Today’s through-line runs from a model lab’s new technical models, to a cloud provider’s expansion of its eighth-generation accelerator family, to a robotics foundation model aimed at dexterous work, to a lightweight model built for agent execution, and to a shift in how enterprises can buy software: Anthropic on November 9 released Claude Fable 5.2 and Claude Mythos 5.2, the same underlying model, with improvements in coding, reasoning and vision, outputs up to about 30% faster than 5.1, and a lower price than the previous generation. Google the same day, in a post by Amin Vahdat, introduced TPU 8tx and TPU 8i Max, expanding the eighth-generation TPU lineup to four systems in a single year. NVIDIA at the Conference on Robot Learning (CoRL) unveiled GR00T 2, an open frontier world action model built on Cosmos, and discussed simulation and evaluation tools including Isaac Lab 3.0. Meta launched Muse Ember, its smallest and most efficient Muse model, for tool calls, coding subtasks, routing, extraction and long-document parsing. OpenAI listed ChatGPT Enterprise in AWS Marketplace, allowing organizations to buy and manage the product through their existing AWS purchasing and billing relationship.

1. Anthropic launches Claude Fable 5.2 and Claude Mythos 5.2

On November 9, 2026, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, describing the models as designed to tackle the most daunting technical challenges. The company wrote that they are the same underlying model: Fable has strong cybersecurity and biology safeguards, which allows Anthropic to make it available to all users; organizations pursuing beneficial applications in cybersecurity or biomedical research can apply for Mythos through the recently expanded Life Sciences and Cyber Verification Programs.

Anthropic said Fable 5.2 and Mythos 5.2 show improvements in areas such as coding, reasoning and vision, generate outputs up to about 30% faster than Claude Fable 5.1 and Claude Mythos 5.1, and are available at a lower price: $8 per million input tokens and $40 per million output tokens. An optional fast mode speeds token generation by up to about 2.5 times and is priced at $16 per million input tokens and $80 per million output tokens.

In prerelease testing, Anthropic cited examples it presented as illustrations of the models’ capabilities: engineers used them to complete software migrations that would otherwise have been infeasible; the models designed three working electronic devices from scratch; they solved two open problems in the physics of black holes; they made substantial progress on robotics evaluations; and a team of about 1,000 Fable 5.2 agents formalized the proof of Fermat’s Last Theorem in 30 hours with minimal human intervention. In the same setup, Opus 5.5 completed about 26% of the formalization at 18 hours.

Benchmarks Anthropic reported—showing results with safeguards on unless otherwise noted—include FrontierCode 2.0 at Xhigh effort of 49.1%, which the company said was the highest score of any released model it tested; FrontierSWE v2 at 69.9%; AutomationBench at 51.4%; Terminal-Bench-Science 0.1 at 74.6% for Fable and 75.3% for Mythos; CyScenarioBench at 76.3% for Mythos; CVE Rediscovery at 89.6% for Mythos; OSWorld 2.1 (partial) at 89.3%; RoboDojo Sim at 64.9%; Humanity’s Last Exam with tools at 74.2%; and Remote Labor Index at High effort of 41.3%.

Anthropic recommended that users start with Claude Opus 5.5, its strongest daily driver, for most coding and knowledge work, then move up to Fable or Mythos—or down to Claude Sonnet 5.5 or Claude Haiku 5.5—depending on the scope and complexity of the task. Fable 5.2 can also serve as Advisor, where smaller models consult it on harder problems, or as Orchestrator, where it delegates simpler tasks to smaller models.

On safety, the company said it implemented its strongest-ever protections against distillation attacks and that data harvested by distillers has been reduced by 90% to 99%; preserved thinking now applies to all accounts. During training, Anthropic took steps to limit advances in research capabilities that could most directly accelerate AI development—including accelerator-kernel optimization and improvements to large-language-model architectures—to about the level of Opus 5.5, while leaving related capabilities such as agentic coding, hardware development and interpretability research intact. Enterprise Frontier Safeguards is now in private beta with customers including Bridgewater, Salesforce and Stripe.

Claude Fable 5.2 is available today on platforms including Amazon Web Services, Google Cloud and Microsoft Azure, under the model id claude-fable-5-2. Mythos is available through the verification programs; Claude Enterprise customers can also use Claude Security, now powered in part by Mythos 5.2.

Takeaway: Anthropic has split one underlying model into Fable for general use and Mythos for verified organizations, pairing a lower price and faster output with the hardest technical work—while holding the capabilities most likely to accelerate AI’s own development to about the level of Opus 5.5, and still recommending Opus 5.5 as the everyday default.

2. Google introduces TPU 8tx and TPU 8i Max, expanding the eighth-generation lineup to four systems in one year

On November 9, 2026, Amin Vahdat, Google’s senior vice president and chief technologist for AI and infrastructure, wrote that the company is introducing TPU 8tx and TPU 8i Max, expanding the eighth-generation TPU portfolio to four systems in a single year: 8t for large-scale training, 8i for real-time reasoning, the new 8tx for confidential and secure inference as well as small-scale training, and 8i Max for massive context windows and what Google called the highest compute density per rack. Vahdat wrote that TPUs now have nearly 100 external customers, including frontier labs, trading firms, retailers and universities. The Verge reported the same day on how the two new systems are positioned. The blog also noted a broadcast at 9 a.m. Pacific Time that day.

Of TPU 8tx, Google said the system builds confidential compute into the hardware layer, keeping sensitive data and model weights protected and isolated even while in use. It is built on the same silicon as 8t and paired with 5th Generation AMD EPYC hosts, and it adopts the Boardfly network created for real-time reasoning, delivering up to about 2x lower latency than 8t with a near-zero performance penalty. The company said organizations can fine-tune models on private data, distill larger models into smaller ones and train agents with reinforcement learning on a single system, without migrating sensitive data between separate clusters.

Of TPU 8i Max, Google said it combined two low-latency TPU 8i chips onto a single package, delivering double the compute, double the SRAM and double the high-bandwidth memory (HBM) bandwidth of TPU 8i, so data centers can deploy more computing power per square foot and per watt and keep large context windows from stalling against the memory wall.

On software, Google said it has open-sourced TorchTPU so that PyTorch code and Hugging Face models run natively on TPUs; a TPU agent suite is in preview; it introduced the XLA Serving Engine and the Agent Compute Engine; and it is open-sourcing TPU Sync, described as the model-state layer for TPU clusters. Cloud Pathways is now generally available on Google Kubernetes Engine, as is GKE dynamic slicing.

Takeaway: Rather than waiting for the next generation of silicon, Google has carved two new systems out of the same eighth-generation chips—one by wrapping confidential compute around inference and small-scale training, the other by packaging two chips together to chase long context and rack density.

3. NVIDIA unveils GR00T 2, an open world action model built on Cosmos, at CoRL

On November 9, 2026, NVIDIA introduced GR00T 2 at the Conference on Robot Learning (CoRL) in Austin, Texas. The company blog described it as an open frontier world action model built on NVIDIA Cosmos, one that predicts future video and robot actions together. The research page said the model was trained on diverse robot and human data; the pretraining mixture includes 402,730 hours of egocentric human video, 10,184 hours of handheld-gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation, with 100,000 hours of open-source human egocentric data released as Human-1M-100K.

Metrics NVIDIA cited include completing 40 of 50 evaluated tasks at least once without task-specific demonstrations, and an average success rate of 71% with one demonstration per task, rising to 95% with five. The research page also reported 40% average success in a zero-shot setting, and said agents can use an application programming interface to run skills in sequence and recover from errors. Around the 95% figure, reporting citing NVIDIA’s research materials said the evaluated tasks were selected among those the model could learn with five episodes at a success rate of at least 80%, so that average should be read together with this filter.

Partners testing the model include Nexuni for robotic laundry handling, Foxconn for factory automation in server manufacturing, and Siemens across industrial use cases.

NVIDIA also released Isaac Lab 3.0 with a Newton backend, allowing the simulation framework to model deformable objects and other more intricate physics. The company said that on a vision-based dexterous-manipulation benchmark, Isaac Lab 3.0 with Newton achieved up to about 2.5 times the training throughput of Isaac Lab 2.3 with PhysX. AlphaNIST uses a single policy to learn all 20 tasks on the NIST Assembly Task Board, reaching more than 98% training success compared with about 81% for separate per-task policies, and using roughly 39 times less GPU time. The company also introduced industrial benchmarks including DexBench and VABAR, and Graph-as-Policy (GaP), an agentic robotics framework developed with UC Berkeley.

Takeaway: NVIDIA has put “seeing what happens next” and “deciding how to move” into one open model, and has used large amounts of first-person human video to fill gaps in robot demonstrations—while reporting high few-shot success rates that come with a footnote about how the evaluated tasks were selected.

4. Meta launches Muse Ember, a lightweight model for agent execution

On November 9, 2026, Meta introduced Muse Ember, describing it as the smallest and most efficient Muse model, designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long-document parsing. The company claimed that Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting, and that it sits on the efficiency frontier of the Artificial Analysis leaderboard. TechCrunch reported the launch the same day, noting that Ember is designed to pair with the larger Muse Wonder inside an agentic harness.

For prompts of up to 200K tokens, pricing is $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens; longer prompts cost $0.30, $0.012 and $1.50 respectively. The model has a 1 million-token context window and can produce up to 512K output tokens, with native support for text, images, audio and video. It is available as muse-ember-1.0 on Muse Code and the Meta Model API, and also on Google Cloud and Oracle Cloud.

Evaluation highlights Meta cited at maximum reasoning effort include AutomationBench-AA at 37.8, compared with 13.9 for Haiku and 20.2 for Luna; Agents’ Last Exam at 61.6; Terminal-Bench 4.0 at 40.9; Coding Agent Index-AA at 46; OSWorld 2.1 partial at 69.5; and an Artificial Analysis Index of 36. The company said Muse Ember costs about $0.14 per Artificial Analysis task, compared with about $2.39 for Muse Spark 1.3.

In positioning the model, Meta wrote that “Muse Wonder plans, Muse Ember handles the volume.” It said Ember is the first model trained using Muse Wonder Full, and that evaluations under the Meta Superintelligence Scaling Framework found it remained below the framework’s high-risk thresholds. Partners including Box, Cline and Cloudflare offered early-testing comments on its use for enterprise content, subagent coding and high-throughput workloads; Cloudflare said it ran up to about 30% faster at the tail than the comparable model it tested.

Takeaway: Meta is not asking Ember to unseat its strongest flagship model; it is placing the new system in a division of labor—Wonder plans, Ember executes at volume—and competing on intelligence per dollar and cost per task rather than on peak scores alone.

5. OpenAI lists ChatGPT Enterprise in AWS Marketplace

On November 9, 2026, OpenAI said ChatGPT Enterprise is now available through AWS Marketplace. The company wrote that the listing gives organizations a simpler way to discover, buy and manage ChatGPT Enterprise using their existing AWS purchasing and billing relationship, and to apply eligible spend toward their AWS commitments. The offering includes ChatGPT, ChatGPT Work, Codex and dots; listing information indicates that pricing is set through private offers.

OpenAI said customers buying through AWS Marketplace receive the same models, features and functionality available directly from OpenAI, with new features released at the same time. The products are positioned as complementary to OpenAI models already available in Amazon Bedrock: developers can build applications and agents with those models in Bedrock, while teams across the business use ChatGPT Enterprise in daily work. Eligible customers can also access ChatGPT for Healthcare and ChatGPT for Financial Services through AWS Marketplace.

Takeaway: This is not a model release so much as an opening of a procurement path—placing enterprise ChatGPT on the cloud-commitment bill organizations are already paying, and reducing purchasing friction rather than publishing a public list price.

---

One-line takeaway: Anthropic released Claude Fable 5.2 and Mythos 5.2, the same underlying model, with Fable generally available and Mythos limited to verified organizations, saying outputs are faster than 5.1 and prices are lower; Google introduced TPU 8tx for confidential inference and TPU 8i Max for massive context windows; NVIDIA unveiled the world action model GR00T 2 and Isaac Lab 3.0 at CoRL; Meta launched Muse Ember, a lightweight model for agent execution; and OpenAI listed ChatGPT Enterprise in AWS Marketplace so that eligible spend can count toward existing cloud commitments.