Claude Fable 5.2 API

Price check: official, OpenRouter and Wokey

USD per million tokens. Official and OpenRouter rates were checked by hand on 2026-11-10, excluding tax; the Wokey rate is read from the live price list.

Checked 2026-11-10 · Official pricing page · OpenRouter endpoint list

Use it in Claude Code

This model supports /v1/messages on Wokey. Set these variables, then start claude.

export ANTHROPIC_BASE_URL="https://api.wokey.ai"
export ANTHROPIC_AUTH_TOKEN="$WOKEY_API_KEY"
export ANTHROPIC_MODEL="claude-fable-5-2"

Claude Fable 5.2 at a glance

Release date

2026-11-09, alongside Mythos 5.2; the two are the same model, and Fable 5.2 carries cyber and biology safeguards and is open to every customer

Positioning

For the hardest open-ended technical work and long-horizon agentic tasks. Anthropic still says to start everyday coding and knowledge work on Opus 5.5 and move up when your evals fall short

Context and output

1M context (default and maximum); 128K max output; output up to 30% faster than Fable 5.1, with latency now listed as “moderate” instead of “slower”

Input and cutoff

Text and image input; knowledge cutoff June 2026; the tokenizer introduced with Opus 4.7

Thinking and effort

Adaptive thinking is always on and cannot be disabled; five effort levels (low to max), defaulting to high on the API and medium in Claude Code and the Claude apps

Official price

$8 input / $40 output per million tokens, 20% below Fable 5.1’s $10 / $50; cache read $0.20 (2.5% of input); 5-minute cache write $10, 1-hour $16; Batch $4 / $20; fast mode $16 / $80 for up to 2.5x output speed

Safeguards and data

About 55% fewer safeguard blocks than Fable 5.1 (cyber about 60%, biology about 40%); blocked cyber requests fall back to Opus 4.8 and biology requests to Opus 5; 30-day retention, and zero data retention (ZDR) only with authorization

Launch demos

An Anthropic engineer had Fable 5.2 build an operating system from an empty folder: it coordinated over 1,000 subagents for eight days and produced a prototype that boots on a laptop; 1,000 Fable 5.2 agents fully formalized the proof of Fermat’s Last Theorem in 30 hours, where Opus 5.5 in the same setup had done 26% after 18 hours

Model ID

claude-fable-5-2 (anthropic.claude-fable-5-2 on Bedrock; the same ID on Google Cloud, Microsoft Foundry and Claude Platform on AWS)

Checked 2026-11-10 · Sources: Anthropic · What’s new · Migration guide

Official benchmarks: Fable 5.2 vs Fable 5.1 vs Opus 5.5 vs GPT-6 Astra vs Gemini 4 Carbon

From Anthropic’s Fable 5.2 launch page; these are vendor-reported results, and the GPT-6 Astra and Gemini 4 Carbon scores are as listed by Anthropic. Each model runs at its highest effort, with safeguards on for Fable and Opus. Several benchmarks use different versions or settings from the Opus 5.5 launch table (OSWorld offline partial, Chartography without tools), so the two tables are not merged; “—” means no published score.

Effort exceptions: on FrontierCode 2.0 Fable 5.2 and Fable 5.1 ran at xhigh and Gemini 4 Carbon at medium; on Remote Labor Index Fable 5.2 ran at high. Mythos 5.2 scored 75.3% on Terminal-Bench-Science. RoboDojo Sim was run internally by Anthropic. The launch page also lists three cyber and biology benchmarks that report Mythos 5.2, Mythos 5.1 and Opus 5.5 without safeguards rather than Fable 5.2, so they are left out. Fable 5.2 is not first only on AA-Briefcase Elo (1785 vs Opus 5.5’s 1800) and AutomationBench (51.4% vs Gemini 4 Carbon’s 53.9%).

Which to use: Fable 5.2, Opus 5.5 or Fable 5.1

vs Opus 5.5

Fable 5.2 lists at twice Opus 5.5’s $4 / $20, with the same $0.20 cache read. It leads eleven of the twelve rows above, such as 49.1% vs 45.5% on FrontierCode 2.0 and 41.3% vs 21.3% on Remote Labor Index. On Artificial Analysis Intelligence Index v5.0.1, Fable 5.2 scores 65 at max for $26.54 per task and Opus 5.5 scores 59 for $10.04. Anthropic still calls Opus 5.5 its daily driver, but also published a lower-effort comparison: Fable 5.2 at high scores 45.3% on FrontierCode, close to Opus 5.5’s best of 45.5%, at about half the cost. Compare cost at the same score, not at the same effort.

vs Fable 5.1

It is 20% cheaper per token, keeps cache reads at 2.5% of input, outputs up to 30% faster and triggers about 55% fewer safeguard blocks, so there is no reason to stay on Fable 5.1. Scores rose almost everywhere: FrontierCode 2.0 went from 32.4% to 49.1% and Remote Labor Index from 17.9% to 41.3%. At max it uses more tokens, though: Artificial Analysis measured 275M output tokens and $26.54 per task, against 163M and $12.99 for Fable 5.1. Migration is mostly a model-ID change; check the four breaking changes below.

Choosing effort

The default is high on the API and medium in Claude Code. Anthropic says most of the capability gain is already there at high; keep xhigh for the hardest long-running coding and agentic work and max for measuring the ceiling. On Artificial Analysis, Fable 5.2 scores 62 at both xhigh and high and 65 at max. At medium and above, a prompt line like “answer directly without deliberating” no longer stops thinking, so lower the effort to save tokens. At xhigh and max it may start subagents to review its own work, which raises cost noticeably. One of the most shared launch-day posts on X was simply “Don't use Fable 5.2 Max.”

Fable 5.2 plans, Opus 5.5 executes

The most-liked user post on X about the launch (Matt Shumer, about 2,800 likes) sums up the pattern: Fable 5.2 for planning, Opus 5.5 for execution. That matches Anthropic’s own positioning: start everyday work on Opus 5.5, and use Fable 5.2 as an advisor that smaller models consult or an orchestrator that hands them tasks. In Claude Code, plan with /model fable and switch back to Opus 5.5 to implement, so the higher Fable price goes to the hardest step only.

Effort documentation

Third-party and community reviews

• Artificial Analysis: On Intelligence Index v5.0.1 it scores 65 at max, #1 of 74 models: 6 points above Opus 5.5 (59 at max), 8 above GPT-6 Astra (57) and 16 above Fable 5.1 (49); xhigh and high both score 62. The cost is money and time: a max task costs $26.54, the most of the 73 models measured and about 2.6x Opus 5.5 at max and 2x Fable 5.1, with 275M output tokens against a 72.8M median, about 44 minutes per task and roughly 103 tokens per second. Every Wokey meter is 24% of the official rate, so the same max task comes to about $6.34.
• Early-access customers (Anthropic launch page): Vercel: work that took weeks with other models finished in a few days. Oracle: SQL from two other models returned 5–8x the expected values, while Fable 5.2’s was correct. Snowflake: the strongest model on its 103-task data-eng-bench. Jump Trading: a full 10% lower error than the next best model. Fundamental Research Labs: 91% on its eval against 86% for the next best. These are customer statements quoted by the vendor.
• Hacker News (790-point thread): Most of the thread is not about Fable 5.2 but about whether Opus 5.5 was degraded before the launch; one commenter re-ran 20 private tasks on Opus 5.5 and called the changes noise. One reply, without a source, lists max-effort cost per task as $11.46 for Fable 5.1 and $44.70 for Fable 5.2, and another commenter felt the gap only pays off for design and review. Many questioned the claim that 1,000 agents formalized Fermat’s Last Theorem in 30 hours as unverified.
• ARC Prize and Cursor: ARC Prize verified 81.2% on ARC-AGI-3 in the standard harness, a new high there, at a cost of $21.9K, and 93.8% on ARC-AGI-2 at $0.98 per task. On Cursor’s CursorBench it tops the board at 60% at max effort and scores 56.9% at the default, up from 49.2% for Fable 5.1.
• Epoch AI: A record 174 on the Epoch Capabilities Index (ECI), above the previous high of 169 from GPT-6 Astra; Epoch says the score is above what the pre-Astra trend for reasoning models would have predicted. It comes from pre-launch testing on Epoch’s own benchmarks only and may change as more results come in.
• Hands-on tests on X: Launch-day tests split two ways. At top effort it usually costs more: Theo spent 76 minutes and about $42 on one build and found it not significantly better than Opus 5.5, especially for the price; a Chinese user running the same prompt at top effort on both got 15 minutes and $8.2 from Opus 5.5 versus 60 minutes and $18 from Fable 5.2; and on a Claude Code bug fix, Fable 5.2 used about twice Fable 5.1’s tokens, $4.40 a run against $3.28. At lower effort some report savings: one developer compared it with Opus 5.5 on 17 coding tasks and 68 scored runs and found the same quality at about half the API cost, and on another team’s fault-diagnosis task Fable 5.2 at medium matched Opus 5.5 at high with about a third of the tokens and half the time.

Migrating from Fable 5.1 or Opus 5.5: requests that return 400

These changes come from Anthropic’s What’s new page and migration guide. Moving from Fable 5.1 is mostly a model-ID change, but the four breaking changes below can make requests that used to work fail.

• The prefix check applies to every account: if the system prompt, tools or earlier messages before a returned thinking block have changed, the request returns 400 “Invalid `signature` in `thinking` block. The block is bound to a different conversation.”, and retrying the same body does not clear it. Fable 5.1 only checks accounts created on or after 2026-08-31, so code that worked on an older account can fail on Fable 5.2. Set thinking.block_binding.prefix_mismatch_behavior: "drop_block" (with the thinking-binding-controls-2026-08-01 beta header) to drop the blocks instead. Nothing changes if Claude Code or the Agent SDK manages your history.
• Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks: when a conversation moves from Fable 5.2 to Fable 5.1, Opus 5.5 or a fallback model, the blocks are dropped silently and not billed, and later turns run without that reasoning; moving onto Fable 5.2 keeps it. The blocks are also bound to the account that produced them, so replaying them through another account drops them too.
• The earlier computer use tool computer_20251124 returns 400 everywhere except Amazon Bedrock; declare computer_toolset_20260801 instead. With the advisor tool, a Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor, and even Fable 5.1 returns 400.
• From Opus 5.5: the default effort moves from medium to high, so requests that omit it spend more tokens; rate limits are separate and lower; organizations under zero data retention get 400 on every request; organizations with customer-managed encryption keys (CMEK) lose structured outputs.
• Limits carried over from Fable 5.1 still apply: thinking: {"type": "disabled"} or {"type": "enabled", "budget_tokens": N}, tool_choice of any or tool, and non-default temperature, top_p or top_k all return 400. Through Wokey the gateway strips the incompatible thinking fields before forwarding, so the request does not fail, but thinking still runs and bills as output tokens.
• Behavior changes: at medium and above a prompt asking for less thinking no longer works, and a small max_tokens can stop with max_tokens before any text; at low and medium a long task can end early with a list of remaining work; there may be little text between consecutive tool calls.
• Refusals return stop_reason: "refusal" with stop_details, and fallbacks: "default" (beta) retries on Opus 4.8 or Opus 5; the fallback model does not receive Fable 5.2’s thinking blocks.

Request body after migration (/v1/messages)

{
  "model": "claude-fable-5-2",
  "max_tokens": 64000,
  "thinking": { "type": "adaptive", "display": "summarized" },
  "output_config": { "effort": "high" },
  "messages": [{ "role": "user", "content": "Find and fix the flaky test in this repo" }]
}

Catalog pricing snapshot

Prices are per 1M tokens; request-time pricing controls settlement.

Official price source: https://platform.claude.com/docs/en/models/fable-5-2/overview

Volume pricing: pay $6,000 / $10,000 on every Claude model in a day and every Claude model is 10% / 20% off the next day — automatic, with no cap. See the rules

Model capabilities

Context

1,000,000 tokens

Max output

128,000 tokens

Streaming

Supported

Tools

Supported

Vision

Supported

Supported API forms

/v1/messages

Send a request

/v1/messages

curl https://api.wokey.ai/v1/messages \
  -H "x-api-key: $WOKEY_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-fable-5-2","max_tokens":1024,"messages":[{"role":"user","content":"Hello"}]}'

Use Claude Fable 5.2 in your tools

The gateway is https://api.wokey.ai (OpenAI-compatible clients usually take https://api.wokey.ai/v1), and the model ID is claude-fable-5-2. Open the guide for the client you use.

• Use Claude Fable 5.2 in Claude Code: Anthropic Messages · base URL and API key setup
• Use Claude Fable 5.2 in Claude Desktop: Anthropic Messages · base URL and API key setup
• Use Claude Fable 5.2 in pi: Anthropic Messages · base URL and API key setup
• Use Claude Fable 5.2 in Z Code: Anthropic Messages · base URL and API key setup
• Claude API cost calculator: what Claude Fable 5.2 costs per month: Compare monthly cost on the Anthropic API, OpenRouter and Wokey for your usage and cache hit rate.
• OpenRouter alternative: Wokey vs OpenRouter: Compare token prices, payment fees, API forms, and upstream sources.
• Verifiable AI API: check a Claude Fable 5.2 response came from the official upstream: Which responses carry a tee.proof, and what a proof does and does not show.
• Verify a Claude Fable 5.2 response online: AI API relay verifier: Paste a response with its tee.proof and check it locally in your browser; nothing is uploaded.

Example production call

These usage values come from a completed Claude Fable 5.2 call that succeeded on its first attempt and was reconciled with its bill.

Total input

291883

Output

2486

Cached share of input

99.3%

Call duration

27.91 s

Usage field example

Compiled from this call’s billing record. input_tokens is uncached input; cache reads and writes are separate.

{
  "model": "claude-fable-5-2",
  "usage": {
    "input_tokens": 2,
    "output_tokens": 2486,
    "cache_read_input_tokens": 289984,
    "cache_creation_input_tokens": 1897
  }
}

Token accounting

These are the final per-million-token rates recorded for this call.

Billed amount

$0.042165

Cost = sum of tokens × their rate ÷ 1,000,000, rounded to six decimals after summing.

Source of this call

The transport record points to api.anthropic.com with OAuth authorization. Usage was metered upstream and the request returned HTTP 200.

Verify Prompt Cache

1. Keep the system, tools, and conversation prefix fixed and mark cache_control according to the cache rules.
2. Check cache_creation_input_tokens on the first response, then reuse the prefix and check cache_read_input_tokens on the next.
3. Reconcile uncached input, cache reads, cache writes, and output separately. Total input is the sum of the three input buckets.

Frequently asked questions

Related models and documentation

• Claude Fable 5
• Claude Fable 5.1
• Claude Opus 5
• Claude Opus 4.8
• API documentation
• All models and pricing
• Provider Node