Anthropic Fable 5.2: what it is, what it costs, when to use it

Official specs and API pricing

Read from our model registry entry for claude-fable-5-2, which was re-checked against Anthropic's model overview, what's-new page and pricing docs. Every rate is USD per million tokens.

Specs and rates verified against Anthropic's model overview, what's-new and pricing pages on 2026-11-11. The global Anthropic pricing baseline we use for the other models on this site was last re-read on 2026-10-08.

Full Fable 5.2 pricing breakdownEvery Claude API rate we track

What Fable 5.2 is actually for

Anthropic positions Fable 5.2 at the top of the stack for work that runs long and stays open-ended — not as a default daily driver.

• Agent sessions that run for hours

  The 1M-token window and always-on thinking are aimed at multi-hour agents that keep a long history and a large tool surface in play. Anthropic's docs name agent sessions that run for hours as the first typical use.

• Multistep deep research

  Research tasks that fan out across sources and come back together. In Anthropic's launch testing, Fable 5.2 worked largely autonomously on two open problems in black-hole physics that had been laid out but never completed.

• Large, ambiguous software work

  Migrations, prototype systems and hard bugs. Anthropic reports a team of agents running an earlier version of Fable 5.2 migrating a critical production service from Python to Rust, and an engineer asking it to design a prototype operating system from an empty folder over eight days.

• Analysis carried through to a deliverable

  Anthropic describes the target as analysis that ends in a finished document, spreadsheet or deck rather than a partial answer — the pattern behind its reported hardware, robotics and formalisation results.

• Vision-heavy and computer-use work

  Input is text and images. On Anthropic's launch table, Fable 5.2 scored 89.3% on OSWorld 2.1 (offline set, partial) and 73.4% on Chartography without tools — both vendor numbers, not independent evaluations.

These are Anthropic's own descriptions of the intended use, not our recommendations. If a task finishes in one turn, a cheaper model is almost always the right call.

What Anthropic says changed from Fable 5.1

Vendor claims from the November 9, 2026 launch post and system card. We have not reproduced them as independent results.

• Up to 30% faster output, at a lower price

  Anthropic states Fable 5.2 generates output up to 30% faster than Fable 5.1 while cutting input and output prices by 20%. An optional fast mode takes token generation up to 2.5x faster again, at double the base rate.

• Fewer unnecessary blocks

  Cyber and biology safeguards are described as equally robust as Fable 5.1's but with about 55% fewer blocks in early testing across claude.ai, Claude Code and Cowork. A blocked request falls back to Opus 4.8 for cyber topics and Opus 5 for biology.

• Fermat's Last Theorem in 30 hours

  A team of 1,000 Fable 5.2 agents formalised the proof in about 30 hours with minimal human intervention, against 11 days for Anthropic's first formalisation by an unreleased research model. With the same 1,000-agent setup, Opus 5.5 completed 26% of the work in 18 hours.

• Three working electronic devices

  From a one-paragraph prompt, Fable 5.2 designed three working devices — a voice-assistant hub, a USB speakerphone and a handheld virtual pet — writing the specs, routing four-layer boards, modelling 3D-printed enclosures in CAD and programming the firmware with Diode. All three worked without electrical modification.

• Robot control it was not trained for

  Anthropic reports substantial gains on physical robot control, including tasks the model was never explicitly trained on. On the public RoboDojo simulation benchmark Fable 5.2 scored 64.9% against 39.4% for Opus 5.5 in the same launch table.

All figures in this section are Anthropic's published launch numbers, quoted as claims. No independent third-party evaluation of Fable 5.2 existed when this page was last checked on 2026-11-11.

Price list, in the shape Anthropic publishes it

USD per million tokens next to the multiple of the base input price. Fast mode and Batch are alternatives to each other, not multipliers you can stack.

Input is 20% and output 20% below Fable 5.1's $10/$50. Cache hits bill 0.025x the base input price — the same multiplier Fable 5.1 used. Against Opus 5.5, input is 2.0x the price, which is the whole trade: Fable 5.2 is the model you move up to, not the one you default to.

There is no long-context surcharge: the full 1M-token window bills at the standard rates. Fast mode rows are exactly 2x the base row, which is how Anthropic publishes them.

Full Fable 5.2 pricing breakdownPrice it on your own token counts

Fast mode: double the price, up to 2.5x the speed

An optional mode that trades money for latency, available per request.

Fast mode speeds token generation by up to 2.5x and bills exactly double: $16.00 per million input tokens and $80.00 per million output tokens. It is available for Fable 5.2 on the Claude API, on Google Cloud (both Anthropic- and Google-managed) and in Microsoft Foundry as a separate model, claude-fable-5-2-fast. It does not combine with the Batch API, and it is not available for Fable 5.1.

Fast mode changes latency, not capability — Anthropic does not claim better answers at the higher price.

Fast mode pricing across modelsEffort levels: quality against price

API rules that will bite you

Fable 5.2 keeps the always-on thinking rules from Fable 5.1 and adds four breaking changes of its own.

• One model ID, one Bedrock variant

  Use claude-fable-5-2 on the Claude API, Claude Platform on AWS, Google Cloud and Microsoft Foundry, and anthropic.claude-fable-5-2 on Amazon Bedrock. Foundry also lists claude-fable-5-2-fast for the fast variant.

• Effort defaults to high

  All five levels are supported — low, medium, high, xhigh, max — and the API default is high, not medium. Anthropic's advice is to start at high, step to xhigh only where you have measured a gain on your hardest long-running work, and reserve max for ceiling measurements.

• Adaptive thinking cannot be switched off

  Sending thinking with type disabled, or type enabled with budget_tokens, returns a 400. Omit the field or send type adaptive, and control depth with effort. temperature, top_p and top_k must be omitted too.

• No forced tool use, and an older computer-use tool is gone

  tool_choice types any and tool return a 400 — use auto with strict tool use or JSON outputs for schema-valid JSON. The computer_20251124 tool also returns a 400 on every platform except Amazon Bedrock; declare the computer_toolset_20260801 toolset instead.

• 30-day retention is mandatory

  Fable 5.2 requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorises it. It is a Covered Model. If your compliance posture forbids retention, this model is out before you compare prices.

• Four breaking changes from Fable 5.1

  The prefix check now applies to every account — an edited history returns 400 'The block is bound to a different conversation'. Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks, so a model switch loses the reasoning. Computer use moves to the toolset. And a Fable 5.2 executor needs a Fable 5.2 or Mythos 5.2 advisor.

Summarised from Anthropic's what's-new page for Fable 5.2 and its migration guide, both read on 2026-11-11. The full rule set — including prefix_mismatch_behavior and the thinking-binding-controls beta — is in those docs.

Fable 5.2 vs Opus 5.5, row by rowFable 5.1 vs Fable 5: what changed last generation

Benchmarks: Anthropic's launch numbers

Two rows from the vendor's launch table, quoted with attribution. The full table, including the rows that compare Mythos against Opus without safeguards, is on the comparison page.

On agentic coding (FrontierCode 2.0, highest available effort) Anthropic reports 49.1% for Fable 5.2 against 45.5% for Opus 5.5. On the Remote Labor Index it reports 41.3% against 21.3% — the largest gap in the table, and the one that best matches Anthropic's long-horizon framing. Anthropic also states that at Xhigh effort Fable 5.2 scored the highest of any released model it tested.

These are vendor-reported figures with safeguards on and each model's highest available effort. They are not an independent evaluation, and we have not reproduced them. Treat them as Anthropic's claims, and run your own evals before switching.

Fable 5.2 vs Opus 5.5, row by rowFable 5 benchmarks (previous generation)

Where it sits in the lineup

Fable is now a tier above Opus, not a replacement for it.

Anthropic's guidance is explicit: start on Claude Opus 5.5, the strongest daily driver, and move up to Fable 5.2 only when Opus at higher effort still misses your bar — or move down to Sonnet 5.5 or Haiku 5.5 for cheaper work. Fable 5.2 can also serve as an advisor that smaller models consult, or as an orchestrator that delegates simpler subtasks. Mythos 5.2 is the same model with the cyber and biology safeguards removed, available only to organisations verified through Anthropic's Cyber or Life Sciences programmes.

Claude Opus 5.5: specs and pricingFable 5.2 vs Opus 5.5, row by row

Keep reading