Claude Fable 5.2 and Mythos 5.2: Price, Benchmarks, Plan Access and How to Use Them (2026)

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, its new top models. They landed on 9 November 2026, and they are cheaper than the 5.1 versions they replace.

The headline claim is big: a team of 1,000 Fable 5.2 agents formalised Fermat’s Last Theorem in Lean in 30 hours. But most readers want simpler answers: what it costs, whether you get it on your plan, and when to use it instead of Opus 5.5.

This guide covers all of that, with the full benchmark table, API setup, plan access, safeguards and the limits Anthropic mentions in the small print.

What Anthropic Launched

Anthropic calls Fable 5.2 and Mythos 5.2 its models “for the most ambitious, open-ended technical work”. They succeed Fable 5.1 and Mythos 5.1, and Anthropic says output is up to 30% faster than the 5.1 pair.

The Fable line started with Fable 5 and Mythos 5 on 9 June 2026 at $10 / $50 per million tokens. The 5.2 release keeps the same idea but drops the price by 20%.

Anthropic also gave Fable 5.2 two named roles in multi-model setups:

• Advisor: Haiku, Sonnet or Opus consult Fable 5.2 on hard steps.
• Orchestrator: Fable 5.2 plans the job and hands simpler sub-tasks to smaller, cheaper models.

Fable 5.2 vs Mythos 5.2: Same Model, Different Rules

This confuses a lot of people, so here it is plainly. Anthropic says the two are “the same underlying model”. The only difference is how many safety blocks sit on top.

Note: For normal coding, writing, research and data work you will never notice the difference. The blocks only touch high-risk cyber and biology requests.

Specs and Pricing

Here is everything you need for budgeting. Prices are Anthropic first-party API rates per million tokens; Bedrock and Vertex set their own prices.

What a job actually costs

A run that reads 200,000 tokens and writes 20,000 costs about $1.60 + $0.80 = $2.40 at standard rates. The same job on Opus 5.5 ($4 / $20) is about $1.20.

That is why effort matters so much. Anthropic says Fable 5.2 at low effort matched Opus 5.5’s best Humanity’s Last Exam score at under a tenth of the cost per task. So a cheap Fable setting can beat an expensive Opus one.

Tip: Long system prompts and repeated documents should use prompt caching. Reads cost $0.20 instead of $8. My prompt caching guide shows the setup.

Benchmarks: Fable 5.2 vs Opus 5.5, GPT-6 Astra and Gemini 4 Carbon

All numbers below come from Anthropic’s launch page, so treat them as vendor-reported. Fable 5.2, Fable 5.1 and Opus 5.5 ran with safeguards on unless marked.

Two rows are worth a closer look. Gemini 4 Carbon still wins AutomationBench, and on the AA-Briefcase Elo score Opus 5.5 (1800) edges Fable 5.2 (1785). So Fable is not first on every chart.

Cyber and biology scores (Mythos 5.2)

• CyScenarioBench: 76.3% vs 61.7% for Mythos 5.1 and 61% for GPT-6 Cyber.
• CVE Rediscovery (1,829 CVEs): 89.6%, finding 10.2% more vulnerabilities than Opus 5.5.
• De novo protein binder design (24 h): 85.1% vs 79.1% for Mythos 5.1.

Independent tracker DataLearner also lists Anthropic’s system-card scores of 93.0% on SWE-bench Pro and 67.6% on Terminal-Bench 4.0. It notes that xhigh effort beat max on FrontierCode (49.1% at about $14 per task vs 45.7% at about $45), partly because 22 of 660 max attempts timed out.

For the competing models, see my Gemini 4 Carbon benchmarks post and the OpenAI API pricing guide for GPT-6 Astra costs.

The Fermat’s Last Theorem Result

This is the part that went viral. Anthropic says a team of 1,000 Fable 5.2 agents fully formalised Fermat’s Last Theorem in Lean in 30 hours, with minimal human help.

The agents started from short prompts, built their own way of coordinating, and used no maths references. Progress was scored only against a list of intermediate theorem statements.

Note: Lean is a proof checker, so the proof itself is machine-verified. The caveat is that this was a single run per model, reported by Anthropic.

Other showcase results

• Operating system: one engineer had Fable 5.2 coordinate over 1,000 subagents for eight days to build a prototype OS that boots on a laptop.
• Python to Rust: agents migrated a production service that now starts about 26x faster.
• Hardware: from a one-paragraph prompt, it designed three working devices (voice hub, USB speakerphone, handheld virtual pet) with four-layer boards, working with partner Diode.
• Physics: Anthropic says it solved two open black-hole problems tied to holographic duality (preprints posted).

How to Use Fable 5.2 (App, Claude Code and API)

In the Claude app

1. Open claude.ai or the Claude app on a paid plan.
2. Pick Claude Fable 5.2 in the model menu under the chat box.
3. Watch your usage. On Pro and standard Team seats, Fable can use up to 50% of your weekly limit. Eligible paid plans also get one to three Fable resets, based on how long you have subscribed.

In Claude Code

1. Update Claude Code to the latest version.
2. Run /model and choose Fable 5.2, or start with claude --model claude-fable-5-2.
3. Set effort to match the job. Claude Code and the app default to medium; raise it only for hard tasks.

Plan prices are in my Claude Code pricing guide.

Through the API (Python)

pip install anthropic
export ANTHROPIC_API_KEY="sk-ant-..."

import anthropic

client = anthropic.Anthropic()

with client.messages.stream(
    model="claude-fable-5-2",
    max_tokens=64000,
    output_config={"effort": "high"},
    messages=[{"role": "user", "content": "Review this repo plan and list risks."}],
) as stream:
    msg = stream.get_final_message()

if msg.stop_reason == "refusal":
    print("Blocked by a safeguard:", msg.stop_details)
else:
    print(msg.content[-1].text)

Notice what is missing: no thinking setting and no temperature. Thinking is always on, and the old sampling knobs are rejected. Effort is the control you use.

New to the SDK? Start with my Claude API Python tutorial, and see the reasoning effort guide for how each level changes cost.

Turning on fast mode

Add speed="fast" to a normal Messages request. It costs double ($16 / $80) for up to 2.5x faster output, and works in Claude Code, the Claude Platform, Claude Models on Google Cloud and Microsoft Foundry.

Fable 5.2 vs Opus 5.5 vs Sonnet 5.5: Which to Pick

• Pick Fable 5.2 when a task runs for hours, spans many files, or failed on Opus.
• Stay on Opus 5.5 for most daily work. It costs half as much and is close on many charts.
• Mix them: use Fable as the orchestrator or advisor and let Haiku 5.5 or Sonnet do the grunt work.

If you are weighing Claude against cheaper Chinese models, my DeepSeek vs Claude comparison covers price per task.

Safeguards, Refusals and Fallbacks

Fable 5.2 runs safety classifiers that block high-risk cyber work and most research biology. Anthropic says these blocks fire about 55% less often than on Fable 5.1 (about 60% fewer in cyber and 40% fewer in biology).

• Blocked cyber request: answered by Opus 4.8 instead.
• Blocked biology request: answered by Opus 5 instead.
• Distillation protection: Anthropic says Fable 5.2 cuts the data copycat labs can harvest by 90-99%.

Safety test results (vendor-reported)

Warning: Anthropic also says both models voiced suspicion that they were being tested more often than earlier models. That makes safety evaluations harder to read, and outside groups such as the UK AI Security Institute and METR took part in testing.

How to Get Mythos 5.2 Access

1. Cybersecurity teams: apply to Anthropic’s Cyber Verification Program.
2. Biomedical researchers: apply to the Life Sciences Verification Program.
3. Claude Enterprise customers: you can use Claude Security, now powered by Mythos 5.2, without joining the cyber program.

Mozilla says Mythos 5.2 roughly doubled Mythos 5.1’s vulnerability-discovery numbers, and one security firm found one in six of its confirmed bugs were missed by every other frontier model it tried.

Limits and What We Still Do Not Know

• No zero data retention by default. Fable 5.2 needs 30-day retention unless Anthropic authorises otherwise.
• Some AI R&D skills were deliberately limited. Kernel optimisation and LLM architecture research now perform around Opus 5.5 level.
• Long turns. Hard requests can run many minutes, so plan for streaming and timeouts.
• Usage caps on plans. Anthropic gave the 50% weekly share but no exact message counts.
• Health questions. DataLearner lists HealthBench at 58.8% for Fable 5.2 vs 64.3% for Opus 5.5.
• Unknown: independent leaderboard results are still arriving, and Bedrock/Vertex pricing is set by those clouds.

Troubleshooting Common API Errors

My Take

The price cut is the real story for most of us. Fable 5.2 is 20% cheaper than 5.1 and, on Anthropic’s numbers, a big step up, especially on long agent tasks like Remote Labor Index (41.3% vs 17.9%).

Still, I would not switch everything over. Anthropic itself tells you to start on Opus 5.5, which costs half as much. My plan would be Opus 5.5 by default, Fable 5.2 at high effort for the jobs Opus fails, and Fable as an orchestrator for big multi-agent runs.

The Fermat result is impressive because Lean checks every step. But it is one run, so I will wait for independent agent benchmarks before calling Fable 5.2 the best model everywhere. Gemini 4 Carbon still leads AutomationBench, after all.

FAQ

When was Claude Fable 5.2 released?

Anthropic released Fable 5.2 and Mythos 5.2 on 9 November 2026.

How much does Claude Fable 5.2 cost?

$8 per million input tokens and $40 per million output tokens on the API. Fast mode costs $16 / $80, and cache reads are $0.20.

What is the difference between Fable 5.2 and Mythos 5.2?

They are the same model. Mythos 5.2 has fewer cyber and biology safeguards and is only for organisations approved through Anthropic’s verification programs.

Is Fable 5.2 available on Claude Pro?

Yes. Pro and standard Team seats can use Fable for up to 50% of their weekly usage, and eligible paid plans get one to three Fable resets.

What is the context window of Fable 5.2?

1 million tokens, with up to 128K output tokens per response.

Should I use Fable 5.2 or Opus 5.5 for coding?

Opus 5.5 for everyday coding, as Anthropic recommends. Move to Fable 5.2 for long, multi-file or multi-agent work, or when Opus gets stuck.

Did Claude really prove Fermat’s Last Theorem?

The theorem was proved by Andrew Wiles in the 1990s. Fable 5.2 agents formalised that proof in the Lean proof checker in 30 hours, according to Anthropic.

Why does Fable 5.2 refuse some security questions?

Its classifiers block high-risk cyber tasks like exploit writing. Blocked cyber queries are answered by Opus 4.8 instead, and verified security teams can apply for Mythos 5.2.

Sources

• Anthropic: “Claude Fable 5.2 and Mythos 5.2” launch page (9 Nov 2026), anthropic.com/claude-fable-and-mythos-5-2: benchmarks, pricing, Fermat, safeguards, availability
• Anthropic: “Claude Fable 5 and Mythos 5” (9 Jun 2026): original Fable/Mythos launch and $10 / $50 pricing
• Anthropic Claude API documentation (model table, Fable 5.2 API behaviour, fast mode): context window, max output, cache prices, model IDs
• DataLearner model page for Claude Fable 5.2 (updated 10 Nov 2026): batch and cache pricing, system-card scores, FrontierCode cost per task, HealthBench
• AI Weekly, “AI News Today, November 9” (aiweekly.co): launch coverage