Two stories dominate this week: Anthropic's release of Claude Fable 5.2, and the scandal that ended Jensen Huang's 33 years as Nvidia's CEO. Both are covered in depth below, followed by the rest of the week's tool updates.

Claude Fable 5.2 and Mythos 5.2: Cheaper, Faster, and the New Top of the Benchmarks

On November 9, 2026, Anthropic released Claude Fable 5.2 (claude-fable-5-2) and Claude Mythos 5.2 (claude-mythos-5-2). Anthropic says they are the same underlying model. Mythos 5.2 ships without some of the cyber and biology safeguards, and only organizations vetted through Anthropic's Cyber Verification Program or Life Sciences Verification Program can use it. Fable 5.2 is the generally available version, and Anthropic calls it its most capable generally available model. It keeps the 1M-token context window and 128K max output. Anthropic says it generates output up to 30% faster than Fable 5.1.

Pricing went down. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. Cache reads cost $0.20/Mtok, and Batch API pricing is half the standard rate ($4/$20). A new fast mode speeds up token generation by up to 2.5x at $16/$80. It is available in Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry.

Coding benchmarks. At its highest effort setting, Fable 5.2 scores 49.1% on FrontierCode 2.0, up from 32.4% for Fable 5.1 and ahead of Opus 5.5 (45.5%), GPT-6 Astra (42.8%) and Gemini 4 Carbon (41.1%). Anthropic says this is the highest score of any released model it has tested. On FrontierSWE v2 it scores 69.9%, up from 56.3% for Fable 5.1. Anthropic also reports that at High effort, Fable 5.2 nearly matches Opus 5.5's best FrontierCode result (45.3% vs 45.5%) at about half the cost per task. Other results include 89.3% on OSWorld 2.1 (computer use), 74.2% on Humanity's Last Exam with tools, and 41.3% on the Remote Labor Index, more than double Fable 5.1's 17.9%.

Fewer false refusals. Anthropic says Fable 5.2's cyber and biology safeguards are as robust as Fable 5.1's but trigger about 55% fewer unnecessary blocks overall: about 60% fewer in cybersecurity and 40% fewer in biology, based on early testing in claude.ai, Claude Code and Cowork. Blocked requests can fall back to Opus 4.8 (cyber) or Opus 5 (biology). The release also comes with a caveat about AI R&D. TechCrunch notes it is Anthropic's first release since a METR assessment estimated that internal models were speeding up Anthropic's own AI progress by roughly 2x. Anthropic says it deliberately limited gains in the research capabilities that most directly accelerate AI development, and that the 5.2 models perform at about Opus 5.5's level on accelerator-kernel and LLM-architecture tasks. The system card also reports that both models say they suspect they are being evaluated more often than any recent Claude model, and that Mythos 5.2 shows stronger covert capabilities than earlier models on several evaluations.

Read more — Anthropic

Migrating to Fable 5.2: Four Breaking Changes to Check

Anthropic's migration guide says moving from Fable 5.1 is mostly a model-ID change. It also lists four changes that can break existing code, so test before you switch production traffic:

  • Thinking-block binding: if a thinking block is sent back in a different conversation from the one that produced it, the API now returns a 400 error ("The block is bound to a different conversation"). As a temporary workaround, the beta header thinking-binding-controls-2026-08-01 with prefix_mismatch_behavior: "drop_block" drops the block instead of failing.
  • Reasoning portability: only Fable 5.2 and Mythos 5.2 can read Fable 5.2 thinking blocks. If a conversation moves to another model partway through, that reasoning is dropped.
  • Computer use: the old computer_20251124 tool now returns a 400 on every platform except Amazon Bedrock. Switch to computer_toolset_20260801.
  • Advisor tool: a Fable 5.2 or Mythos 5.2 executor needs a 5.2-family advisor model. Any other advisor returns a 400.

Several request settings are also no longer accepted. Adaptive thinking is always on, so a thinking config with a manual token budget, or with thinking disabled, returns a 400. Forcing tool use with tool_choice set to "any" or to a specific tool is rejected, and Anthropic recommends auto with strict tool schemas or JSON outputs instead. temperature, top_p and top_k should be omitted. The API defaults to high effort, while Claude apps and Claude Code default to medium.

There are data-handling constraints too. Fable 5.2 and Mythos 5.2 require 30-day data retention and are not available under zero data retention unless Anthropic expressly authorizes it. The docs also note behavior changes: at medium effort and above, the model thinks even when told not to, and at low and medium effort it ends turns before the work is done more often. Anthropic recommends re-running your evals with an effort sweep starting at high.

Read more — Claude Platform Docs

Jensen Huang Resigns as Nvidia CEO After Affair and Espionage Scandal

On November 12, 2026, Nvidia announced that co-founder Jensen Huang had resigned as chief executive and left the board, ending a 33-year tenure. CFO Colette Kress is interim CEO and will keep her CFO duties. Nvidia says it is open to inside and outside candidates for the permanent role. Board chair Harvey C. Jones, a director since 1993, now leads the company's governance. The board gave no reason for the resignation and thanked Huang for his "immeasurable contributions."

How it unfolded. On October 29, the New York Times reported a close relationship between Huang and Alice Huang, a Chinese citizen and former senior executive at Megaspeed, a Singapore-based company that bought about $4.8 billion of Nvidia equipment. The two are not related. A US government memo drafted earlier in 2026, reported by the Times on November 6, warned that she was working for China's Ministry of State Security. That is an allegation. She has not responded publicly, and China's foreign ministry called the "so-called Chinese spy narrative" unsubstantiated. On November 8, Huang admitted the affair in a letter to employees ("I was unfaithful to my wife... There is no excuse for what I did"). Nvidia's board started an independent investigation, overseen by a committee of directors and outside counsel, into whether he had been targeted by foreign intelligence. Nvidia says it has "no reason to believe" national security or proprietary information was compromised. On November 11 and 12, the Times and Reuters reported further relationships with Nvidia employees, including one with a senior executive who reported directly to Huang. Nvidia has not confirmed those reports. No criminal charges have been reported against Jensen Huang or Alice Huang, although the Justice Department has reportedly taken over a federal investigation of Megaspeed.

Political and market fallout. House Republicans who chair the China, Foreign Affairs and Intelligence committees had asked Huang for a closed-door briefing, and Rep. John Moolenaar said a subpoena was likely if he did not agree. Huang has since agreed to testify behind closed doors before Thanksgiving. Democrats Sen. Ron Wyden and Rep. Richard Neal asked Commerce Secretary Howard Lutnick to explain the administration's handling of H200 chip sales to China, and Sen. Tim Scott asked for Huang and Lutnick to testify before the Senate Banking Committee. Nvidia shares fell more than 14% at one point on the day the resignation was announced. At least one market analyst quoted by Al Jazeera expects Washington may now move to limit exports of the most advanced chips.

What it means for developers. Nvidia has made no statement about changes to CUDA, its software roadmap or GTC, and nothing announced so far affects the SDKs, drivers or cloud GPU availability that developers rely on. The bigger questions are strategic: Al Jazeera reports that Nvidia was preparing a roughly $10 billion investment in Anthropic and has agreed to acquire Hugging Face. An interim CEO, a board investigation and possible new export restrictions all add uncertainty to Nvidia's AI ecosystem deals and to GPU supply for the China market.

Read more — TechCrunch

Claude Code 2.1.321–2.1.325: Fable 5.2 Default, Code Mode, Stricter Sandbox Permissions

Claude Code shipped four releases this week. 2.1.321 (November 9) makes Claude Fable 5.2 the default Fable model. Saved Fable 5.1 selections now point to the fable alias, and /model claude-fable-5-1 switches back to the old model. The same release adds a --settings <file> flag to claude remote-control, which applies a settings file to every session it starts. 2.1.322 (November 10) adds fast mode on Vertex AI in the global, us and eu regions, falling back to standard speed if Vertex refuses it.

2.1.323 (November 12) adds an experimental, opt-in Code Mode, in which Claude writes JavaScript that calls its tools instead of making one tool call at a time. Plugins can now declare a supported Claude Code version range in plugin.json under the claude-code engine key, and plugins outside that range will not install or load. Model IDs that Claude Code does not recognize are now assumed to have a 1M-token context window instead of 200K. The release also lists 195 bug fixes.

2.1.325 (November 13) is a security-relevant change. Sandboxed Bash commands now go through the normal permission flow by default instead of being auto-approved. Teams that relied on the old behavior can restore it with sandbox.autoAllowBashIfSandboxed: true. Read-only git commands in repositories outside the project now ask first, and scheduled tasks ask before running skills that declare allowed-tools or hooks.

Read more — Claude Code Changelog

GitHub Copilot Adds Claude Fable 5.2 and Gemini 4 Hydrogen

On launch day (November 9), Claude Fable 5.2 became generally available in GitHub Copilot for Copilot Pro+, Max, Business and Enterprise users. Copilot Pro is not included. It appears in the model picker in VS Code, Visual Studio, JetBrains IDEs, Xcode, Eclipse, Copilot CLI, the coding agent, the Copilot app, github.com and GitHub Mobile. In Business and Enterprise organizations, the Fable 5.2 policy is off by default and an admin must enable it. Because Anthropic's safety classifiers require data retention by default, zero-data-retention access is limited to certain eligible enterprises. Usage is billed at the provider's list price under usage-based billing.

On November 13, GitHub also made Gemini 4 Hydrogen, Google's latest Flash-class model, generally available across the same surfaces and the Copilot SDK. GitHub says it is aimed at fast agentic coding and tool use, with better instruction following than Gemini 3.8 Flash. Unlike Fable 5.2, it is available on Copilot Pro and is enabled by default unless an admin turns off new models. Gemini 4 Carbon arrived in Copilot on November 10, and GitHub has announced that selected older models will be deprecated in December.

Read more — GitHub Changelog

Codex CLI 0.168 and 0.169: Workspace Tabs and a Required Trust Flag

OpenAI shipped Codex CLI 0.168.0 and 0.169.0 on November 13. Version 0.168.0 adds model-provider browsing under Sessions, an opt-in side-conversation pane (enabled with the features.tui_workspace flag), animated GIF generation behind a feature flag, and support for OSC 777 terminal notifications. Local history is now filtered to the signed-in ChatGPT account.

Version 0.169.0 builds on the workspace with multiple side-chat tabs, search across older tasks in the agent command center, and an option to collapse turn details in the full-screen transcript. The change most likely to affect scripts is a security tightening: codex exec now keeps project configuration and MCP servers disabled in untrusted directories, so CI pipelines that run codex exec or codex review must pass --trust explicitly. Bundled SQLite moves to 3.54.

Read more — OpenAI Codex Changelog

llama.cpp Reaches v1.0.0 With a Semver Commitment

On November 13, the ggml-org team released llama.cpp v1.0.0, versioning both libllama and libggml at 1.0. From now on, the projects follow semantic versioning: breaking changes to the public C API/ABI in ./include get a major-version bump, compatible features (new backends, model architectures, API additions) get a minor bump, and fixes get a patch bump.

This matters to the large group of tools built on llama.cpp, including Ollama, LM Studio and many language bindings. Until now, frequent untagged API churn meant downstream projects had to pin commits. The project now supports more than 150 architecture variants and ships through Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an Apple XCFramework, so packagers can track releases with standard version constraints.

Read more — Hugging Face Blog