Claude Fable 5.2: Pricing, API ID & Migration Checklist

Direct answer: Claude Fable 5.2 is available now with the API model ID claude-fable-5-2. It has a 1-million-token context window, up to 128,000 output tokens, always-on adaptive thinking, and standard API pricing of $8 per million input tokens and $40 per million output tokens. Use it for unusually difficult, long-running technical work—not as the automatic default for every prompt.^[1][2]

Copy-paste migration checklist

Use this before moving production traffic from Fable 5.1, Mythos 5.1, Opus 5.5 or an older Opus model:

• [ ] Change the model ID to claude-fable-5-2.^[3]
• [ ] Set effort explicitly; start with medium or high, then test your own workloads.^[3][5]
• [ ] Raise max_tokens if thinking can consume the available output budget.^[3]
• [ ] Keep conversation history append-only and return thinking blocks unchanged.^[3]
• [ ] Do not use forced tool choice with any or a named tool; those settings return a 400 error.^[3]
• [ ] Remove temperature, top_p and top_k unless Anthropic documents a supported value.^[3]
• [ ] Replace computer_20251124 with computer_toolset_20260801 where required.^[3]
• [ ] Handle stop_reason: "refusal" and inspect stop_details.category.^[3][5]
• [ ] Check rate limits before moving live traffic.^[3]
• [ ] Confirm the 30-day data-retention requirement is acceptable for the target workspace.^[2][3]
• [ ] Re-run quality, cost and latency evaluations instead of assuming settings tuned for an older model will transfer.^[3][5]
• [ ] Roll out gradually and keep a tested fallback path.^[3]

Anthropic’s migration guide says the move from Fable 5.1 is mostly drop-in, but preserved-thinking validation, the older computer-use tool, advisor compatibility and model-switching behavior can break integrations.^[3]

Claude Fable 5.2 at a glance

Item	Claude Fable 5.2
API model ID	claude-fable-5-2
Released	November 9, 2026
Context window	1 million tokens
Maximum output	128,000 tokens
Standard input	$8 / million tokens
Standard output	$40 / million tokens
Cache read	$0.20 / million tokens
Batch API	50% discount on input and output
Fast mode	$16 input / $80 output per million tokens
Thinking	Adaptive and always on
Default API effort	high
Inputs and output	Text and images in; text out
General access	Claude Fable 5.2
Verified-program access	Claude Mythos 5.2

The specification, lifecycle status and pricing above come from Anthropic’s current model documentation. Anthropic lists the model as active and says retirement will be no sooner than November 9, 2027.^[2]

Fable 5.2 versus Mythos 5.2

Fable 5.2 and Mythos 5.2 use the same underlying model. Fable adds cybersecurity and biology safeguards so it can be offered broadly. Mythos is limited to organizations accepted through Anthropic’s verification programs for approved cybersecurity or life-sciences work.^[1][2]

That distinction matters operationally:

• Choose Fable 5.2 for generally available advanced coding, research, analysis and long-horizon agent work.^[1][2]
• Consider Mythos 5.2 only if your organization has a legitimate covered use case and has completed the relevant verification process.^[1][2]
• Do not assume buying a paid Claude plan automatically grants Mythos access.^[1][2]

Should you use Fable 5.2 or Opus 5.5?

Anthropic recommends starting most coding and knowledge tasks with Opus 5.5, then moving up to Fable or Mythos when the scope and difficulty justify it. Fable 5.2 costs twice as much as Opus 5.5 at standard API rates: $8/$40 versus $4/$20 per million input/output tokens.^[1][2]

A workload that consumes 1 million input tokens and 200,000 output tokens would cost about $16 on Fable 5.2 or $8 on Opus 5.5, before caching, tools, regional multipliers or discounts. Treat that as a budgeting example, not a prediction of how many tokens either model will need.

Use this decision rule:

Workload	Start with
Routine chat, extraction or classification	Haiku or Sonnet
Normal production coding and knowledge work	Opus 5.5
Difficult multi-hour coding or deep research	Fable 5.2
Verified high-risk cyber or biomedical work	Mythos 5.2, if approved
Latency-sensitive work where Fable quality is required	Test Fable 5.2 fast mode

Anthropic reports that Fable 5.2 generates output up to 30% faster than Fable 5.1 and offers an optional fast mode that can increase token generation speed by up to 2.5 times. These are provider-reported results; test latency and total task time in your own environment.^[1]

Minimal Claude API example

import anthropic

client = anthropic.Anthropic()

message = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=4096,
    output_config={"effort": "medium"},
    messages=[
        {
            "role": "user",
            "content": "Review this migration plan, identify the three highest risks, and return a practical test checklist."
        }
    ],
)

text = next(block.text for block in message.content if block.type == "text")
print(text)

Select response blocks by type rather than assuming the first block contains text. Fable 5.2 can return a thinking block before a text block, and thinking tokens count toward max_tokens.^[2][3]

If you use Amazon Bedrock, AWS documents the model ID global.anthropic.claude-fable-5-2 for its example and says the model can be tested from Bedrock → Test → Playground. AWS also notes that Bedrock fallback handling differs from automatic switching: your integration may need to retry a declined request itself.^[4]

Five migration checks that prevent common 400 errors

1. Do not disable thinking

Adaptive thinking is always on. Requests using thinking: {"type": "disabled"} or a manual thinking budget return a 400 error. Control depth and cost with the effort setting instead.^[3]

2. Do not force a tool call

tool_choice values auto and none are supported. Forcing any or a named tool returns a 400 error. Keep tool selection automatic, give a clear instruction, and validate tool inputs in your application.^[2][3]

3. Keep earlier turns unchanged

Thinking blocks are bound to the system prompt, tools and conversation history that preceded them. If your code edits earlier turns and then sends the old thinking block again, the prefix check can reject the request. Keep history append-only or use Anthropic’s documented context-management mechanisms.^[3]

4. Update computer use

The older computer_20251124 declaration can fail on Fable 5.2. Anthropic’s migration guide directs affected integrations to computer_toolset_20260801 and a compatible agent loop.^[3]

5. Leave room for the answer

Thinking and visible text share the max_tokens allowance. A small limit can end the request after thinking but before any visible answer. Increase the limit or lower effort, then measure the effect on quality and cost.^[3]

Effort settings: a practical starting point

Anthropic provides five effort levels. Its prompting guide positions them this way:^[5]

• low: well-specified, high-volume or quick tasks.^[5]
• medium: interactive coding, chat and everyday knowledge work.^[5]
• high: difficult coding, research and long unattended tasks.^[5]
• xhigh: the hardest long-running agent work, where your evaluations prove a gain.^[5]
• max: rare ceiling tests where time and token cost are secondary.^[5]

Do not use xhigh or max because the labels sound better. Anthropic warns that gains above high are uneven and that higher effort can trigger more subagent use, which can materially increase cost.^[5]

A simple evaluation grid

Run the same representative tasks at medium, high and xhigh, then record:

Test	Medium	High	Xhigh
Task success rate
Human review score
Input tokens
Output tokens
Total cost
Time to first useful result
Total completion time
Tool failures or retries

Choose the lowest setting that reliably clears your quality threshold. A benchmark win does not prove that a model is better on your repository, data or workflow.

Production rollout plan

Phase 1: offline test

1. Select 20–50 real tasks with known acceptable results.
2. Remove secrets and customer data unless the environment is approved for them.
3. Run the current model and Fable 5.2 with fixed prompts and tools.^[3][5]
4. Compare task success, reviewer preference, token use, latency and failure modes.^[3][5]
5. Inspect every tool action and code diff, not only the final summary.^[5]

Phase 2: shadow traffic

1. Send a small sample of production requests to Fable 5.2 without using its output live.
2. Log model ID, effort, token use, stop reason, refusal category and tool errors.^[3][5]
3. Check whether preserved-thinking or model-routing changes create failures.^[3]
4. Estimate cost at actual traffic volume.

Phase 3: controlled live rollout

1. Start with a low percentage of eligible tasks.
2. Keep risky and irreversible actions behind human confirmation.^[5]
3. Route ordinary work to a cheaper model unless Fable produces a measured benefit.^[1][2]
4. Define rollback conditions before increasing traffic.
5. Review quality and cost after enough tasks—not after one impressive demo.

Data-retention and safety limitations

Anthropic’s documentation states that Fable 5.2 requires 30-day data retention and is not available under zero-data-retention arrangements unless Anthropic expressly authorizes it. An incompatible workspace can receive a 400 error. Confirm the policy and platform-specific setup before sending sensitive material.^[2][3]

Fable can also decline requests under safety classifiers. Anthropic documents refusal categories including cyber, biology, frontier-model development, reasoning extraction and general harms. Build refusal handling around structured response fields rather than trying to infer a refusal from visible text.^[5]

What Anthropic’s launch benchmarks do—and do not—prove

Anthropic reports gains across its coding, research, computer-use, visual and scientific evaluations. For example, its launch table lists Fable 5.2 at 69.9% on FrontierSWE v2 and 49.1% at xhigh effort on FrontierCode 2.0.^[1]

Those results are useful evidence about the model’s intended strengths, but they are provider-run evaluations. They do not guarantee fewer bugs, lower total cost or better business outcomes in your system. Use them to decide what to test, not to skip testing.

FAQ

What is the Claude Fable 5.2 API model ID?

Use claude-fable-5-2 on the Claude API. Cloud platforms can use platform-specific identifiers; AWS’s Bedrock example uses global.anthropic.claude-fable-5-2.^[2][4]

How much does Claude Fable 5.2 cost?

Standard Claude API pricing is $8 per million input tokens and $40 per million output tokens. Anthropic lists cache reads at $0.20 per million tokens and a 50% Batch API discount on input and output.^[2]

Is Claude Fable 5.2 available on Claude subscriptions?

Anthropic says Fable 5.2 is available across its platforms and announced temporary extra usage for eligible paid plans at launch. Subscription availability and limits can change, so check the current plan screen before relying on a promotional allowance.^[1]

Can anyone use Claude Mythos 5.2?

No. Anthropic limits Mythos 5.2 to organizations verified through programs covering approved cybersecurity or life-sciences uses.^[1][2]

Does Fable 5.2 replace Opus 5.5?

Not for most workloads. Anthropic recommends Opus 5.5 as the starting point for ordinary coding and knowledge tasks, with Fable reserved for harder, more open-ended technical work.^[1]

Can I turn thinking off?

No. Adaptive thinking is always on. Use the effort setting to control depth, latency and token use.^[2][3]

Should I switch production traffic immediately?

No. Change the model ID in a test environment, run your own evaluations, check retention and rate limits, verify tools and fallbacks, then use a gradual rollout.

Sources

1. Introducing Claude Fable 5.2 and Claude Mythos 5.2 | Anthropic
2. Claude Fable 5.2 model overview | Claude Platform Docs
3. Claude Fable 5.2 migration guide | Claude Platform Docs
4. Introducing Claude Fable 5.2 on AWS | AWS
5. Prompting Claude Fable 5.2 | Claude Platform Docs