Claude Fable 5.2: 20% Cheaper, Top Benchmarks — Worth It?

Anthropic released Claude Fable 5.2 on Monday, November 9, 2026 — the official announcement describes it as the company’s most capable model for demanding technical work — and the headline is a rare one in AI pricing: the company’s most powerful model got 20% cheaper at the same time it got faster and smarter. Fable 5.2 now costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50, and it sits first on the independent Artificial Analysis Intelligence Index with a score of 65 — ahead of Gemini 4 Carbon at 60 and Claude Opus 5.5 at 59.

But the price cut is not the whole story, and for most freelancers and small teams it is not even the most important number. At the effort level that earned that top score, a single benchmark task costs an average of $26.54 to run — about 2.6 times what Opus 5.5 costs for the same task. Fable 5.2 is cheaper per token than its predecessor and still the most expensive way to get a job done in Anthropic’s lineup. Whether you should switch comes down to what kind of work you hand it, which effort setting you use, and whether your plan includes it at all.

What Anthropic actually launched on November 9

Fable 5.2 arrived with a sibling, Claude Mythos 5.2, and the first thing to understand is that they are the same underlying model wearing two different sets of safeguards. Fable is the broadly available version: it carries cybersecurity and biology safety classifiers that block high-risk work like penetration testing, exploit generation, and most research biology. Mythos drops those safeguards, and you cannot just sign up for it — it is limited to organizations vetted through Anthropic’s Cyber Verification Program or Life Sciences Verification Program.

For almost everyone reading this, Mythos is irrelevant, and that is by design. Fable 5.2 is the model you can actually use. It is generally available on the Claude API under the model ID claude-fable-5-2, in Claude apps and Claude Code, and through Amazon Bedrock, Google Cloud, and Microsoft Foundry. It keeps Fable 5.1’s 1 million-token context window and 128,000-token maximum output, with a June 2026 knowledge cutoff. Anthropic says it generates output up to 30% faster than the 5.1 models, and it adds an optional fast mode that generates tokens up to 2.5 times faster still — at double the price, which we will get to.

Anthropic also tightened the safeguards’ aim rather than loosening them. In early testing across Claude, Claude Code, and Cowork, Fable 5.2’s classifiers blocked about 55% fewer requests than Fable 5.1’s — cybersecurity blocks fell about 60% and biology blocks about 40% — while Anthropic says the safeguards are as robust as before. Fewer false alarms, not fewer guardrails.

The price cut, in plain dollars

Here is the pricing table that matters, all per million tokens:

A 20% cut on the headline rate is real money, but notice the second row of the comparison: Fable 5.2 still costs exactly twice as much per token as Opus 5.5. That is the number that decides your bill. The launch discount narrows the gap with the old Fable; it does nothing to the gap with the model most people should be comparing it against.

Cache pricing softens the blow for coding work, which is Fable’s home turf. Prompt-cache reads cost $0.20 per million tokens, so re-reading a large codebase you have already loaded is nearly free — the expensive part is always the fresh output. The Batch API, for work that can wait, cuts both rates in half to $4 and $20, which is exactly Opus 5.5’s standard rate — and our Fable vs Opus cost comparison walks through the same per-task numbers. If your workload is asynchronous — overnight test generation, batch document processing, scheduled analysis — Batch API Fable costs the same per token as real-time Opus, which changes the decision completely.

Fast mode deserves a warning label. It costs $16 input and $80 output per million tokens, double the standard rate, for up to 2.5 times faster token generation. That is a token-generation claim, not a promise that your task finishes 2.5 times sooner — a long agentic task spends much of its time thinking, calling tools, and waiting on your systems, none of which fast mode speeds up. Fast mode is also limited to select organizations on certain platforms, including Claude Code, the Claude Platform, Google Cloud, and Microsoft Foundry.

What it actually costs to run a job

Per-token prices are abstract. Per-task costs are the bill. Here is the math, worked from the verified numbers, so you can rerun it with your own token counts.

The benchmark task. Artificial Analysis publishes what it costs to run its Intelligence Index at each effort level. At maximum effort, Fable 5.2 averages $26.54 per task — the highest per-task cost of the 73 models on its chart — against about $10.21 for Opus 5.5 and $12.99 for Fable 5.1. That is the price of the headline score of 65. It also takes about 44 to 45 minutes per task at that setting, against a 12.9-minute median across comparable models. Artificial Analysis describes the model at maximum effort as “notably slow per task and very verbose.”

A typical coding session. Say an agent session reads 20,000 fresh input tokens of instructions and context, then writes 30,000 output tokens of code and explanation. On Fable 5.2 at standard rates that is 20,000 × $8/1,000,000 = $0.16 of input plus 30,000 × $40/1,000,000 = $1.20 of output: $1.36 total. The same session on Opus 5.5 costs $0.08 + $0.60 = $0.68. Exactly half. Run ten such sessions a day and the monthly difference is about $204 on Fable versus $102 on Opus — before you count the cached codebase reads, which are nearly free on both.

A monthly budget. Take a freelancer spending $100 a month on API calls, with a typical 3-to-1 ratio of input to output tokens. Fable 5.2’s blended rate is (3 × $8 + 1 × $40) ÷ 4 = $16 per million tokens, so $100 buys about 6.25 million tokens. Opus 5.5’s blended rate is (3 × $4 + 1 × $20) ÷ 4 = $8 per million, so the same $100 buys 12.5 million tokens — twice the work. The question is never “is Fable better?” It is “is Fable’s extra quality on this specific task worth half my token budget?”

One long answer. A single maximum-length response of 128,000 output tokens costs about $5.12 in output alone on Fable 5.2. Break long outputs into sections and you pay the same per token, but you get checkpoints you can actually review — and you stop paying when the first section shows the model misunderstood the brief.

The benchmark picture — and what it doesn’t prove

Fable 5.2’s launch numbers look strong, and they are worth reading carefully rather than repeating. On Anthropic’s own published benchmark table, Fable 5.2 scores 49.1% on FrontierCode 2.0, an agentic coding test that measures whether a model’s code changes would be merged by open-source maintainers — against 32.4% for Fable 5.1 and 45.5% for Opus 5.5. It also scores 69.9% on FrontierSWE v2 and 64.9% on a robotics benchmark at maximum effort, more than double GPT-6 Astra’s 29.8% and Fable 5.1’s 27.6%.

The independent numbers tell a more nuanced story. Artificial Analysis ranks Fable 5.2 first of 74 models with an Intelligence Index of 65 at maximum effort, ahead of Gemini 4 Carbon at 60, Opus 5.5 at 59, and GPT-6 Astra at 57, and Vals AI ranks it first of 36 models at 71.31%. But the same Artificial Analysis page shows the catch: at High and extra-high effort, Fable 5.2 still scores 62 — still first overall — while using fewer output tokens per task than Opus 5.5 at maximum effort. Anthropic says Fable 5.2 at High effort matches Opus 5.5 at its highest setting on coding, at roughly half the cost per task. That High-effort result is the most useful number in the whole launch.

Fable 5.2 does not win everything. Anthropic’s own table puts it at 51.4% on AutomationBench, behind Gemini 4 Carbon at 53.9%, and third-party coverage notes it did not lead in legal, tax, or certain multimodal programming tasks. Every benchmark here comes with the same caveat: Anthropic’s numbers are vendor-reported, run by Anthropic or evaluators it selected, and none have been independently replicated yet.

Then there are the launch demos, which need proportion. Anthropic says a team of about 1,000 Fable 5.2 agents formalized the proof of Fermat’s Last Theorem in Lean in about 30 hours, and another 1,000-agent team built a prototype operating system capable of booting on a laptop within eight days. Those are Anthropic’s own experiments, designed to show what coordinated agents can do — not independent proof, and not a promise that one Fable session will do the same for your project. Judge the model on the coding and reasoning benchmarks, not the moonshots.

One rumor worth clearing up: there is no Claude Fable 5.5. Weeks of leaks and social posts predicted a “Fable 5.5” launch in late September or early October, and none of them named the model or the date correctly. Anthropic released Fable 5.2 instead. If you still see “Fable 5.5” circulating, it is stale speculation, not a second model you missed.

When Fable 5.2 pays for itself — and when Opus 5.5 wins

Anthropic’s own guidance is refreshingly blunt: it still recommends Claude Opus 5.5 as the default for most work, and positions Fable 5.2 for long-horizon agentic tasks. Here is the decision table:

The pattern is simple: Fable earns its premium on long, hard, open-ended work where the quality of the result matters more than the token count. It loses on anything routine, short, or chatty, where Opus delivers nearly the same answer for half the price. A freelancer’s sensible default is Opus for the day-to-day and Fable for the one or two jobs a week where the model needs to think hard for a long time.

The effort setting is the real price control

If you take one practical lesson from this launch, make it this: the effort setting, not the model choice, is where your Fable bill is decided. Adaptive thinking is always on in Fable 5.2, and the effort setting controls how deep it reasons. The default is high.

At maximum effort, Fable 5.2 scores 65 on the Artificial Analysis index and costs $26.54 per task. At High and extra-high effort, it scores 62 — still first overall — while using fewer output tokens per task than Opus 5.5 at maximum effort. Three index points buy you a doubling of cost.

Set your effort deliberately. Use High for serious coding and research, drop to lower effort for anything where a good answer beats a perfect one, and reserve Max for the rare task where you would genuinely pay $26 to get the best available answer — a production bug you cannot reproduce, a security review, a decision you cannot reverse. And when you compare Fable with Opus, compare Fable at High effort with Opus at Max: that is the comparison Anthropic itself says lands at roughly half the cost per task for comparable coding quality.

This is also why the “20% cheaper” headline misleads in both directions. A careless Fable user running everything at Max effort will spend more than a careful Fable 5.1 user did. A deliberate one running High effort on the right tasks will spend far less than either.

What Pro and Max subscribers actually get

Subscription access is where most readers will meet Fable 5.2, and the launch comes with a one-week window that ends soon. Pro costs $20 a month and Max starts at $100 a month — our Claude Pro, Max and Ultra plans guide breaks down what each tier includes — and Fable follows the same plan rules as Fable 5 and 5.1, with a launch offer layered on top.

Through November 16, 2026 at 11:59 p.m. Pacific time, Pro subscribers and standard Team seats can use Fable models within their existing weekly limit — up to 50% of that limit can go to Fable — plus one Fable limit reset. Max and Ultra subscribers, premium Team seats, and premium legacy Enterprise seats already have Fable included up to 50% of their weekly limit, and they get one to three Fable resets depending on how long they have been paying. Those resets expire March 8, 2027, and only go to accounts with paid billing in place before 10 a.m. Pacific on November 9.

After November 16, Pro goes back to usage credits for Fable — pay-as-you-go at API rates — and any unused Pro reset expires. Max at $100 a month becomes the cheapest plan that includes Fable without usage credits. If you are on Pro, this week is genuinely the cheapest Fable you will get: try it before November 16, watch Settings and then Usage to see how fast Fable drains your limit, and use the reset if you hit the wall mid-task. A once-a-month Fable task is cheaper on Pro credits than on a $100 plan.

One more subscription note: GitHub Copilot added Fable 5.2 the same day for Pro+, Max, Business, and Enterprise users, but admins must switch on a policy that is off by default. If you use Copilot at work and cannot see the model, that policy switch — not your plan — is the likely reason.

The upgrade traps: pinned models, retention, and fast mode

Three things can quietly cost you money or break your setup when you move to Fable 5.2.

Three things can quietly cost you money or break your setup when you move to Fable 5.2.

Pinned model IDs keep paying the old rate. If your code, agent configuration, or integration calls claude-fable-5-1 by name, it keeps running 5.1 at $10/$50 until you change the ID to claude-fable-5-2. Audit every place you named the model — Claude Code settings, API calls, agent configs, CI scripts — and update them deliberately rather than assuming an upgrade happened.

The 30-day retention requirement. Fable 5.2 requires 30-day data retention by default, and API requests from organizations without it fail with a 400 error. Customers already on zero-data-retention arrangements can continue under those arrangements, but everyone else needs the retention setting in place. If your clients’ data cannot be retained for 30 days, Fable 5.2 may not be usable for their work at all, whatever the benchmarks say.

Fast mode is a premium, not a default. At $16/$80, fast mode doubles your token cost for up to 2.5 times faster token generation, and it is only available to select organizations on certain platforms. Treat it like a rush-delivery fee: worth it when the deadline justifies it, wasteful as a standing setting.

The honest limits

A fair guide says what the model does not do. Fable 5.2 is not the best model for every task: it trails Gemini 4 Carbon on AutomationBench, and it did not lead on legal, tax, or certain multimodal programming evaluations. It is slow at maximum effort — about 44 minutes per Artificial Analysis task — and verbose, which is a real cost when you pay per output token. Its safeguards will block legitimate cybersecurity and biology work, and the only remedy is Mythos access through a verification program most readers will never qualify for.

There is also a context worth naming: this launch landed in the same week Anthropic set an IPO price range valuing the company at up to $1 trillion, with trading expected around November 19. None of that changes what Fable 5.2 costs or does. Buy the model for your work, not for the company’s week in the news. Our earlier coverage of the IPO roadshow explains the terms if you are curious — but the pricing decision stands on its own.

Should you switch today?

If you are on Pro, yes — try it this week, before the November 16 launch offer ends. Spend up to half your weekly limit on Fable, run one genuinely hard task at High effort, compare the result with what Opus 5.5 gives you on the same prompt, and check how much of your limit it ate. That experiment costs you nothing beyond the subscription you already pay, and it answers the only question that matters: does Fable’s extra quality show up in your work?

If you pay per API token, switch selectively. Move your hardest weekly task to Fable 5.2 at High effort, keep everything else on Opus 5.5, and watch the bill for a month. A 2x token price is worth it only when the task is hard enough that Opus’s answer costs you a second attempt, a debugging round, or a missed edge case. For routine work, Opus at half the price is still the right default — Anthropic says so itself.

And if you run agents all day, set the effort level before you set the model. High effort Fable beats maximum-effort Opus on coding quality at roughly half the per-task cost.

Frequently Asked Questions

How much does Claude Fable 5.2 cost?

Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1 — a 20% cut. Prompt-cache reads cost $0.20 per million tokens, the Batch API halves both rates to $4 and $20, and an optional fast mode costs $16 and $80 per million tokens.

Is Claude Fable 5.2 better than Opus 5.5?

On the hardest tasks, yes — Fable 5.2 scores 65 on the Artificial Analysis Intelligence Index against 59 for Opus 5.5, and 49.1% on FrontierCode 2.0 against 45.5%. But it costs exactly twice as much per token, and at maximum effort a benchmark task costs about $26.54 against roughly $10.21 for Opus.

What is the difference between Fable 5.2 and Mythos 5.2?

They are the same underlying model with different safeguards. Fable is broadly available with cybersecurity and biology safety classifiers; Mythos drops those safeguards and is limited to organizations vetted through Anthropic’s Cyber Verification Program or Life Sciences Verification Program. Pricing, context, and output limits are identical.

Can Claude Pro subscribers use Fable 5.2 for free?

Not exactly. Through November 16, 2026 at 11:59 p.m. Pacific, Pro subscribers can use Fable models within their existing weekly limit — up to 50% of it — plus one Fable limit reset, as a launch offer. After that, Fable returns to pay-as-you-go usage credits on Pro.

What is Fable 5.2’s fast mode, and is it worth it?

Fast mode generates tokens up to 2.5 times faster at double the price — $16 input and $80 output per million tokens. It speeds up token generation, not the whole task, so a long agent run will not finish 2.5 times sooner. Use it for deadline-driven work, not as a default.

Why does Fable 5.2 require 30-day data retention?

Fable 5.2 requires a 30-day data retention period by default for safety monitoring, and API requests without it fail with a 400 error. Customers already on zero-data-retention arrangements can continue under those arrangements. If your client work cannot be retained for 30 days, check your retention setting before adopting Fable 5.2.