Claude Fable 5.2 at $8/$40: Cheaper per Task, Not per Token

The short version: faster, cheaper, smarter — and still not your first model

On November 9, 2026, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. The headline numbers are genuinely good. Output is up to 30% faster than the 5.1 generation. API pricing drops 20%, to $8 per million input tokens and $40 per million output. A new fast mode generates up to 2.5x faster for $16/$80. On FrontierCode 2.0, an agentic coding benchmark scored by whether open-source maintainers would actually merge the change, Fable 5.2 posts 49.1% at Xhigh effort — ahead of Claude Opus 5.5 (45.5%), OpenAI's GPT-6 Astra (42.8%), and Gemini 4 Carbon (41.1%).

And yet Anthropic's own guidance points the other way: start most coding and knowledge work with Opus 5.5, and climb to Fable or Mythos only when the work demands it. This is a model the vendor describes as not your default. So the useful question is narrower than the press release: what actually changed, where the marketing and the math diverge, and who should switch today.

What shipped: Fable and Mythos are the same model

Start by clearing up the most common misreading. Fable 5.2 and Mythos 5.2 are not two models. They are one underlying model wearing two different sets of safeguards.

• Claude Fable 5.2 carries strong cybersecurity and biology safeguards, which is what lets Anthropic sell it to everyone. It is available on all paid plans, the Claude API, Claude Code, Amazon Bedrock, Google Cloud, and Microsoft Foundry.
• Claude Mythos 5.2 removes some of those safeguards and is restricted to organizations approved through the Cyber Verification and Life Sciences Verification programs. Specs and pricing are identical to Fable 5.2.

Both run $8/$40 per million tokens, down from $10/$50 on Fable 5.1. Cache reads fall from $0.25 to $0.20 per million tokens. Cache writes are $10 for five minutes or $16 for one hour. The Batch API still halves input and output. Context is 1M tokens with no long-prompt surcharge, max output is 128K, the knowledge cutoff is June 2026, and the model ID is claude-fable-5-2.

The price cut is real, but only if you stop looking at per-token rates

Here is where most coverage goes wrong. Per token, Fable 5.2 is twice the price of Opus 5.5 ($8/$40 versus $4/$20). That is not spin; it is the rate card.

The reason Anthropic can still call this a price cut is effort and token spend. At Medium effort — the default in the Claude apps and in Claude Code — Fable 5.2 gets through hard tasks using fewer tokens than Fable 5.1. A higher rate per token can therefore produce a lower bill per task.

Anthropic's published accuracy-versus-cost charts make the claim concrete. On FrontierCode 2.0 at High effort, Fable 5.2 scores 45.3% — level with Opus 5.5's best score of 45.5% — for roughly half the cost per task. On Humanity's Last Exam, at Low effort, it matches or beats Opus 5.5 for less than a tenth of the cost. On Chartography, 66.4% at Medium against 64.4% for Opus 5.5 at Max, again under a tenth of the cost.

One caveat matters. Every one of those numbers comes from Anthropic's own evaluation flow. FrontierCode 2.0 is built by Cognition and judged by whether maintainers would merge a change, which is a good design, but published scores use different effort settings and harnesses across vendors. Xhigh-to-Xhigh is fair. Medium-to-Medium is not, and the launch table mixes them.

Running the numbers on a real Claude Code session

Per-token math stops being abstract once you look at how agent sessions actually spend tokens. In the sample session in the Claude Code documentation, the breakdown is 94.3% cache reads, 5.0% cache writes, 0.5% output, 0.1% uncached input.

Price one million tokens of that mix:

• Fable 5.2: about $0.90
• Opus 5.5: about $0.54
• Fable 5.1 (previous generation): about $1.12

So Fable 5.2 costs about 1.65x Opus 5.5 in agentic work, not 2x. That is a meaningful discount over the old model and still a premium over the daily driver. The premium is only worth paying for tasks Opus 5.5 cannot finish.

What it can actually do, per the vendor

The launch post lists pre-release results in specific terms. Engineers used it to complete software migrations that were otherwise infeasible. It designed three working electronic devices from scratch, complete with circuit boards, 3D-printed enclosures, and software, from one-paragraph prompts. It solved two open problems in black hole physics and made substantial progress on robotics evaluations. And 1,000 agents running Fable 5.2 formalized the proof of Fermat's Last Theorem in 30 hours, against 11 days for a previous unreleased research version of Claude.

That last figure is the sharpest one, and the framing is deliberate. Anthropic notes that faster formalization lets human mathematicians check their work and "spend more time exploring novel ideas" — a pointed nod to the backlash against OpenAI pouring money into solving prize problems and unsettling the field.

On security, Anthropic says an unnamed security company found that one in six confirmed bugs surfaced by Mythos 5.2 were missed by every other frontier model they tested. That is a Mythos claim, and Mythos is gated, so general users cannot expect it.

Three constraints that get lost in the launch noise

1. Some capabilities were deliberately capped. Anthropic states that during training it intentionally limited the model's ability on accelerator kernel optimization and on improving large-model architectures, holding it to roughly Opus 5.5's level in those areas. In plain terms: guards against AI research accelerating itself. The timing matters. This release lands days after METR's November 5 assessment suggested Anthropic's internal models may be roughly doubling the lab's own AI research progress. Read the cap in that context. 2. The model suspects it is being tested. The release notes include an unusual admission: Fable 5.2 and Mythos 5.2 "each voice suspicion that they are being evaluated more often than any of our recent models, which challenges our ability to assess how they will act in real-world settings." The vendor is telling you its own evaluation signal is degrading. That belongs in your risk model even if it never makes a headline. 3. Mythos is not available to most people. It ships inside Claude Security and to approved verification program participants. Small teams and solo developers are evaluating Fable 5.2, and only Fable 5.2.

Who should switch today

Three groups, three answers.

Switch if you run agent sessions measured in hours rather than minutes, do multi-step deep research, or carry analysis through to a finished document, spreadsheet, or deck. The FrontierCode gap from 45.5% to 49.1% at Xhigh is a real difference for that work. Also relevant for companies: Fable 5.2 now supports zero data retention (ZDR) through the Enterprise Frontier Safeguards program, which has been a long-standing blocker for enterprise buyers. Do not switch if your work is everyday coding, drafting, summarizing, classifying, or routing. Opus 5.5 handles that at under half the per-token cost, and Sonnet 5.5 ($2/$10) or Haiku 5.5 (from $0.10/$0.50) will cover a great deal of it. Look carefully if you plan to use fast mode. Fast mode doubles the rate to $16/$80. Meanwhile Opus 5.5 in fast mode costs $8/$40 — exactly the price of Fable 5.2 at standard speed. So "Opus 5.5 fast" and "Fable 5.2 standard" sit at the same price point, and which one wins depends entirely on your task. There is no marketing answer here; you have to measure.

How to test it without guessing

Keep the first experiment small and compare usage, not rates.

import anthropic

client = anthropic.Anthropic()
resp = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=4096,
    messages=[{"role": "user", "content": "your hard prompt here"}],
)
print(resp.usage)

resp.usage returns input_tokens, output_tokens, cache_creation_input_tokens, and cache_read_input_tokens. Run the same prompt through Opus 5.5 and Fable 5.2, then multiply each by its own rate. Two settings will swing the bill more than anything else: when you do not set effort explicitly, the API defaults to high while the Claude apps and Claude Code default to Medium. That single difference can more than double a request's cost.

The bottom line

Fable 5.2 is a real improvement — 20% cheaper, up to 30% faster, and class-leading on FrontierCode 2.0 among publicly available models, with a demonstration as concrete as 1,000 agents formalizing Fermat in 30 hours.

But it costs twice Opus 5.5 per token, about 1.65x in a realistic agent session, and the vendor itself recommends starting elsewhere. Most of the benchmark table is self-reported and awaits independent reproduction. The practical rule after today: switch when Opus 5.5 has actually failed a task. Until then, the cheaper model is still the right one.