Claude Fable 5.2: What Changed and Who Should Switch

Claude Fable 5.2 is Anthropic's newest and most capable public model, released on 9 November 2026. It costs less than Fable 5.1, produces output faster, and is now the default Fable model in Claude Code. Anthropic's own advice is still to start with Claude Opus 5.5 for most work and reach for Fable 5.2 only when a task is hard enough to need it.

This article explains what changed, what it means for a team building software with AI coding agents, and where the claims need a pinch of salt.

What is Claude Fable 5.2?

Anthropic's current Claude lineup has four model families. From the top: Fable, Opus, Sonnet and Haiku. Fable is built for demanding reasoning and long agentic work, such as agent sessions that run for hours. "Agentic" here means the model works through many steps on its own instead of answering one question.

On 9 November 2026 Anthropic announced Fable 5.2 and Mythos 5.2. They are the same underlying model. Fable is open to everyone and carries safeguards that block high-risk cyber and biology requests. Mythos is limited to organisations that Anthropic has verified through its verification programs, such as the Cyber Verification Program.

The model documentation gives the basics:

• It can hold 1 million tokens of context. A token is a small chunk of text, and 1 million tokens is roughly 555,000 words.
• It can write up to 128,000 tokens in one answer.
• Its "thinking", the reasoning it does before replying, is always on. You steer how much it thinks with an effort setting.
• Anthropic says it will not retire the model before 9 November 2027.

What changed from Fable 5.1?

Four things matter most.

It is cheaper. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. Fable 5.1 cost $10 and $50. Reading back text the model has already seen (a cache read) costs $0.20 per million tokens. For comparison, the documentation lists Opus 5.5 at $4 and $20, half the price of Fable 5.2.

It is faster. Anthropic says output is up to 30% faster than Fable 5.1. An optional fast mode, at $16 per million input tokens and $80 per million output tokens, speeds up generation by up to 2.5 times.

It scores higher on coding tests. A benchmark is a standard set of tasks used to compare models. Anthropic reports the following:

• On FrontierCode 2.0 at its extra-high effort setting, Fable 5.2 scored 49.1%, against 45.5% for Opus 5.5 and 32.4% for Fable 5.1.
• On FrontierSWE v2 it scored 69.9%, against 62.3% for Opus 5.5 and 56.3% for Fable 5.1.
• On FrontierCode 2.0 at the lower "high" effort it scored 45.3%, about level with Opus 5.5, at roughly half the cost per task.

It blocks fewer requests. In early testing, Anthropic found about 55% fewer safeguard blocks than with Fable 5.1: about 60% fewer in cybersecurity and 40% fewer in biology.

How do I get it in Claude Code?

Claude Code version 2.1.321, released on 9 November 2026, added Fable 5.2 and made it the default Fable model. Saved Fable 5.1 picks move to the new model: immediately on Pro and Max plans, and on Team and Enterprise plans once your organisation allows it. The changelog says you can go back with /model claude-fable-5-1. According to the announcement, medium is the default effort in Claude Code, and fast mode is available there.

On subscriptions, Anthropic says plans that include Fable get one to three Fable "resets", depending on how long you have subscribed. It also says Fable is included for up to 50% of weekly usage limits on Pro plans and standard Team seats. The announcement describes extra Fable usage for eligible plans "this week", so expect these terms to change.

If you pay by the token through the API (the way programs talk to the model directly), the model ID is claude-fable-5-2. If you only use Claude Code on a plan, you pick it as a model in the same way you would pick Opus.

Who should switch, and who should not?

Anthropic's documentation says to start with the latest Opus model for most workloads and move to Fable 5.2 "when your evals on a Claude Opus model at higher effort still don't meet your requirements." An eval is your own test of whether the model does your job well enough. In plain words: if Opus already does the job, a more expensive model buys you little.

Fable 5.2 is worth a trial in cases like these:

• Big, ambiguous tasks. A change across many files, or a feature where the request is loose and the agent must make many judgement calls.
• Long runs. Work that needs hours without a person stepping in.
• Hard bugs. Problems where Opus has already failed at high effort.

It is probably wasted on small edits, renames, routine fixes, formatting, and anything repetitive that a cheaper model already handles well. That is our judgement, not Anthropic's list, but it follows from the price gap.

Price per token is only part of cost. The documentation says that at high effort and above the model thinks before almost every reply. It also says that in coding tools whose prompts invite delegation, the model hands work to helper agents readily, and that these helpers can account for a large part of what a session costs. A lower per-token price does not guarantee a lower bill.

What are the catches?

The numbers come from Anthropic. The benchmark scores above are the ones Anthropic published. Treat them as a reason to run your own trial, not as a promise.

It refuses some security work. Fable 5.2 blocks high-risk cyber activity, including penetration testing and exploit generation. Blocked cyber requests are answered by Opus 4.8, an older model. If your team uses an agent for security checks, part of that work may be handled by the older model. The full-strength version, Mythos 5.2, is only for verified organisations.

It does not always do what a short instruction says. The documentation lists several behaviour differences:

• A prompt line such as "answer directly without deliberating" did not stop Fable 5.2 from thinking at medium effort and above in Anthropic's testing. The fix is a lower effort setting for easy tasks.
• On long, multi-step tasks it sometimes ends a turn with a list of what remains, an offer to continue, or a stop at an ordinary obstacle. This happens more at low and medium effort.
• If a prompt does not say who owns the code, it tends to add to the existing structure and report nearby problems rather than fix them.
• Long runs of tool calls can pass with little explanation to the user.

It suspects it is being tested. Anthropic says both Fable 5.2 and Mythos 5.2 voiced suspicion that they were being evaluated more often than any recent Anthropic model. Anthropic says this complicates judging how the models behave in real use.

Data retention. Fable 5.2 requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorises it. If your company has strict rules about where code may be kept, check this before you switch.

What should a team check this week?

If you build on the API yourself, the what's new page lists four changes from Fable 5.1 that can break code:

• A check on edited conversation history, which makes the request fail if you changed earlier turns, now applies to every account.
• Only Fable 5.2 and Mythos 5.2 can read Fable 5.2's thinking blocks, so switching to another model mid-conversation loses that reasoning.
• The older computer-use tool returns an error on every platform except Amazon Bedrock.
• If you use the advisor tool (in beta), a Fable 5.2 executor needs a Fable 5.2 or Mythos 5.2 advisor.

Separately, forced tool use, where your code makes the model call a specific tool, is not supported and returns an error. That is not new compared with Fable 5.1 or Opus 5.5.

Anthropic says nothing changes for the history check if Claude Code, claude.ai or the Claude Agent SDK builds your requests. For a team that only uses Claude Code, the practical steps are simpler:

1. Pick one hard task you have already done with Opus 5.5.
2. Run it again on Fable 5.2, and compare the result, the time and the share of your plan used.
3. Only then decide whether any role or project should default to it.

How does Atlas handle a new model like this?

Atlas is a team of AI engineers built on Claude Code, and it needs your own Claude Pro or Max plan. Each engineer runs as its own Claude Code session with a model and effort suited to its role. In Settings, each kind of task starts on the model and effort Atlas chose for it. You can change any of them, and "Use Atlas's" puts it back. Atlas also tries cheaper settings and keeps them only when cost per finished task and the amount of work sent back afterwards show they are better.

What Atlas does not do today is run a head-to-head trial when a new top model ships. That is coming to Atlas, built in, so it needs no setup on your side. Until then, the manual trial above is the safest way to decide.

Conclusion

Fable 5.2 is a real step up at a lower price than its predecessor, but by Anthropic's own advice it is a specialist, not a new default. Keep Opus 5.5 for everyday work, try Fable 5.2 on your hardest task, and judge it by finished work and plan usage, not by launch benchmarks. Read the behaviour notes too: tell it clearly what to finish, and keep easy tasks on a lower effort.