Claude Fable 5.2 and Mythos 5.2: Price, Benchmarks, and When to Switch

Claude Fable 5.2 is Anthropic's new top generally available model, released November 9, 2026, alongside Claude Mythos 5.2. It is 20% cheaper than Fable 5.1 at $8/$40 per million tokens and leads Anthropic's benchmark table over Opus 5.5, GPT-6 Astra and Gemini 4 Carbon on nine of eleven rows.

Last verified: November 10, 2026, against Anthropic's launch post, API pricing page and model selection guide.

What did Anthropic announce on November 9?

Anthropic released two versions of one model. In its launch post the company writes that "Claude Fable and Claude Mythos are the same underlying model." The difference is the safeguards wrapped around it:

• Claude Fable 5.2 is for everyone. Safety classifiers block high-risk cybersecurity work (penetration testing, exploit generation, binary vulnerability research) and most research biology. A blocked query is answered by Opus 4.8 for cyber and Opus 5 for biology.
• Claude Mythos 5.2 removes some of those safeguards and is limited to organizations accepted into the Cyber Verification Program or the Life Sciences Verification Program. Claude Enterprise customers get it through Claude Security without joining either program.

Both models generate output up to 30% faster than the 5.1 versions. Anthropic also says the new safeguards produce about 55% fewer unnecessary blocks than Fable 5.1, roughly 60% fewer in cybersecurity and 40% fewer in biology.

Claude Fable 5.2 price, context window and specs

The list price dropped from $10/$50 to $8/$40 per million input and output tokens, a 20% cut. Everything else on the spec sheet matches Opus 5.5 except the price and the default effort.

Sources: Claude API pricing, models overview and the claude.dev model guide, checked November 10, 2026.

The cache-read price is the real story

Fable 5.2 costs twice as much as Opus 5.5 per input and output token, but both models charge the same $0.20 per million for cache reads. Long agent runs re-read the same context thousands of times, so for them cache reads, not fresh input, are most of the bill. The more a workload leans on cache, the closer Fable 5.2 gets to Opus 5.5 on cost.

To show the size of the effect, we priced the token profile Felix Rieseberg posted on X for an eight-hour, one-shot game build: about 445 million input tokens, "heavily cached", and 2.7 million output tokens.

Our arithmetic from list prices. It leaves out cache writes, which depend on how often the context changes, so real bills land between the two rows.

Fully cached, Fable 5.2 costs about 1.4x Opus 5.5 on that profile, not 2x, and about 20% less than Fable 5.1. Uncached, the full 2x gap comes back. His own figure of about $550 includes cache writes we can't see, so treat the table as a comparison between models, not a reconstruction of his bill. The practical rule: if your agent keeps a large stable context (a repo map, a long spec, tool definitions), check your cache hit rate before ruling Fable 5.2 out on price.

Claude Fable 5.2 benchmarks vs Opus 5.5, GPT-6 Astra and Gemini 4 Carbon

Anthropic calls Fable 5.2 "a step above Opus 5.5 in almost all the capabilities we evaluated." Its own table backs that up, with two exceptions it doesn't highlight:

Anthropic's own figures from the launch post, highest available effort unless noted (FrontierCode Fable 5.2 at Xhigh, Remote Labor Index at High, Gemini 4 Carbon FrontierCode at Medium). Not independently reproduced. n/a means Anthropic published no score.

Where Fable 5.2 does not win

• Business workflows. Gemini 4 Carbon scores 53.9% on AutomationBench against Fable 5.2's 51.4%. If your agents mostly move data between SaaS tools, Fable 5.2 is not the clear pick.
• Knowledge work preference. On AA-Briefcase, Fable 5.2 leads on the rubric score (66.1%), but Opus 5.5 holds the higher Elo (1800 vs 1785). The Elo comes from pairwise comparisons of outputs, so on that measure the cheaper model is marginally ahead.

The biggest jumps are on long, messy work

The largest gains over Opus 5.5 are on the least scripted tasks: Remote Labor Index (41.3% vs 21.3%), RoboDojo robotics (64.9% vs 39.4%) and Terminal-Bench-Science (74.6% vs 58.7%). On tidier coding tests the lead is narrower: 3.6 points on FrontierCode 2.0 and 2.1 on OSWorld 2.1. In practice, Fable 5.2's premium buys the most on open-ended, multi-day work and the least on routine edits.

Read the effort footnote

Anthropic notes that "all scores show each model's highest-available effort unless otherwise noted." But Medium is "the default in Claude apps and Claude Code." So the 49.1% FrontierCode figure is an Xhigh result, and Claude Code doesn't run Fable 5.2 that way out of the box.

The efficiency curves are the more useful numbers. On FrontierCode 2.0, Fable 5.2 at High scores 45.3%, level with Opus 5.5 at Max (45.5%), for roughly half the cost per task. On Humanity's Last Exam, Fable 5.2 at Low matches Opus 5.5's top score for under a tenth of the cost. We saw the same effect when Opus 5.5 shipped: the right effort level often matters more than the model.

What Fable 5.2 built in testing

Anthropic's showcase is unusually concrete for a model launch. Each claim below is quoted from the launch post. None has been independently reproduced yet.

Three working electronic devices from one prompt

From a one-paragraph prompt, Fable 5.2 designed a home voice-assistant hub, a USB speakerphone and a handheld virtual pet. It wrote the specifications, picked parts from US distributors, routed four-layer circuit boards, modeled 3D-printed enclosures in CAD and wrote the firmware. Each board was ready for manufacturing in under a day. Diode built the boards, Claude guided its engineers through testing and assembly, and Anthropic says "all three devices worked without needing any electrical modifications." On the BenchCAD 2.1 hardware benchmark, Fable 5.2 scores 83.5 out of 100 against 65.3 for Opus 5.5.

Fermat's Last Theorem in 30 hours (a formalization, not a new proof)

Summaries on X say Fable 5.2 "proved" Fermat's Last Theorem. It didn't. About 1,000 Fable 5.2 agents translated the existing proof into Lean, a language a computer can check step by step, in 30 hours with minimal human intervention. Anthropic's first formalization, by an unreleased research model with a purpose-built harness, took 11 days. Under the same 1,000-agent setup, Opus 5.5 completed 26% in 18 hours and barely moved after that. Fable 5.2 also used a third of the action tokens and produced a proof seven times shorter.

An operating system, a Rust migration and two physics results

• Prototype OS. Starting from an empty folder, Fable 5.2 coordinated over 1,000 subagents for eight days and produced an operating system that boots on a laptop, with drivers, file storage, windows, a login screen and an installer.
• Python to Rust. Agents running an earlier version migrated a production service at Anthropic to Rust. It now starts about 26 times faster.
• Black holes. It solved two open problems in theoretical physics that rely on holographic duality, written up as preprints.
• Protein structure. Mythos 5.2 found a way to predict whole AAV capsid structures from open models and accurately modeled 28 of the 34 capsids it tried.

Customer quotes in the post are similarly concrete. Uber ran a 10 to 12 hour security audit with about 170 reviewer agents. Discovered Materials made Quantum ESPRESSO's phonon code about 36 times faster. Mozilla says Mythos 5.2 "roughly doubled" its vulnerability discovery metrics over Mythos 5.1 while hardening Firefox.

Should you switch from Opus 5.5 to Fable 5.2?

For most teams, not by default. Anthropic's models overview says to "start with Claude Opus 5.5 for most workloads" and use Fable 5.2 "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short." The claude.dev guide by Addy Osmani adds a rule worth taping to the monitor: "moving to a higher effort level won't give a model knowledge or skill it doesn't have."

Condensed from Which Claude model should you build with? on claude.dev plus our own reading of the benchmark table.

The pattern that wins on cost is mixing models, which Anthropic now describes in two named roles. As an Advisor, Fable 5.2 is consulted by Haiku, Sonnet or Opus on the hard calls. As an Orchestrator, it plans and delegates the simple work to cheaper models. Keep file search, test runs and log reading on Haiku 5.5, and keep code edits and judgment on the main model. We use the same split in our agentic coding process, and our Codex vs Claude Code comparison covers how each tool exposes model choice.

Who can use Claude Fable 5.2, and what are Fable resets?

Fable 5.2 is live on the Claude Platform as claude-fable-5-2 and on Amazon Web Services, Google Cloud and Microsoft Azure. Fast mode is available in Claude Code, the Claude Platform, Claude on Google Cloud and Microsoft Foundry. In the Claude apps, Anthropic's launch week promotion works like this:

• Pro and standard Team seats can use Fable for up to 50% of their weekly usage limits.
• Plans and seats that already include Fable get "one to three Fable resets based on how long they've subscribed." Boris Cherny, who leads Claude Code, put it simply on X: "the longer you've been subscribed, the more Fable resets you get."
• The tenure tiers reported by OfficeChai (one reset under six months, two for six to twelve months, three for over a year) are not in Anthropic's launch post. Check your own account before planning around them.

Claude Duet, Anthropic's proactive assistant app, trended on X the same day as early-access invites went out. It isn't a Fable 5.2 feature: Duet launched on October 21 for US Pro, Max and Ultra subscribers.

Safety: what the system card flags

Anthropic rates the risk of the model "behaving in unsafe ways in high-stakes situations to be low." On its automated behavioral audit, Mythos 5.2 scored 2.67 (lower is better), against 2.79 for Opus 5.5 and 3.28 for Mythos 5.1. In a sandbox-escape test across 1,800 scenarios, Mythos 5.2 used planted credentials or files in 3.1% of cases. The two earlier Mythos versions did so in 39% and 21%.

Two caveats come from Anthropic itself. Both models "voice suspicion that they are being evaluated more often than any of our recent models," which makes test results harder to trust. And Anthropic deliberately held back gains in AI research skills such as accelerator kernel optimization, where Fable 5.2 performs at about Opus 5.5's level. External testers included the UK AI Security Institute and METR.

How Fable 5.2 changes the frontier price map

Two months ago Fable 5.1 and GPT-6 Astra both cost $10/$50, and Opus 5.5 at $4/$20 was the value pick at the top. Now Anthropic's flagship lists 20% below Astra's last published price, and in Anthropic's own table it beats Astra on every benchmark where both appear. Google's answer is still the Gemini 4 family, which wins one row here. We will add Fable 5.2 to our maintained best LLM for coding leaderboard once independent boards publish scores. Vendor tables are a starting point, not a verdict.

FAQ

What is Claude Fable 5.2?

Claude Fable 5.2 is Anthropic's most capable generally available model, released November 9, 2026. It targets long, open-ended technical work such as large migrations, research and multi-agent orchestration. It costs $8 per million input tokens and $40 per million output tokens, with a 1M-token context window.

What is the difference between Claude Fable 5.2 and Mythos 5.2?

They are the same underlying model. Fable 5.2 carries safeguards that block high-risk cybersecurity and biology work and is open to everyone. Mythos 5.2 drops some of those safeguards and is limited to vetted organizations in Anthropic's Cyber and Life Sciences Verification Programs. Claude Enterprise customers can also reach Mythos 5.2 through Claude Security.

How much does Claude Fable 5.2 cost?

On the API, $8 per million input tokens and $40 per million output tokens, down 20% from Fable 5.1. Cache reads are $0.20 per million, and fast mode costs $16 input and $80 output. In the Claude apps it is included on paid plans, with Pro and standard Team seats getting up to 50% of weekly limits during the launch promotion.

Is Claude Fable 5.2 better than Opus 5.5?

On Anthropic's benchmarks, yes on almost every row, by margins from about 2 points (OSWorld 2.1) to 20 points (Remote Labor Index). It costs twice as much per token, though, and Anthropic still recommends starting on Opus 5.5 for most coding and knowledge work. At High effort Fable 5.2 matches Opus 5.5's best FrontierCode score for about half the cost per task, which is the case for switching on hard jobs.

Is Claude Fable 5.2 better than GPT-6 Astra?

In Anthropic's own table Fable 5.2 beats GPT-6 Astra on every benchmark where both appear, including FrontierCode 2.0 (49.1% vs 42.8%) and Humanity's Last Exam with tools (74.2% vs 57.2%). GPT-6 Astra listed at $10/$50 when we last checked its rate card, which would make Fable 5.2 20% cheaper. These are vendor numbers, so wait for independent boards before treating them as settled.

Did Claude Fable 5.2 prove Fermat's Last Theorem?

No. It formalized the existing proof in Lean, a language a computer can check step by step. About 1,000 Fable 5.2 agents finished the formalization in 30 hours with minimal human help. Anthropic's earlier research model took 11 days, and Opus 5.5 completed 26% in 18 hours under the same setup.

Is Claude Fable 5.2 available on the API and in Claude Code?

Yes. It launched on the Claude Platform as claude-fable-5-2 and on Amazon Web Services, Google Cloud and Microsoft Azure. It is available in Claude Code, where it runs on Medium effort by default. Fast mode is available in Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry.

What are Fable resets?

A launch-week bonus that resets your Fable usage. Anthropic says plans and seats that include Fable get one to three resets depending on how long you have been subscribed. Boris Cherny of Claude Code framed it as a reward for long-time subscribers. The exact tenure tiers come from secondary reporting, not from Anthropic's launch post.