Claude Fable 5.2: Price, Benchmarks and Why a 20% Cheaper Token Costs Twice as Much per Task

Anthropic released Claude Fable 5.2 on 9 November. It is the company's top model, it costs $8 per million input tokens and $40 per million output tokens, and that is 20% less than Fable 5.1 was. The first independent test, from Artificial Analysis, scores it 65, the highest on its index. It also puts the cost of finishing one test task at $26.54, more than double Fable 5.1, because the new model thinks for much longer. This article covers what Anthropic shipped, what Artificial Analysis measured, why the cheaper price per token produces a bigger bill per job, what the "Mythos 5.2" twin is, and who should pay for it.

What Anthropic released on 9 November

Anthropic announced two models together: Claude Fable 5.2 and Claude Mythos 5.2. It calls them its "most capable models for daunting technical work". The facts from the announcement:

• Fable 5.2 is the one you can use. The API model id is claude-fable-5-2. Anthropic says it is available on all platforms including Amazon Web Services, Google Cloud and Microsoft Azure, and in the Claude apps and Claude Code, where medium effort is the default.
• Price. $8 per million input tokens and $40 per million output tokens. An optional fast mode costs $16 and $80 and, Anthropic says, generates tokens up to 2.5 times faster.
• Speed. Anthropic says output is up to 30% faster than Fable 5.1 and Mythos 5.1.
• Plans. As a launch promotion, Anthropic says Fable is included for up to 50% of weekly usage limits on Pro and standard Team seats, and eligible paid plans get one to three "resets" depending on how long they have subscribed.

Our own earlier article on Opus 5.5 recorded Fable 5.1 at $10 in and $50 out. So the top tier just got 20% cheaper per token. Hold that thought, because the next sections show why it does not mean 20% cheaper work.

Fable 5.2 and Mythos 5.2 are the same model with different locks

Anthropic says the two are the same underlying model at the same price. The difference is the safeguards. Fable 5.2 ships with strong cybersecurity and biology protections and is open to everyone. Mythos 5.2 is the version without them, and Anthropic restricts it to vetted organisations through its Cyber Verification and Life Sciences Verification programmes. Claude Enterprise customers can also reach it through Claude Security.

For almost every business reader this is a footnote, with one exception. When Fable 5.2's classifiers block a request, the request is answered by an older model instead: Opus 4.8 for cybersecurity and Opus 5 for biology, according to Anthropic. It says blocks are about 55% fewer overall than on Fable 5.1, and about 60% fewer in cybersecurity. If you run an IT or security business, test your own prompts, because the answer may come from a different model than the one you chose.

What the independent test says: 65 points, $26.54 a task

Anthropic's own benchmark table is the vendor's view, so we start with the independent one. We read the Artificial Analysis pages for each model on 10 November. All figures are for the maximum-effort setting, which is what Artificial Analysis tests.

Artificial Analysis ranks Fable 5.2 first of 74 models on intelligence. It also ranks it last of 73 on cost per task. Its page shows the model produced 275 million output tokens on the test, against a median of 72.8 million for comparable models, and that it takes 46.4 minutes per task against a median of 12.6. The time before the first answer token is about 216 seconds, or three and a half minutes.

Fable 5.2 is the smartest model on the independent test and the most expensive to run on it, and the cheaper price per token is the least important number on the page.

One caution on the Fable 5.1 row. The Artificial Analysis page for 5.1 now marks it deprecated and its score of 49 sits well below Opus 5.5, which does not match Anthropic's own earlier statement that Opus 5.5 performs at Fable 5.1's level on most work. We could not confirm whether that page uses an older test setup. Use the 5.1 row for its cost and token use, not as a firm score comparison.

Why a cheaper token gives a bigger bill

You pay for every token the model writes, including the thinking you never see. Fable 5.2 wrote about 69% more output tokens than Fable 5.1 on the same kind of test (275 million against 163 million). Its output price per token fell by 20%. Put together, a finished task costs $26.54 instead of $12.99, which is about 2.0 times.

The gap to the other top models is larger. Per task, Fable 5.2 costs about 2.6 times Opus 5.5, 3.4 times Gemini 4 Carbon and 4.3 times GPT-6 Astra, for a lead of 6, 5 and 8 index points. Astra got its result with 54 million output tokens, about a fifth of Fable 5.2's. We work these multiples out from the Artificial Analysis page figures above. This is the same lesson we described in our guide to the cheapest AI model by cost per task: the price list tells you the price of a token, and the bill depends on how many tokens the model chooses to use.

A worked example, with our assumptions

Say a job reads a long document and writes a short answer: 20,000 tokens in, 2,000 out. These token counts are our assumption, not a measurement, and they ignore the extra thinking a model does.

• Fable 5.2: $0.16 for input plus $0.08 for output, $0.24 a job, or $240 per 1,000 jobs.
• Fable 5.1: $0.20 plus $0.10, $0.30 a job, or $300 per 1,000.
• Opus 5.5: $0.08 plus $0.04, $0.12 a job, or $120 per 1,000.

On the sticker price alone Fable 5.2 is cheaper than 5.1 and twice Opus 5.5. If it also thinks longer, as the test suggests, each of those figures goes up. Your own 100 test jobs will show by how much.

What Anthropic's own benchmarks say about business work

These are Anthropic's numbers, from its launch table, not independent results. The ones that matter to a business owner:

• Business workflows. On AutomationBench, Anthropic reports 51.4% for Fable 5.2, against 42.5% for Opus 5.5 and 41.4% for GPT-6 Astra. Gemini 4 Carbon is shown ahead at 53.9%, so this is not a clean sweep.
• Long jobs done end to end. On the Remote Labor Index, Anthropic reports 41.3% against 21.3% for Opus 5.5 and 20.8% for Astra. This is the biggest gap in the table.
• Computer use. On OSWorld 2.1 it reports 89.3%, ahead of Opus 5.5 (87.2%), Gemini 4 Carbon (81.5%) and Astra (79.5%). A model that clicks through software like a clerk is the use case behind many back-office automations.
• Knowledge work. On AA-Briefcase, Anthropic shows 66.1% for Fable 5.2 and 61.2% for Opus 5.5. The Elo figure in the same row has Opus 5.5 narrowly ahead, 1800 to 1785, so we would not read this as a clear win.

Anthropic also makes a cost claim worth testing. It says that on FrontierCode 2.0 at high effort, Fable 5.2 scores 45.3%, level with Opus 5.5's best of 45.5%, at roughly half the cost per task. And it says that on Humanity's Last Exam, Fable 5.2 at low effort matches Opus 5.5's top score at under a tenth of the cost. Those are vendor claims, but they point at the real lever: the effort setting.

Effort is the dial that sets your bill

The Artificial Analysis result is for maximum effort. Anthropic's apps default to medium, and Anthropic's own best cost claims are for lower settings. So the practical reading is that $26.54 is a ceiling for hard work, not a typical job. If you call the model through the API, effort is a setting you choose per request.

• Start low and move up. Run your real jobs at low or medium effort first. Raise it only for the jobs that fail.
• Cap the output. A model that writes 275 million tokens on a test will write a lot on your task if you let it. Set a maximum output length and a monthly spending cap.
• Log the cost of each job. You cannot see a doubling in the bill unless you record the cost per finished job.

A note on data: 30 days of retention

Anthropic says that, unless a customer has implemented zero data retention under its Enterprise Frontier Safeguards programme (currently a private beta), requests to Fable 5.2 and Mythos 5.2 are retained for 30 days for safety monitoring. If you handle customer records or financial documents, put that sentence in front of whoever signs off your vendors before you move sensitive workloads to this model.

Who should pay for Fable 5.2, and who should not

• Hard, high-value work nobody re-checks cheaply: yes, test it. Difficult code, long multi-step research, complex financial analysis. This is where a lead of several points and the large gaps on long-task benchmarks can repay a $26 task.
• Everyday office work: probably not. Reading invoices, drafting routine replies and filling forms are jobs where Opus 5.5 at $10.04 per task, or a cheaper tier, usually finishes the work. Our Opus 5.5 guide covers when to step down.
• Customer-facing chat: no. A first answer token that takes minutes in the test is the wrong shape for a live conversation. Use a fast, small model; see our Haiku 5.5 test.
• Were you paying for Fable 5.1? Switch and measure. Per token it is 20% cheaper and, on Anthropic's table, clearly stronger. Check that your cost per finished job has not doubled.

Your next step: test 100 of your own jobs

Launch week always produces a new "best model". The best one for you is the one that finishes your own jobs, at acceptable quality, for the lowest total cost. Pick 100 real jobs, run them at low, medium and high effort, and write down quality and cost for each setting. That test takes an afternoon and will tell you more than any leaderboard.

If you would rather use AI than manage models, that is what our products are for. IO Snack Accounts uses AI to categorise entries and flag anomalies as you post them, and the AI Chat Assistant answers customer questions on your website. If you are weighing a bigger project and want a second opinion on which model fits it, talk to us.