How Much To Run AI
Hosted API

Claude Fable 5.2: Pricing, API and Flagship Comparisons

Claude Fable 5.2 is Anthropic’s November 9, 2026 model for difficult reasoning and long-running technical work. It accepts text and images and produces text, with a 1M-token context window. Standard API input/output costs $8/$40 per million tokens; Fast mode doubles those rates. Start with Opus 5.5 for everyday work, and consider Fable when a harder task justifies deeper reasoning or a different effort setting.

Content reviewed:

Fixed-token request costs: Fable $1.20, Opus $0.60, Astra short-context $1.50 and Argon $0.30 through November 30, 2026
Calculated USD token costs for 100K uncached input and 10K output tokens. Argon’s offer ends November 30, 2026; its example then costs $0.60.

Use this model

API docs and console

When is Claude Fable 5.2 worth using?

  • Resolve a difficult repository issue after a cheaper model’s patch fails the acceptance tests.
  • Carry a research question through tool use, source checking and a finished technical deliverable.
  • Review complex diagrams, charts and screenshots alongside the code or documents they describe.
  • Advise a less expensive coding agent at a difficult decision point instead of running every step on Fable.

How to use Claude Fable 5.2

Use claude-fable-5-2 through the Claude Messages API, or select Fable in Claude Code and the Claude apps where your plan provides access. The model guide also lists Amazon Bedrock, Google Cloud and Microsoft Foundry; each platform has its own account and quota setup. API usage and a subscription’s included allowance are separate budgets.

For a first API request, ask for a bounded analysis before adding a tool loop. This example requests a migration checklist at medium effort. The output limit includes thinking as well as response text, so inspect the stop reason and usage before deciding that a response is complete.

GitHub lists Fable 5.2 for Copilot Pro+, Max, Business and Enterprise, with a gradual rollout across supported clients. Business and Enterprise administrators must enable the model policy. Copilot bills it at provider list pricing under usage-based billing; access through a Copilot plan does not make API usage unlimited. The announcement also describes default safety-classifier data retention and a separate eligible-enterprise zero-retention route.

A first Claude API request (requires your own API credential)
curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-fable-5-2",
    "max_tokens": 4096,
    "output_config": {"effort": "medium"},
    "messages": [{"role": "user", "content": "Plan a migration from a synchronous job runner to a durable queue. List the failure cases and the tests to run before switching traffic."}]
  }'

Claude Fable 5.2 API pricing and request costs

The following USD prices are for the Claude API, checked November 10, 2026. At standard speed, 100,000 uncached input tokens and 10,000 billed output tokens cost $0.80 + $0.40 = $1.20. With the same usage, Fast mode costs $2.40 and asynchronous Batch costs $0.60. These are token-budget examples; tools, additional turns and retries add to the session bill.

Caching separates prefix writes and reads from ordinary input. A five-minute write of a 100,000-token prefix costs $1; a subsequent standard-speed read costs $0.02. Add new uncached input and output for each call. A one-hour write of the same prefix costs $1.60. Cache savings depend on reusing the prefix before it expires.

Claude API operationUSD per 1M tokens
Standard uncached input / output$8 / $40
5-minute / 1-hour cache write$10 / $16
Standard cache read$0.20
Batch input / output$4 / $20
Fast input / output$16 / $80

Does Fast mode make the whole task 2.5 times faster?

Fast mode uses the same model with a faster inference configuration. Anthropic advertises up to 2.5 times higher output-token speed, rather than a matching reduction in time to first token or end-to-end task duration. Database calls, test runs and human review still take time.

On the Claude API, add speed: "fast" to the request. Fast mode has its own rate limits and premium pricing; caching and data-residency modifiers apply on top. Microsoft Foundry instead uses a separate claude-fable-5-2-fast deployment. Choose it when waiting for generation is expensive enough to justify the extra token spend.

Choose effort before comparing speed or cost

The Claude API defaults to high effort; Claude Code and Claude apps default to medium. Set effort explicitly when comparing runs. Adaptive thinking is always on, and higher effort can spend more tokens or launch additional subagents in a coding harness. Account for those calls when budgeting a session.

When migrating from Opus 5.5, omitting effort silently changes the API default from medium to high. Remove unsupported sampling controls and forced tool-choice settings rather than copying the old request unchanged. A small output limit can be consumed by thinking before a text answer appears; adjust the limit or lower effort for the task.

What do the Fable 5.2 benchmarks show?

Artificial Analysis lists Fable 5.2 at 65 on Intelligence Index v5.0.1 for its Max, Default Fallback configuration. The same page reports $26.54 per index task and 46.4 minutes per task. Those are evaluation averages for that configuration, not a quote for a coding issue or a latency promise for your application.

Anthropic’s launch tests show improvements across coding, reasoning and vision, and report prerelease formalization of the proof of Fermat’s Last Theorem in 30 hours. The reported task formalized an existing proof.

Claude Fable 5.2 vs other flagship models

The table holds token volumes fixed: 100,000 uncached text input tokens and 10,000 billed output tokens, standard processing, without tools or cache. It isolates the price difference. Real models use different numbers of tokens to complete the same task, especially at high reasoning effort.

Model / rate periodInput / output per 1MExample request
Claude Fable 5.2$8 / $40$1.20
Claude Opus 5.5$4 / $20$0.60
Gemini 4 Hydrogen$1.50 / $7.50$0.225
GPT-6 Astra (short context)$10 / $50$1.50
Gemini 4 Argon through November 30, 2026$2 / $10$0.30
Gemini 4 Argon from December 1, 2026$4 / $20$0.60

Claude Fable 5.2 vs Claude Opus 5.5

Opus costs half as much per standard input and output token. Fable is worth testing when Opus cannot finish a hard task reliably, or when a lower Fable effort reaches the required result with fewer tokens. Anthropic still recommends Opus as the everyday starting model.

In Anthropic’s FrontierCode 2.0 evaluation, Fable at high scored 45.3% against Opus’s best 45.5%, at roughly half the cost per task. Fable at xhigh reached 49.1%. Test effort settings on your task: a higher token rate can still produce a lower bill if the model uses fewer tokens to reach the required result.

Claude Fable 5.2 vs GPT-6 Astra

The available OpenAI model is gpt-6-astra. Its standard short-context rates are $10/$50, making Fable 20% cheaper at identical uncached token volumes. Astra prompts above 272K input tokens apply doubled input/cache rates and 1.5 times output rates to the full request. Keep the prompt length with any cost comparison.

Anthropic reports Fable high ahead of Astra’s best FrontierCode score, 45.3% versus 42.8%, at about the same task cost. Across its broader suite, Artificial Analysis reports higher task costs for Fable: Astra max scores 57 at $6.24 and 23.0 minutes per task, while Fable’s Max, Default Fallback configuration scores 65 at $26.54 and 46.4 minutes. Choose the extra capability when it changes the accepted result; do not carry the FrontierCode cost claim across unrelated tasks.

Claude Fable 5.2 vs Gemini 4 Argon

Argon is an alternative for long engineering, research and professional workflows. Its introductory $2/$10 rates run through November 30, then rise to $4/$20 on December 1. Google bills thinking output and offers implicit rather than explicit context caching for Argon. The fourfold token-price difference from Fable during the offer can matter for repeated long tasks.

Artificial Analysis currently lists Argon high at 50, $9.95 per task and 29.2 minutes per task, using $4/$20 rates. Fable’s listed configuration scores higher but also spends more per task. That AA cost does not use Argon’s introductory quote, so keep the benchmark accounting separate from the price table above.

Claude Fable 5.2 vs Gemini 4 Hydrogen

Hydrogen is the lower-priced option for everyday coding and repeated tool calls. At the fixed token volumes above, its Standard API bill is $0.225 against Fable’s $1.20. That difference isolates token rates; it does not account for different reasoning usage, retries or whether the model finishes the task.

Use Hydrogen for a bounded implementation trial and compare Fable when the task needs deeper reasoning or the cheaper run fails its acceptance tests. Keep the same repository revision and tool environment, and count all attempts before choosing the default model for an agent.

Claude Fable 5.2 vs Gemini 4 Carbon

Carbon is the Gemini model in Anthropic’s launch comparison. Fable leads that table on FrontierCode 2.0, 49.1% at xhigh versus Carbon’s 41.1% at medium, while Carbon leads AutomationBench, 53.9% versus 51.4%. These are publisher-reported tests with different model settings. They support a task-specific trial rather than a claim that one flagship wins every category.

A coding-task comparison outside the launch benchmarks

Atomic Agent’s developers tested a 1,400-line Python stock library with planted bugs and a missing feature. Their prerelease harness used OpenRouter pinned to Anthropic, fresh project copies and three runs per setting. All four configurations passed 59 hidden tests, so this task did not separate the models on quality.

At medium effort, Opus averaged 4.2 minutes and $0.97; Fable averaged 4.2 minutes and $1.87. At high, Opus averaged 9.6 minutes and $1.70, while Fable took 26.6 minutes and $7.70. Timing had a scheduling confound: Opus medium ran sequentially earlier, while other runs overlapped. Fast mode was not tested. These developer-reported results show why the effort setting deserves its own trial before switching an agent’s default.

Fable 5.2 or Mythos: which access route fits?

Fable and Mythos 5.2 share an underlying model and token prices. Fable is the general-access option with safeguards; Mythos serves approved cybersecurity and biomedical work through verification programs. A higher score in a Mythos-specific security or biology test does not describe the general-access Fable configuration. Use the program that covers the authorized task.

Before you use it

  • Fable 5.2 is a hosted, closed-weight model; there is no downloadable checkpoint to size for a local GPU.
  • The model requires 30-day data retention unless Anthropic expressly authorizes an exception. Check this before submitting a restricted workload.
  • A fallback response can come from a different permitted Claude model. Keep the actual model, stop reason and billed usage in your task records.

How to get started

Try one difficult task with a fixed acceptance check. Start at medium effort, compare Opus on the same task, and increase effort only when the result needs it. Record accepted outputs, tool and subagent calls, elapsed time and the full bill before choosing a default.

Sources

Sources (15)

Calculators