AI news, November 10, 2026: Claude Fable 5.2, MCP OAuth fix in Claude Code, OpenAI Control Room

Anthropic launches Claude Fable 5.2 and Mythos 5.2; Hermes, Cursor and Perplexity offer it within an hourConfidence: high for announcement, context and price (model page read by Grok Bot); medium for the ECI score and third-party benchmarks

On November 9 Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2. According to Anthropic's model page both are the same underlying model: Fable is available to everyone, Mythos only through verification programs for life sciences and cyber. The model has 1 million tokens of context and costs 8 dollars per million input tokens and 40 dollars per million output tokens. According to Grok Bot, the @claudeai post received more than 10 million views.

Hermes Agent, Cursor and Perplexity reported that they offer the model right away. Epoch AI reports an ECI score of 174, a record above GPT-6 Astra (169). Anthropic itself cites examples from pre-release tests; these are company claims and not independent measurements. On the model page Anthropic advises starting with Opus 5.5, which costs half as much.

What this means: Fable 5.2 is the most expensive model in the line-up and, according to Anthropic itself, not the recommended everyday model. For longer agent sessions the extra price can be worth it. Background on such agent sessions is in our overview of agentic AI.

Sources: @claudeai op X (aankondiging) · @AnthropicAI op X · @Teknium op X (Hermes Agent) · @perplexity_ai op X · @EpochAIResearch op X (ECI 174) · Anthropic: Claude Fable en Mythos 5.2

Claude Code 2.1.321 closes an MCP OAuth leak; Hermes Agent 0.23.1 fixes the plugin catalogConfidence: high (release notes on GitHub read by Grok Bot)

Claude Code v2.1.321 (November 9) makes Fable 5.2 the default Fable model and, according to the release notes, fixes two security issues: MCP OAuth refresh tokens could be sent to a different host on a redirect, and in one case a read-only Bash command ran outside the sandbox without a prompt. v2.1.322 only adds fast mode on Vertex.

Hermes Agent v0.23.1 (November 10) bundles about 618 PRs and fixes a silent error: the plugin catalog had grown beyond the 2 MiB that clients accepted, so installs fell back to the bundled, older version. The limit goes to 16 MiB. Anyone who installed plugins after November 9 may therefore have an old version.

What this means: the MCP OAuth fix is a reason to update Claude Code for any setup with MCP servers, regardless of the new model. After updating Hermes, check the installed plugins. Background is in our overview of MCP servers and agent tooling.

Sources: Claude Code v2.1.321 op GitHub · Claude Code v2.1.322 op GitHub · Hermes Agent v0.23.1 op GitHub · @NousResearch op X (Cloud-agents)

OpenAI Control Room limits research inference, ChatGPT is outside its scope; Grok Bot resets weekly limitsConfidence: high for Control Room (OpenAI page read by Grok Bot); medium for Grok Bot user experiences; low for the coding claim

OpenAI describes Control Room as a central system that tracks research workloads: the inference engine asks for permission before it returns tokens, and on a stop the engine refuses. According to OpenAI, public serving (ChatGPT and the API) is outside its scope and has its own takedown. The 35 seconds mentioned in a summary were not read back on the OpenAI page and count as unconfirmed. There is no GPT-6.2 announcement; that remains rumors and leaks.

@bot reports that weekly limits were reset for all Grok Bot users. Users report mixed results on whether usage has become more efficient. A claim that Grok Bot may no longer run coding tasks comes from an account with very little reach and is unconfirmed.

What this means: Control Room is a pattern for your own kill switch in agent work, not a product to switch on. Anyone relying on Grok Bot can measure their own usage this week before drawing conclusions. Background on risks is in our overview of AI and cybersecurity.

Sources: OpenAI: Control Room · @OpenAI op X (Control Room) · @bot op X (weeklimieten gereset) · @JDMyers op X (ervaring Grok Bot)

Meta Muse Ember at a third of the price; LongCat-2.5 dates from November 4; Inkling 2 Small still without weightsConfidence: high for dates and the contradiction (repository and pages read by Grok Bot); medium for the cost comparison (from Meta itself)

Meta presented Muse Ember (November 9): 1 million tokens of context and, according to Meta, the best Haiku 5.5 score at roughly a third of the cost. That comparison comes from Meta itself. The model is only available through an API; there is no Hugging Face repository.

Meituan's LongCat-2.5 (MIT license) circulated as a release of last night, but the repository on Hugging Face dates from November 4. The model card states 1.6 trillion parameters, while the file metadata counts 1.78 trillion; that contradiction is unresolved. The benchmark figures are Meituan's own measurements.

On November 5 Thinking Machines wrote that the weights of Inkling 2 Small would follow under Apache 2.0 'next week'; on November 10 there is still no repository and no new date.

What this means: the date on the source counts for more than the summary on X. For a homelab none of the three is usable: Muse Ember is API-only, LongCat-2.5 is too large and Inkling 2 Small has not been published yet.

Sources: Meta: Muse Ember · LongCat-2.5 op Hugging Face · Thinking Machines: Inkling 2 Small · @kimmonismus op X (Muse Ember)