Claude Fable 5.2: Price, Benchmarks, and Which Effort Level Is Worth Paying For

Anthropic released Claude Fable 5.2 on November 9, 2026. The API model ID is claude-fable-5-2. It costs $8 per million input tokens and $40 per million output, 20% less than Fable 5.1, and it ranks first on the Artificial Analysis Intelligence Index and the Vals Index, the two independent rankings with results we found as of November 10.

The token price is only half of the bill. Fable 5.2 can also spend far more tokens on the same work, and how many depends on the effort level you pick.

Fable 5.2 vs Fable 5.1: what changed

Anthropic describes the release as an improvement in coding, reasoning, and vision, with output up to 30% faster. Artificial Analysis shows the max effort versions at 101 and 87 tokens per second, a gap of about 16%.

Fable 5.1 stays available, so nothing forces a move today. Anthropic lists it as a legacy model and recommends Opus 5.5 as the place to start for most work.

Is Fable 5.2 really cheaper?

Per token, yes. Per task at the highest setting, no.

Artificial Analysis runs nine evaluations for its Intelligence Index and reports what each model spends on them. These are the figures for Fable 5.2 at all five effort levels, next to other models, as listed on November 10:

The max effort comparison explains the complaints you will see this week. Fable 5.2 at max used 275 million output tokens across the index, against 163 million for Fable 5.1 at max, and about three times the input tokens. A 20% cut in the token price cannot offset that, so cost per task went up 2.04 times.

That is not the whole story, and three other rows matter more if you are choosing a setting:

• Fable 5.2 at high beats Fable 5.1 at max and costs less. It scores 62 against 49, at $9.08 against $12.99.
• Fable 5.2 at high also beats Opus 5.5 at max. It scores 62 against 59, at $9.08 against $10.04. Anthropic says that on FrontierCode 2.0, high scores about level with Opus 5.5 at max for roughly half the cost per task. On the Artificial Analysis index the score holds up, and the saving is about 10%.
• Fable 5.2 at medium costs about the same as Opus 5.5 at high. It scores 56 against 52, at $3.80 against $3.46.

GPT-6.1 Sol stays in a different price range. At max it scores 55 for $1.27 per task, about a third of what Fable 5.2 at medium costs for one point less.

Which effort level to use

Fable 5.2 has five levels: low, medium, high, xhigh, and max. Thinking is always on and cannot be turned off, so effort is your main control over the bill.

High is the sensible default for hard work. Anthropic’s docs say most of the capability gain is already there at high. The Artificial Analysis numbers agree, with high and xhigh tied at 62.

Medium is the one to test for routine agent work. It costs 58% less than high and drops six points on the index. Anthropic made medium the default in Claude apps and Claude Code. On the API the default is high, so the same model can cost much more in your own code than in Claude Code, without anyone choosing that.

Xhigh did not earn its price on this index. It costs 63% more than high for the same index score. Anthropic reserves it for the hardest long-running work, where you have measured a gain.

Max is for measuring the ceiling. Anthropic’s docs reserve it for that and for the rare task where time and tokens do not matter. At 2.9 times the cost of high for three points, it is hard to justify as a daily setting.

Low saves little. At $3.10 it is only 18% cheaper than medium for two fewer points.

Two behaviors to plan for. In an agentic coding product at higher effort levels such as xhigh and max, Anthropic saw the model start subagents to review its own work, which Fable 5.1 did not do on the same tasks. Subagents can be a large part of what a session costs. Long tasks also sometimes end early with a list of what remains, and this happens more often at low and medium.

What we measured ourselves

We ran one coding task through Atomic Agent, our open-source agent, on OpenRouter with the provider pinned to Anthropic. The task is a stock library of about 1,400 lines with 13 planted bugs and one missing feature, scored by 59 hidden tests. Each setting ran three times.

Medium cut the cost by 76% and the time by 84%, with the same tests passing. Fable 5.2 at high spent $5.83 more per run to reach the same score on this task.

It is one task, and both settings scored 59 out of 59, so it says nothing about the ceiling of either level. The full results, with Opus 5.5 at the same two levels, are in our Fable 5.2 vs Opus 5.5 comparison.

Pricing

List prices from Anthropic’s pricing page. OpenRouter lists the same base rates.

The full 1M context is billed at the standard rate. Fast mode raises output speed by up to 2.5 times and does not shorten the wait for the first token. It does not work with the Batch API or Priority Tier. Anthropic’s general rule of 1.1 times the price for US-only inference applies to Claude 4.6 and later.

Independent results from Vals

Vals AI tested Fable 5.2 at high effort and published the results on release day.

The first place is narrow. Claude Sonnet 5.5 is at 70.85% and Gemini 4 Carbon at 70.45%, and Fable 5.2’s own margin is 0.96 points, so the top three are close to a tie at $17.94 per test for Fable 5.2.

The weak spots are the useful part. Legal and tax work rank mid-table or lower, and Opus 5.5 beats it on SWE-bench Multimodal, 70.21% to 62.29%. A first place on a general index does not mean first place for your field.

Benchmarks from Anthropic

Anthropic’s announcement figures, at each model’s highest available effort unless noted. The Fable 5.2 FrontierCode score is at xhigh; at high it is 45.3%. These are the lab’s own numbers.

The gains over Fable 5.1 are large on agentic work. Remote Labor Index more than doubled, and FrontierCode 2.0 went up by about half. Anthropic’s lead example is a Lean formalization of Fermat’s Last Theorem run by 1,000 Fable 5.2 agents in about 30 hours. Opus 5.5 with the same setup reached 26% in 18 hours. Anthropic gave no dollar cost for that run.

Named customers report results on long jobs. Uber describes a 10 to 12 hour security audit with about 170 reviewer agents. Snowflake calls it the strongest model on its 103-task data engineering benchmark. A footnote in the announcement says outside organizations did not review the post and may not agree with its claims.

Fable 5.2 vs Mythos 5.2

They are the same underlying model with the same specs and prices. Fable 5.2 adds stronger cybersecurity and biology safeguards, which is what lets Anthropic offer it to everyone.

Mythos 5.2 (claude-mythos-5-2) is available through the Cyber Verification Program, the Life Sciences Verification Program, and to enterprise customers through Claude Security. Anthropic claims it finds 10.2% more vulnerabilities than Opus 5.5 on CVE Rediscovery, while a footnote says vulnerability identification is roughly on par with Opus 5.5 on Anthropic’s own evaluations. Read both before you pay for access.

Blocks, fallbacks, and data retention

Fable 5.2 blocks high-risk cybersecurity work, such as penetration testing and exploit generation, and most research biology. On the API a blocked request returns HTTP 200 with stop_reason: "refusal". Anthropic says blocked users get a response from Opus 4.8 for cybersecurity and Opus 5 for biology. On the API, that retry is something you set up, through server-side fallback (beta), the SDK middleware, or your own code.

Anthropic says blocks fell about 55% compared with Fable 5.1, with about 60% fewer in cybersecurity and 40% fewer in biology. That figure comes from early testing on claude.ai, Claude Code, and Cowork, and Anthropic does not name the API in it. Vals recorded a 1.88% fallback rate and no refusals in its run.

If your agent runs long jobs, decide what should happen when a step is declined. If you set up a fallback, part of the run is done by a different model, so log when it happens. Some refusal categories are billed even though no output comes back; cyber refusals are not.

Fable 5.2 requires 30-day data retention. Under zero data retention it is available only with Anthropic’s express authorization, and eligible customers can stay on ZDR while they move to Enterprise Frontier Safeguards.

Anthropic also says it deliberately limited some AI research capabilities during training, to about the level of Opus 5.5, in areas like accelerator kernel optimization. If that is your field, test it first.

Where to use it, and plan limits

On the API, Fable 5.2 is on the Claude API, Amazon Bedrock (anthropic.claude-fable-5-2), Claude Platform on AWS, Google Cloud, and Microsoft Foundry. It is also on OpenRouter as anthropic/claude-fable-5.2.

Plan access is a promotion for now. From November 9 to 16, 2026, Pro and standard Team seats include Fable 5.2 for up to 50% of weekly limits, plus one reset. Max, Ultra, and premium seats get one to three resets depending on how long you have subscribed, and these expire on March 8, 2027. Free plans are not eligible. On Pro and standard Team, after the cap or the week, usage bills to credits if you turned them on. Anthropic has not said what happens to plan access after that.

What early users report

These are first-day anecdotes, so treat them as hints.

A recurring complaint on Hacker News is cost per task erasing the price cut, with others pointing to medium as the fairer figure. One user on a $100 plan said a review spawned more subagents than requested and used the whole five-hour allowance in under 20 minutes. Another said you need more ambitious prompts to see a difference from Opus 5.5.

Ethan Mollick, testing with early access, found it very capable across fields and also noticed it missing problems in its own output. ContextBolt ran three small jobs at medium and paid 1.7 to 6.5 times as much as on Opus 5.5 for broadly equal results.

If you are migrating from Fable 5.1

Anthropic lists four changes that can break code. Edits to earlier messages in a conversation with thinking blocks now return a 400 error on every account. Fable 5.2 thinking blocks can be read only by Fable 5.2 and Mythos 5.2. Computer use goes through computer_toolset_20260801, except on Amazon Bedrock, where the earlier tool still works. The advisor tool needs a Fable 5.2 or Mythos 5.2 advisor. The details are in the migration guide and in our comparison.

One small change affects prompts. Lines asking the model to think less no longer stop thinking at medium and above, so use the effort setting for that.

How to try Fable 5.2

To repeat our test on your own tasks, download Atomic Agent for Mac, Windows or Linux, or install the terminal version. The public release is a little older than the pre-release build we tested, so your numbers may differ.

curl -fsSL https://atomicagent.io/install | sh
atomic-agent

On Windows, run irm https://atomicagent.io/install.ps1 | iex in PowerShell. The source is on GitHub.

Connect OpenRouter with your own key and choose anthropic/claude-fable-5.2 from the live model list; see our OpenRouter guide. We pinned the provider to Anthropic, and OpenRouter also serves this model through Azure. Atomic Agent sends no effort value by default, so you get high. To run at medium, add one line to your openrouter entry under llm.providers in ~/.atomic-agent/config.json:

"extraBody": { "reasoning": { "effort": "medium" } }

The setting applies to every model on that provider entry, so change it when you switch models. More in our models guide. If you are comparing agents rather than models, see our list of open-source Claude Code alternatives.

The takeaway

Fable 5.2 ranks first on the Artificial Analysis and Vals indexes and is a clear step up from Fable 5.1. The lower token price does not mean a lower bill at the top setting, where a task costs twice as much as on 5.1. Set effort yourself. Use high for hard work, test medium for routine agent runs, and treat xhigh and max as settings you turn on after measuring a gain. For most work, Anthropic still points to Opus 5.5 first, and our side-by-side test shows why.