Claude Fable 5.2: faster and cheaper, but Opus 5.5 is still the right pick for most work

On 9 November Anthropic released Claude Fable 5.2, together with Claude Mythos 5.2. Fable is the most capable model anyone can use. Mythos is the same model without some of the safeguards on cyber and biology tasks, and only organizations that passed one of Anthropic's verification programs get it.

What's different from 5.1? In short: faster, cheaper and blocked by mistake less often. Anthropic itself adds that for most work you should start with Opus 5.5. And if you switch through the API, there are four things to check in your code.

What's better than 5.1 #

• Faster: according to Anthropic, 5.2 generates output up to 30% faster than 5.1.
• Cheaper: $8 per million input tokens and $40 per million output tokens, down from $10 and $50. Cache reads cost $0.20 per million tokens, and the Batch API halves the price ($4/$20).
• Fast mode: new for Fable. Up to 2.5 times faster at double the rate ($16/$80). Available in the Claude API, Claude Code, Google Cloud and Microsoft Foundry.
• More efficient: at medium effort, 5.2 needs fewer tokens than 5.1 on hard tasks. Anthropic doesn't give a percentage.
• Fewer blocks: in early testing across claude.ai, Claude Code and Cowork, the safety classifiers blocked about 55% fewer requests than with 5.1: roughly 60% fewer on cybersecurity and 40% fewer on biology.

That last point may matter most to developers: less chance that an ordinary security question ends up with a different model. Blocked questions still get rerouted: cyber questions to Opus 4.8, biology questions to Opus 5.

The benchmarks, with a caveat #

Anthropic publishes a big table. A few numbers that matter to developers:

Benchmark	Fable 5.2	Opus 5.5	Fable 5.1
FrontierCode 2.0 (agentic coding)	49.1% (xhigh)	45.5%	32.4%
FrontierSWE v2 (agentic coding)	69.9%	62.3%	56.3%
Terminal-Bench-Science 0.1	74.6%	58.7%	52.6%
OSWorld 2.1 (computer use)	89.3%	87.2%	83.5%

I find another result more interesting than the top score: on FrontierCode 2.0, Fable 5.2 at high effort scores 45.3%. That's about the same as Opus 5.5 at its best setting, but according to Anthropic at roughly half the cost per task. Per token Fable costs twice as much; it just needs fewer tokens and fewer attempts on those tasks.

The most striking example from the announcement: a team of 1,000 Fable 5.2 agents fully formalized Fermat's Last Theorem in 30 hours. With the same setup, Opus 5.5 got to 26% after 18 hours and barely moved after that.

The caveat: all of these numbers come from Anthropic. Independent measurements are still to come.

Fable or Opus 5.5? #

Anthropic is remarkably candid here. Its own announcement says Fable is mainly meant for "the most ambitious, open-ended technical work". For most coding and knowledge work, it recommends starting with Opus 5.5. The developer docs put it even more concretely: move to Fable 5.2 when your evals on Opus at higher effort still fall short.

The prices side by side:

Model	Input / output per million tokens
Fable 5.2	$8 / $40
Opus 5.5	$4 / $20
Sonnet 5.5	$2 / $10

My rule of thumb after reading the docs: Opus 5.5 for daily work, Fable 5.2 for long, hard agent tasks where Opus gets stuck. Then look at the cost per finished task, not per token.

Effort: medium in the app, high in the API #

This detail is easy to get wrong. In the Claude apps and Claude Code, Fable 5.2 runs at medium effort by default. In the API the default is high. So an API request without effort runs one level higher than what you're used to in Claude Code. On Opus 5.5 the API default is medium.

What Anthropic recommends per level:

• low: quick questions, well-scoped tasks and high-volume subagents.
• medium: interactive coding, chat and everyday knowledge work.
• high: hard code, research and long agent tasks. According to Anthropic, most of the gain is here.
• xhigh: only for the hardest work, and only if you've measured that it helps. It uses many more tokens.
• max: to measure the model's ceiling, or for the rare task where time and tokens don't matter.

Two behavior changes to watch. A line in your prompt like "answer directly without deliberating" no longer works at medium and above: the model thinks anyway. If you want less thinking, lower the effort. And at low and medium, the model sometimes stops long tasks before the work is done, ending with a list of what's left. Anthropic's prompting guide includes a standard text to counter that.

Here's how to set effort explicitly with the PHP SDK:

use Anthropic\Client;

$client = new Client();

$message = $client->messages->create(
    model: 'claude-fable-5-2',
    maxTokens: 16000,
    outputConfig: ['effort' => 'high'],
    messages: [
        ['role' => 'user', 'content' => 'Review this migration for data loss.'],
    ],
);

What breaks when you switch through the API #

If you use Claude Code or claude.ai, you don't need to do anything. If you build API requests yourself, the docs list four changes from Fable 5.1:

1. The prefix check now applies to every account. A Fable 5.2 thinking block is only valid as long as everything before it stays the same: the system prompt, the tools and earlier messages. Change something in the history afterwards and you get a 400 with The block is bound to a different conversation. On 5.1 this only applied to accounts created on or after 31 August 2026, so code that works on an older account today can fail after the switch. The fix is to keep the history append-only: only add, never edit.
2. Only Fable 5.2 and Mythos 5.2 can read 5.2's thinking blocks. If a router or fallback switches to Opus 5.5 or Fable 5.1 mid-conversation, that model runs without the earlier reasoning. The request still succeeds, but the first turn after the switch costs more.
3. The old computer use tool computer_20251124 returns a 400, except on Amazon Bedrock. Use computer_toolset_20260801.
4. The advisor tool: a Fable 5.2 executor only accepts Fable 5.2 or Mythos 5.2 as its advisor.

Coming from Opus 5.5, there are two more things. Fable requires 30-day data retention and isn't available under zero data retention unless Anthropic explicitly allows it. And in organizations with customer-managed encryption keys (CMEK), structured outputs don't work; there you have to validate tool input yourself.

What I would do #

• Using Claude Code? Try Fable 5.2 on a task where Opus 5.5 gets stuck. Pro and Team Standard users can use Fable for up to half of their weekly limit. Start at medium, the default, and go to high when the work is long and hard.
• Running Fable 5.1 through the API? Swap the model ID, but first test whether your code edits the history. Also run an effort sweep, because settings you tuned for 5.1 don't carry over to 5.2 automatically.
• Running Opus 5.5? Leave it, unless your evals show you need more. If you do switch, check data retention and rate limits first: Fable 5.2 has its own limit.

Sources #

• Anthropic: Claude Fable 5.2 and Mythos 5.2
• Anthropic: Claude Fable product page with benchmarks and pricing
• Claude docs: What's new in Claude Fable 5.2
• Claude docs: Prompting Claude Fable 5.2
• MacRumors: Anthropic Launches Claude Fable 5.2 and Mythos 5.2
• Claude on X: FrontierCode results and default effort