Claude Fable 5.2 is out: what changed, what it costs, and when to use it

What Anthropic released

Two names, one model. Claude Fable 5.2 is the version everyone can use, with safeguards that block most cybersecurity and biology misuse. Claude Mythos 5.2 is the same model without those blocks, offered only to organizations accepted into the Cyber and Life Sciences Verification Programs. Both are described in the launch post and the system card.

• →Available today on the Claude apps, Claude Code, the Claude Platform (model id claude-fable-5-2), Amazon Web Services, Google Cloud and Microsoft Azure.
• →Up to 30% faster output than Fable 5.1, at a lower price.
• →An optional fast mode that generates up to 2.5 times faster, at double the price.
• →Paid subscribers on eligible plans get one to three extra Fable usage resets this week, depending on how long they have subscribed.

Claude Fable 5.2 price

Per token it is still one of the most expensive models on the market: four times the median for comparable models. The number that matters is cost per finished task, and that is where 5.2 changes the math.

Claude Fable 5.2 benchmarks

Anthropic's launch table compares Fable 5.2 with its predecessor and the strongest released models from Anthropic, OpenAI and Google. On a phone, the table scrolls sideways.

Three rows matter most for builders. FrontierCode 2.0 asks whether open-source maintainers would merge the change; 49.1% is the highest score of any released model Anthropic has tested. The Remote Labor Index uses real paid freelance jobs, and Fable 5.2 almost doubles Opus 5.5 there. OSWorld 2.1 measures using a computer like a person, at 89.3%. The one row it does not win is AutomationBench, where Gemini 4 Carbon leads with 53.9%.

Cost per task

The cost-per-task claims are the real news. At High effort, Fable 5.2 scores about the same as Opus 5.5 at its best (45.3% vs 45.5% on FrontierCode) for roughly half the cost per task. On Humanity's Last Exam, Fable 5.2 at Low effort matches Opus 5.5's top score for under one tenth of the cost. A more expensive model per token can be the cheaper one per job, because it gets there in fewer steps.

The independent check

Artificial Analysis scores Fable 5.2 (max effort) at 65 on its Intelligence Index, first of the 74 models it lists. The same page flags the downside: at max effort it is slow (about 45 minutes per test task against a 12.9 minute median) and very wordy. So run it at Medium or High for normal work and keep Max for jobs that are worth the wait.

What it did before launch

Anthropic's examples are the clearest sign of where long agent jobs are going:

• →Starting from an empty folder, it coordinated over 1,000 subagents for eight days and produced a working operating system that boots on a laptop, with drivers, file storage, windows, a login screen and an installer.
• →A team of agents running an early version moved a production service from Python to Rust; it now starts about 26 times faster.
• →A team of 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in 30 hours, with minimal human help.
• →It designed three working electronic devices from scratch, with circuit boards, 3D-printed cases and software.
• →Early testers report the same pattern: Uber ran a 10 to 12 hour security audit with about 170 reviewer agents, and Cognition noted it says it does not know instead of making things up.

Claude Fable 5.2 vs Opus 5.5: when to use which

Anthropic's own advice: start with Opus 5.5 for most coding and knowledge work, move up to Fable 5.2 when the job is large or open-ended, and move down to Sonnet 5.5 or Haiku 5.5 for fast, simple work. In practice:

• →Use Fable 5.2 for migrations, large refactors, bugs other models keep failing on, long research jobs, and anything you would leave running unattended for hours.
• →Keep Opus 5.5 for everyday building, edits, writing and chat. It is cheaper per token and fast enough.
• →Use Fable 5.2 as the advisor: Claude Code can have Haiku, Sonnet or Opus consult Fable only on the hard steps, so you pay Fable prices only when it counts.
• →Use Fable 5.2 as the orchestrator: it plans and splits the work, and sends the simple pieces to smaller, cheaper models.

How to try Claude Fable 5.2

• →Claude apps: pick Fable 5.2 in the model menu. Default effort in the apps and in Claude Code is Medium, which already uses fewer tokens than Fable 5.1.
• →Claude Code: switch the model to Fable 5.2 for a session, or set it as the advisor so your usual model calls it only when stuck.
• →API: model id claude-fable-5-2 on the Claude Platform, Amazon Bedrock, Google Vertex AI or Microsoft Azure.

A minimal call with the Anthropic Python SDK. Effort is set inside output_config; start at medium and raise it only for jobs that need it.

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY from your environment

response = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=16000,
    output_config={"effort": "medium"},
    messages=[{"role": "user", "content": "Plan the migration of this service from Python to Rust."}],
)

if response.stop_reason == "refusal":
    print("Declined:", response.stop_details.category)
else:
    for block in response.content:
        if block.type == "text":
            print(block.text)

What it means for you

A stronger model does not fix the usual reason AI builds go wrong. Your AI can keep notes and memory files, but nothing makes it read them before it acts, so it repeats what you already ruled out and redoes work that was finished. Fable 5.2 can coordinate agents for eight days; it still needs to check what you decided before every move.

That is why we built Journals DNA: your AI has to stop and check its memory before it acts, then it builds on what you already decided, and any AI or teammate can pick up the same file and continue. This post was made with Motion DNA, and everything else we build is at vibedna.ai.