting on a different design, a processor that spreads work across thousands of cores, paired with a programming language built for parallel execution from the start. The open source language is a notable part of the plan. New chip architectures often fail not because the silicon is weak but because developers lack tools to use it, and releasing the language openly is an attempt to solve that adoption problem early.

The investor list is telling. Benchmark leading a hardware round is a sign that top venture firms are willing to fund challengers to established chip makers, a category long seen as capital intensive. The participation of the UK Sovereign AI Fund adds a government dimension. Public money is increasingly flowing into computing infrastructure that countries want to control, and a chip company backed in part by that fund fits a wider push to reduce dependence on a small number of suppliers.

Executives should place this in context. A $30 million round is small by chip industry standards, where developing and manufacturing a processor can cost far more. The announcement describes ambition and a team, not a product that is ready to buy, and the company has a long road from design to volume production. Most startups in this field will not succeed against entrenched competitors with enormous scale advantages.

Still, the funding is a data point in a larger story. The scarcity and cost of AI computing are driving investment into alternatives at every layer, from new chip designs to new programming models. Companies planning multi year AI infrastructure strategies should avoid assuming that the current architecture will be the only one available. Keep software portable, track emerging hardware options, and treat any single supplier relationship as a risk to be managed rather than a permanent fact.

Policy & Regulation Story 12 of 12

Scott Wiener Wins Pelosi's Seat and Brings an AI Safety Agenda to Congress

MIT Technology Review reported on November 9 that Scott Wiener won the congressional race for Nancy Pelosi's open San Francisco seat. Wiener, a former California state senator, is one of the most prominent legislators on AI policy in the country, and his arrival in Congress puts a leading author of state level AI safety law into the national debate.

His record explains why the technology industry is watching. Wiener authored California's SB 1047, which Governor Gavin Newsom vetoed in 2024, and then sponsored SB 53, which Newsom signed. The pattern is instructive. A broad bill was rejected, a narrower one became law, and both have influenced federal proposals. For companies, the lesson is that AI regulation tends to arrive in iterations, with early drafts shaping what later versions include.

Funding adds a political layer. MIT Technology Review reported that Wiener's campaign received over $3 million from Public First Action, a nonprofit that supports AI regulation and receives substantial funding from Anthropic. Opponents criticized that support, and Wiener rejects the criticism. The episode shows that AI companies are now participating in elections, and not always on the same side. Some back lighter regulation, while others fund candidates who want stricter rules.

The article described Wiener's stated priorities. He backs a strong independent federal AI regulator, whistleblower protections, testing and auditing requirements, and a kill switch requirement, and he opposes broad federal preemption of state AI laws. That last point matters most for business. Preemption would replace the current patchwork of state rules with a single national standard, which many companies say they want for simplicity. Opposing it signals that Wiener sees state laws as a floor that federal legislation should not remove.

MIT Technology Review also reported that voters approved 40 of 41 anti data center ballot measures on election day. Wiener says those measures reflect local concerns more than fear about AI risk. Whatever the motive, the result shows that the physical footprint of AI, including land, power and water, is becoming a local political issue, and companies planning data center expansion should expect organized resistance in more places.

For executives, the practical response is to prepare for requirements rather than argue about whether they will come. Maintain documentation of how your AI systems are tested, assign clear responsibility for model risk, and set up a channel for employees to raise concerns, since whistleblower protections are among the ideas on the table. Track both state and federal activity, because federal action is uncertain and state capitals may keep writing the rules in the meantime.