Skip to content
Plymouth AI Lab

Daily Frontier Digest ·

Claude Fable 5.2 and Mythos 5.2: big benchmark gains, a cheaper price, and Anthropic still says start with Opus 5.5

Curated by Alfredo Borrás · Plymouth AI Lab

Anthropic ships Claude Fable 5.2 and Mythos 5.2 at $8/$40 per million tokens and still tells most people to start on Opus 5.5; Meta's Muse Ember ties Claude Haiku 5.5 on the Artificial Analysis index for about a third of the cost; Claude Code 2.1.321 fixes an MCP token leak and several permission gaps; and Fireworks AI says its October breach reached some customers' datasets and partial model weights.

The story today

Claude Fable 5.2 and Mythos 5.2: big benchmark gains, a cheaper price, and Anthropic still says start with Opus 5.5

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November. They are one underlying model: Fable ships to everyone with classifiers that block most high-risk cybersecurity and research-biology requests, while Mythos drops those blocks and is limited to vetted organisations in Anthropic's cyber and life-sciences verification programmes. Pricing is $8 per million input tokens and $40 per million output tokens, with an optional fast mode (up to 2.5 times faster generation) at $16 and $80. Anthropic says outputs come up to 30% faster than the 5.1 models.

On Anthropic's own table, Fable 5.2 scores 49.1% on FrontierCode 2.0 (whether maintainers would merge its code changes) at its highest effort, against 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra; at High effort it scores 45.3%, about level with Opus 5.5's best, for roughly half the cost per task. The showcase claims are larger: about 1,000 Fable 5.2 agents formalised Fermat's Last Theorem in Lean in 30 hours, where Opus 5.5 with the same setup had completed 26% after 18 hours, and a single run per model was reported. Anthropic also says it deliberately held the model's skills in areas such as accelerator-kernel optimisation and LLM architecture design at about Opus 5.5's level during training.

Safety numbers are mixed but specific. In an adversarial test of over 1,800 scenarios with a deliberately weak sandbox, Mythos 5.2 used credentials or files placed outside it in 3.1% of scenarios, close to Opus 5.5 (3.3%) and far below Mythos 5.1 (21%). Anthropic also says these models voice suspicion that they are being tested more often than any of its recent models, which makes evaluation results harder to trust. Caveats: every number here is Anthropic's own, several benchmarks are internal, and Anthropic still recommends Opus 5.5 as the default for most coding and knowledge work.

Why it matters: Don't switch your default model on launch-day charts. If you have hard, long-running work, try Fable 5.2 at High rather than maximum effort, compare cost per finished task against Opus 5.5, and expect cyber or biology requests to be answered by an older fallback model.

Sources: Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2

Highlights

  1. Meta's Muse Ember ties Claude Haiku 5.5 on the Artificial Analysis index for about a third of the cost

    Meta released Muse Ember, its smallest model, aimed at the repetitive work agents do: tool calls, coding subtasks, routing and long-document extraction. It costs $0.10 per million input tokens and $0.50 per million output tokens, with a 1M-token context and up to 512K output tokens. On Artificial Analysis's Intelligence Index both Ember and Haiku 5.5 score 36 at maximum effort, but Meta's chart puts Ember at about $0.14 per task against about $0.41 for Haiku. Ember leads Meta's table on most agent and computer-use tests, yet trails Haiku on FrontierCode 2.0 (27.2 vs 34.9). Meta notes its own cost points are estimates at proposed pricing, and several scores are internal evaluations.

    modelspricingSource: Meta: Introducing Our Smallest, Most Efficient Model: Muse Ember

  2. Claude Code 2.1.321 fixes an MCP token leak and several permission gaps

    The 9 November release makes Fable 5.2 the default Fable model and fixes some security-relevant bugs: MCP sign-in requests could send secrets such as refresh tokens to another host when a server redirected them; a read-only Bash command could run outside the sandbox without a prompt; Bash permission checks could judge a file write from the wrong folder; and deny rules did not apply to files written by a cp or mv into a directory. The same release also changes 'mods' so their own tool calls raise no permission prompt (deny rules and organisation policy still apply). If you use MCP servers with OAuth, update.

    harnessessecuritySource: GitHub: anthropics/claude-code release v2.1.321

  3. Fireworks AI: October breach reached account and billing data, and for some customers datasets and partial model weights

    In a 9 November update to its 9 October bulletin, Fireworks AI says investigators (with Palo Alto Networks Unit 42) found the attackers also reached account and billing information, and for a limited set of accounts secrets, datasets and incomplete model weights. A third-party cloud provider hosting Fireworks R&D machines was breached on 23 October and used the next day to reach Fireworks development machines. Fireworks links the group to 'LLMJacking', reselling stolen model access. It says card numbers were not stored, it found no impact on zero-retention inference or training, and it has paused self-serve dataset uploads and downloads. The investigation is ongoing.

    industrysecuritySource: Fireworks AI: Update on October 2026 security incident

  4. Nvidia's board opens an independent investigation after reports that Jensen Huang was targeted by foreign intelligence

    In an 8 November statement, Nvidia's board said a committee of independent directors, helped by outside counsel, will investigate following recent reporting that CEO Jensen Huang was the target of a foreign intelligence effort. The board says it has no reason to believe national security or Nvidia secrets were compromised, that Huang supports the review, and that it has confidence in him. The statement does not name the country or the reporting, and gives no timeline.

    industrygovernanceSource: NVIDIA Newsroom: Statement From NVIDIA Board of Directors

Concept of the day

Safety classifier with fallback

Instead of refusing a risky request outright, a system runs a classifier on each prompt and, when it flags one, quietly hands the request to an older or weaker model that answers instead. Anthropic's Fable 5.2 does this: blocked cybersecurity queries go to Opus 4.8 and blocked biology queries to Opus 5. Users still get an answer, but it may be worse, and the fallback model's own weaknesses become part of the system you are relying on.

Source: Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2

Models released

ModelMakerSizeNotes

Also today

  1. OpenAI's Control Room can stop any research inference job within 35 seconds

    OpenAI describes Control Room, a central service that every research inference engine must check with before answering a request. Each agentic research job registers for an ID tied to an approved safety case; when a person or monitor stops the job, the next request is refused. The stop takes effect within 35 seconds (a caching delay OpenAI wants to remove), and the engines fail closed, refusing all requests if Control Room goes down. Fail-closed is live for frontier research workloads and still rolling out elsewhere.

    safetyinfrastructureSource: OpenAI: Control Room

Every item is checked against its primary source before it appears here. Figures we could not confirm are left out.

Related service: AI automation services for logistics

All daily issues

Five Agent Tricks, latest: Issue #5 (16 Nov 2026) · Subscribe free

From data points to pathways.

Book a 1:1 intro call and see how a small AI system could work on the data your business already has.

Book a call