Claude Fable 5.2 vs Opus 5.5: Is 2 the price worth it

Anthropic released Claude Fable 5.2 on November 9. It's the new top of the Claude lineup, aimed at the hardest coding and reasoning work, and it costs twice as much per token as Claude Opus 5.5.

So the useful question isn't "is it better?" Anthropic's own numbers say it usually is. The question is when it's worth 2× the price, and what changes in your code if you switch.

This post covers what's new, the real price difference, the API details that will bite you on migration, and a simple way to decide. Everything here comes from Anthropic's announcement and docs unless it says otherwise.

What's new in Fable 5.2

Anthropic positions Fable 5.2 for "demanding reasoning and long-horizon agentic work". Compared with Fable 5.1, Anthropic says it:

• scores higher across coding, reasoning and vision benchmarks,
• generates output up to 30% faster,
• refuses less: about 55% fewer safeguard blocks than Fable 5.1,
• can coordinate large numbers of subagents on long-running tasks.

There's also Claude Mythos 5.2, the same underlying model without the cybersecurity and biology safeguards. It's limited to organisations in Anthropic's verification programs, so for most developers, Fable 5.2 is the one you'll use.

The benchmarks (Anthropic's numbers)

These are from Anthropic's launch table, run at each model's highest effort with Fable's safeguards on. They're lab-reported, not independent, so treat them as a starting point for your own evals.

Benchmark	Fable 5.2	Fable 5.1	Opus 5.5
FrontierSWE v2 (agentic coding)	69.9%	56.3%	62.3%
FrontierCode 2.0 (agentic coding)	49.1%	32.4%	45.5%
Terminal-Bench-Science 0.1	74.6%	52.6%	58.7%
OSWorld 2.1 (computer use)	89.3%	83.5%	87.2%
Humanity's Last Exam (with tools)	74.2%	65.6%	67.7%
AA-Briefcase (Elo)	1785	1669	1800

Two things stand out. The jump from Fable 5.1 is large on long agentic tasks. And Opus 5.5 still wins one row, AA-Briefcase Elo, so "newer and pricier" doesn't mean better at everything.

What it costs

	Fable 5.2	Opus 5.5
Input (per million tokens)	$8	$4
Output (per million tokens)	$40	$20
Context window	1M tokens	1M tokens
Max output	128K tokens	128K tokens
Default effort	high	medium

Fable 5.2 is cheaper than Fable 5.1 was ($10 / $50), but still exactly double Opus 5.5 per token. Batch requests are 50% off on both. Fable also has a fast mode at $16 / $80 that Anthropic says runs up to 2.5× faster.

Per token isn't the whole story, though. A stronger model can finish a task in fewer turns and with fewer retries, so it can cost less per task even at a higher per-token price. That's why the number to track is cost per successful task, not cost per token.

The detail most comparisons miss: default effort

Both models have adaptive thinking always on, and you control how hard they think with the effort parameter. But their defaults are different: Fable 5.2 defaults to high, Opus 5.5 to medium.

So if you compare them with no settings, you're comparing Fable thinking harder against Opus thinking less. Some of the quality gap, and a lot of the cost gap, comes from that alone.

Set effort explicitly on both:

import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic();

async function ask(model: 'claude-fable-5-2' | 'claude-opus-5-5', prompt: string) {
  const msg = await client.messages.create({
    model,
    max_tokens: 32000,
    output_config: { effort: 'high' },
    messages: [{ role: 'user', content: prompt }],
  });
  console.log(model, msg.usage.input_tokens, msg.usage.output_tokens);
  return msg;
}

await ask('claude-fable-5-2', 'Refactor this function...');
await ask('claude-opus-5-5', 'Refactor this function...');

This type-checks against @anthropic-ai/sdk 1.8.0. Two notes on it:

• max_tokens covers thinking plus the answer. At high and above, give it room, or long answers get cut off.
• Log usage. Output tokens are where the 2× price shows up, and they're what you need for cost per task.

Allowed effort values are low, medium, high, xhigh and max. Anthropic's advice for Fable 5.2: start at high, try xhigh for the hardest long-running work once you've measured a gain, and keep max for testing the ceiling.

Migration gotchas

If you're swapping the model ID in existing code, check these first. Each one is from the Fable 5.2 docs:

• Forced tool use fails. tool_choice set to any or a specific tool returns a 400. Let the model choose tools instead.
• Drop sampling parameters. Omit temperature, top_p and top_k. Some values return a 400.
• Thinking can't be turned off. Use effort to make it cheaper and faster, not thinking: { type: "disabled" }.
• Data retention. Fable 5.2 requires 30-day retention by default. If your setup depends on zero data retention, check with Anthropic before switching.

So which one should you use?

Anthropic's own model guide is refreshingly direct: start with Opus 5.5 for most workloads, and use Fable 5.2 for demanding reasoning and long-horizon agentic work, or when Opus 5.5 at higher effort still falls short on your evals.

In practice, that gives you a simple test:

1. Pick 10 to 20 real tasks from your own work, not benchmark puzzles.
2. Run both models at the same effort level.
3. Record pass/fail and tokens for each run.
4. Compare cost per successful task.

If Fable passes tasks Opus fails, or finishes in fewer turns, it can pay for itself. If both pass, Opus at half the price is the better default.

We tried it: same prompt, both models

We gave both models the exact same prompt and recorded them side by side. [EDITOR: add the prompt, the effort setting used for each model, and 2–3 sentences on what you saw, then link the reel. Don't add numbers you didn't measure.]

One prompt isn't an eval, so treat it as a first impression, not a verdict.

The limits of all this

• The benchmark table is Anthropic's. Independent results are still coming in.
• Fable's safeguards can route some cybersecurity and biology requests to less capable models, which matters if you work in those areas.
• Prices and defaults change. Check the models page before you build a budget around them.

New models will keep arriving every few months. The part that lasts is having your own evals and a workflow that doesn't depend on any single model. If you want to build that workflow properly, it's what our Agentic Engineering course is about.

Which model are you running as your default right now, and what would make you switch?

Sources: Anthropic: Claude Fable 5.2 · Claude docs: Models overview · Claude docs: Effort · Claude docs: Fable 5.2 overview · MacRumors: Anthropic launches Claude Fable 5.2