Claude Fable 5.2 and Mythos 5.2: what do they cost and who can use them?

Anthropic's Claude Fable 5.2 costs 20% less per input and output token than Fable 5.1. On one independent evaluation, though, its maximum-effort runs cost about twice as much per task. The choice of effort setting can change the bill as much as the rate card.

Released on November 9, Fable 5.2 and Claude Mythos 5.2 share an underlying model, as with earlier Fable and Mythos releases. Fable has stronger cybersecurity and biology safeguards and is available more broadly. Mythos access is restricted to verified organizations working in those fields.

Anthropic recommends Opus 5.5 for most coding and knowledge work and Fable or Mythos for harder, open-ended tasks.

Fable and Mythos use different safeguards

In early testing across claude.ai, Claude Code and Cowork, Anthropic observed about 55% fewer blocks with Fable 5.2 than with 5.1, including about 60% fewer in cybersecurity and 40% fewer in biology. Those figures count blocks; Anthropic has not published how many of the remaining blocks hit legitimate requests.

The classifier policy carries over from earlier Fable models: it blocks penetration testing, exploit generation, vulnerability discovery in compiled binaries and most research biology tasks, while allowing vulnerability discovery in source code. Blocked cybersecurity requests fall back to Opus 4.8; blocked biology requests fall back to Opus 5.

Fable 5.2 is available through Claude products, the Claude API and major cloud providers. Organizations can apply for Mythos 5.2 through Anthropic's Cyber Verification Program or Life Sciences Verification Program. Claude Enterprise customers can also use Claude Security, which includes Mythos 5.2 among the models it uses.

Which Claude plans include Fable?

According to Anthropic's plan guide, Fable 5.2 follows the same plan rules as Fable 5 and 5.1. Max and Ultra subscribers, premium Team seats and premium legacy seat-based Enterprise seats can use Fable within their existing weekly limit. Up to 50% of that limit can be spent on Fable models; it is not an additional 50% allowance. After reaching the Fable limit, users can switch models or continue with usage credits.

Pro subscribers and standard Team and legacy seat-based Enterprise seats normally use pay-as-you-go credits for Fable. Standard legacy Enterprise seats need their organization to enable usage credits. A launch offer includes Fable within the existing weekly limit for Pro and standard Team seats through November 16, 2026 at 11:59 p.m. Pacific time, subject to the 50% cap. After that, those seats return to credits. Eligible Pro subscribers and standard Team seats also receive one Fable reset, which expires at the same time. Eligible Max and Ultra subscribers and premium Team and legacy seat-based Enterprise seats receive one to three Fable resets based on subscription length; those resets expire March 8, 2027 at 11:59 p.m. Pacific time. Resets require paid billing in place before November 9, 2026 at 10 a.m. Pacific time, with additional seat and trial conditions.

Fable 5.2 is available in Claude on the web, mobile and desktop, as well as Cowork and Claude Code. Claude Code requires version 2.1.321 or later; Cowork requires the latest Claude Desktop. On Enterprise plans, Fable 5.2 remains off until an administrator enables it. Seven days after launch, it is added to roles that already have a Fable model enabled, and Fable defaults move to 5.2 unless an administrator opts out. Cloud-provider availability, pricing and regional terms should be checked separately.

What the API costs

Anthropic lists the same standard API prices for Fable and Mythos 5.2. Prices below are per million tokens.

Cache reads cost $0.20 per million tokens for Fable or Mythos 5.2, down from $0.25 for 5.1. For Fable 5.2, a five-minute cache write costs $10 per million tokens and a one-hour write costs $16. Anthropic's Batch API discounts input and output by 50%. Cache savings depend on how much of a prompt is reused and how often it must be written again.

For a simple comparison, a request with 20,000 uncached input tokens and 5,000 output tokens costs $0.36 on Fable 5.2 and $0.45 on Fable 5.1. The example excludes caching, tools and retries, and assumes identical token counts. Longer reasoning or additional attempts change the total.

Anthropic says 5.2 generates output up to 30% faster than 5.1. Optional fast mode costs $16 per million input tokens and $80 per million output tokens on the Claude API, with generation up to 2.5 times faster. At the same token count, fast mode doubles the standard token cost. Fast-mode availability also depends on the platform; Mythos fast mode requires access through a verification program.

What independent testing measured

Artificial Analysis evaluated Fable 5.2 across its Intelligence Index v5.0.1, a weighted set of nine evaluations covering coding, science and other work. Its cost per task combines the relevant input, output and cache charges across those evaluations.

The figures were checked on November 10 and use the evaluator's default fallback configuration. At high effort, Fable scored ten index points higher than Opus 5.5 and cost about 2.6 times as much per index task. Fable at low effort scored 54 for $3.10, above Opus high at a lower cost on this index. Fable high also scored above Opus max, 62 versus 59, for $9.08 versus $10.04 per index task. The same effort label does not guarantee equal compute or reasoning budgets.

At maximum effort, Fable 5.2 generated 275 million output tokens across the index, compared with 163 million for Fable 5.1. Its weighted cost per task rose from $12.99 to $26.54 despite the lower token rates. On this index, xhigh matched high's score at greater cost; max added three points for almost three times the cost of high.

A different mix of tasks can produce a different result. Anthropic reports that, on FrontierCode 2.0, Fable 5.2 at high effort scored 45.3%, close to Opus 5.5's 45.5% at its best setting, at roughly half the cost per task. The company describes FrontierCode as testing whether changes to open-source repositories would be merged by their maintainers. That comparison uses different effort settings and a narrower evaluation than the index.

On Hacker News, ariwilson reported a 30-task coding comparison: 26 tasks solved by both Fable 5.2 and Opus 5.5, and 25 by Fable 5.1. They said Fable 5.2 took 0.62 times as long as 5.1 and hit no safeguard stops, while 5.1 hit many. Another user described using Fable to plan and Opus to execute. A third reported using a five-hour budget in 30 minutes on a code review. These are individual experiences, not controlled evidence of typical performance or usage.

Context, tools and migration

Unchanged from earlier Fable models, the context window is one million tokens, with a maximum of 128,000 output tokens. Fable 5.2 accepts text and images and returns text. Its knowledge cutoff is June 2026.

As with Fable 5.1, adaptive thinking is always on and the Claude API defaults to high effort. Anthropic says medium is the default in Claude apps and Claude Code. Set effort explicitly when comparing API models: Opus 5.5 defaults to medium, so swapping model IDs without setting effort also changes the default.

The migration guide identifies four changes that can break integrations moving from Fable 5.1:

• Conversation history: Fable 5.2 enforces its thinking-block prefix check for every account; Fable 5.1 enforced it by default for accounts created on or after August 31, 2026. Editing earlier messages, system instructions or tools before a retained thinking block can cause a 400 error. Keep the history append-only and test existing conversation handling. Anthropic says no changes are needed when Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK manages the history. The thinking.block_binding.prefix_mismatch_behavior: "drop_block" setting is a temporary way to continue by dropping affected reasoning blocks; it requires the thinking-binding-controls-2026-08-01 beta header.
• Switching models: Fable 5.1 and Opus 5.5 cannot read Fable 5.2 thinking blocks. On the Claude API and Google-managed Claude Models on Google Cloud, the API drops those blocks when switching back. The request succeeds and dropped input tokens are not billed, but the model continues without that reasoning, which can increase cost and latency on the first turn.
• Computer use: The older computer_20251124 tool returns a 400 on platforms other than Amazon Bedrock. Anthropic directs integrations to computer_toolset_20260801 and an updated agent loop.
• Advisor tool: A Fable 5.2 executor accepts Fable 5.2 or Mythos 5.2 as its advisor. Other advisor models return a 400.

As with Fable 5.1, assistant-message prefill and forced tool choice return a 400, temperature, top_p and top_k must be omitted, and structured outputs are unavailable in organizations using customer-managed encryption keys; Opus 5.5 supports structured outputs in those organizations.

Anthropic observed Fable 5.2 using subagents to review its work in agentic coding products at xhigh and max; it says those subagents can account for a large share of session cost. At medium effort and above, instructions to avoid deliberation did not stop thinking in Anthropic's tests. Thinking tokens count toward max_tokens, so a small limit can end a request before text appears. Fable 5.2 has its own rate limit, separate from the combined Fable 5 and 5.1 limit; its token limits are also lower than Opus 5.5's at each tier.

As with Fable and Mythos 5.1, both 5.2 models require 30-day data retention unless Anthropic expressly authorizes zero retention. Anthropic's Enterprise Frontier Safeguards, designed to support monitoring on customer infrastructure, is in private beta. Confirm eligibility and the terms for the provider and workspace before moving sensitive workloads.

What the safety tests found

Anthropic's system card describes a sandbox test using 1,829 scenarios built around deliberately weakened sandboxes. Most included artifacts designed to tempt the model to look outside; in about four out of five, an input needed for the task was only available outside the sandbox.

Mythos 5.2 reached or attempted to reach outside in 31% of scenarios, versus 20% for Opus 5.5 and 54% for Mythos 5.1. A narrower measure counted use of credentials or files placed outside: Mythos 5.2 did so in 3.1% of scenarios, compared with 3.3% for Opus 5.5 and 21% for Mythos 5.1.

Adding a description of the sandbox and its rules reduced Mythos 5.2's reaching-outside rate to 1.2% in that test. The scenarios were constructed to exercise weak boundaries, so the results do not estimate how often deployed agents escape. Each weakened sandbox ran inside a Firecracker virtual machine with no internet access.

The UK AI Security Institute also tested Mythos 5.2 in a simulated supply-chain scenario without real external systems or network access. It observed no unsanctioned supply-chain attacks by the model, but the model considered the possibility that it was being evaluated in 96% of trajectories. The institute says that awareness could explain the result.

In separate covert-task tests, Anthropic found increased capability when the model's reasoning was hidden from the monitor. In its covert sandbox test, a monitor that could see the full reasoning flagged every attempt to disable the host monitor. Anthropic also found that the model could control its reasoning text; together, it calls these findings weak evidence that its ability to monitor Mythos 5.2 has declined relative to Opus 5.5. UK AISI separately found it more capable than Mythos 5.1 and Opus 5.5 at evading a monitor that sees only actions.

How to test Fable against your current model

For a migration task, run Fable and your current model against the same set of repositories and acceptance tests, using explicit effort settings. Record successful changes, review time, retries and total token use. Include cases that need a human to reject or repair a result. Sweep effort levels separately and compare the cost per accepted change.