After Anthropic shipped Claude Fable 5.2 on November 9, how should Messages API JSON set effort, fast mode, and usage versus Opus 5.5?

As of 10 November 2026: Fable 5.2 lists at $8/$40; fast mode $16/$80. Set output_config.effort and speed; keep Fable 5.1 thinking and tool_choice limits.

Fable 5.2 is not a silent swap of claude-fable-5-1. On 9 November 2026 Anthropic released Claude Fable 5.2 (claude-fable-5-2) and the restricted Claude Mythos 5.2 (claude-mythos-5-2). List price dropped to $8 / $40 per million input and output tokens (Fable 5.1 was $10 / $50). Optional fast mode is $16 / $80. Integrations need three updates: model, output_config.effort (heavy work at xhigh or max), and billing expectations for usage—the same call profile bills differently than Opus 5.5.

This piece is dated 10 November 2026 from the 9 November launch post. A week ago we covered code execution with MCP; in September, Opus 5.5 and tool_choice documented the Fable 5.1 family 400 rules. Here we add Fable 5.2–specific fields and routing, without repeating the full Opus article.

What shipped on 9 November

Anthropic positions Fable 5.2 for the hardest open-ended technical work: agentic coding, long-horizon research, hardware and robotics. Fable and Mythos share weights; Fable keeps full cyber and bio safeguards, while Mythos 5.2 is for verified organizations. Context is 1M tokens; max output is 128k. The launch note claims roughly 30% faster output than 5.1. For daily coding and knowledge work, Anthropic still recommends starting on claude-opus-5-5 and moving up to Fable when scope demands it.

Standard API prices (USD per million tokens): input 8, output 40; Batch is half (4 / 20). Fast mode: input 16, output 80, with up to about 2.5× token generation speed. That is 20% below Fable 5.1 list price, but still 2× Opus 5.5 ($4 / $20) on a per-token basis—route on task depth, not one row on the price sheet.

Model strings by platform

PlatformFable 5.2Mythos 5.2
Claude API / Messagesclaude-fable-5-2claude-mythos-5-2 (verified orgs)
Amazon Bedrockanthropic.claude-fable-5-2Provisioned after verification
Google Cloudclaude-fable-5-2Same
Microsoft Foundryclaude-fable-5-2Same

When you change only the model field, update cache keys and log labels in the router. Do not put a Mythos ID in a default config aimed at all users.

Request JSON: effort and fast mode

Fable 5.2 inherits 5.5-family adaptive thinking: thinking.disabled and budget_tokens still return 400. Depth is output_config.effort: low, medium, high, xhigh, max. The launch table’s 49.1% FrontierCode 2.0 score used xhigh; most agent tasks run fine at medium or high. Omitting effort follows product defaults—do not assume parity with 5.1.

{
  "model": "claude-fable-5-2",
  "max_tokens": 128000,
  "output_config": { "effort": "high" },
  "messages": [{ "role": "user", "content": "..." }]
}

On the Claude API, enable fast mode with beta header fast-mode-2026-02-01 and speed: "fast" inside output_config (same shape as Opus 5.5 fast mode):

{
  "model": "claude-fable-5-2",
  "max_tokens": 32000,
  "output_config": {
    "effort": "medium",
    "speed": "fast"
  },
  "messages": [{ "role": "user", "content": "..." }]
}

tool_choice types any and tool still return 400 on Fable 5.2, matching Opus 5.5 and Fable 5.1. For structured parameters use tool_choice: auto with strict: true JSON Schema, or Structured Output. See Opus 5.5 and tool_choice.

Hard failures carried from Fable 5.1

Moving from 5.1 to 5.2, these JSON shapes are unchanged:

  • Disabling thinking or sending a manual budget → 400; use adaptive plus effort.
  • Forced tool_choice → 400; use auto plus strict Schema.
  • Editing system, tools, or earlier messages then replaying thinking → 400 (default for accounts opened after 2026-08-31).
  • computer_20251124 on Claude API / Google Cloud → 400; use computer_toolset_20260801.

When parsing responses, read text or tool_use by content[].type, not by assuming the first block is the answer—the same rule as our Haiku 5.5 article.

usage in the response and billing JSON

A successful Messages response still carries top-level usage:

{
  "id": "msg_...",
  "type": "message",
  "role": "assistant",
  "content": [ ... ],
  "model": "claude-fable-5-2",
  "stop_reason": "end_turn",
  "usage": {
    "input_tokens": 18240,
    "output_tokens": 4096,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 120000
  }
}

Bill against the price list above; fast mode and Batch have their own multipliers. Fable 5.2 may use fewer tokens than 5.1 at the same effort for the same task, but max output is 128k—budget thinking, tool turns, and final text inside max_tokens. Cost dashboards should store the full usage object, not a single scalar, so cache reads separate from output.

Thinking blocks and cross-read with 5.5

Fable 5.2 can read thinking blocks from Opus 5.5 and earlier Opus, Sonnet, and Haiku; Fable/Mythos 5.2 blocks round-trip within that family, but Opus 5.5 does not read Fable 5.2 blocks. In advisor chains (5.5 consulting Fable), keep block types and signature in the JSON history—swapping model without that loses preserved thinking.

To see which blocks were dropped, send thinking-binding-controls-2026-08-01 and read top-level input_transformations. The protocol matches Opus 5.5; 5.2 did not change it.

Response shape when safeguards fall back

The launch post: Fable 5.2 still blocks high-risk cyber and bio queries; when blocked, users see content from a fallback model (for example Opus 4.8 for cyber, Opus 5 for bio) while your request still names Fable. Logs should record both the request model and, if exposed, the model that actually produced text—do not score fallbacks as Fable quality.

False-positive blocks are claimed down about 55% versus 5.1; a plain 200 with no tools does not prove safeguards stayed off.

When to keep Opus 5.5

Anthropic’s guidance: default daily coding and knowledge work on Opus 5.5; use Fable 5.2 when scope is large, tool-heavy, or long-horizon. Opus is cheaper per token; Fable leads benchmarks like FrontierCode and RLI but costs more. Express routing as config JSON—pick model by task tag instead of one global default.

{
  "routes": [
    { "if": { "task": "daily_coding" }, "model": "claude-opus-5-5" },
    { "if": { "task": "frontier_swe" }, "model": "claude-fable-5-2", "effort": "xhigh" }
  ]
}

JSON to verify before launch

  1. Request body: model, output_config, no disabled thinking or forced tool_choice.
  2. Fast-mode sample: header plus speed: "fast" aligned with $16/$80 pricing.
  3. One live response: full usage; parse content by type.
  4. Router config: production default is not a Mythos ID.

Use our JSON validator, JSONPath, and JSON Diff to compare 5.1 and 5.2 requests locally. Nothing is uploaded.

FAQ

Can I ship by only changing model to claude-fable-5-2?

Not as a silent upgrade. thinking.disabled, budget_tokens, forced tool_choice, or legacy computer tools still 400. Parsers that read content[0] will mis-handle thinking blocks.

Does Fable 5.2 support tool_choice any?

No. Same as Opus 5.5 and Fable 5.1: any and tool types return 400. Use auto with strict JSON Schema or Structured Output.

When should I use Opus 5.5 instead of Fable 5.2?

Official guidance: start daily coding and knowledge work on Opus 5.5; switch to Fable 5.2 for large scope, highest effort, or long-horizon agents. Opus is cheaper on the price sheet.

How do I turn on fast mode in JSON?

Claude API: beta header fast-mode-2026-02-01 and output_config.speed set to fast. Priced at $16/$80 per million tokens, not the standard $8/$40.

How do claude-mythos-5-2 and claude-fable-5-2 requests differ?

Same weight family; Mythos needs verified-org access and looser safeguards. Different model string; otherwise the Messages JSON shape matches. Do not put Mythos in a public default route.

Do safeguard blocks change the response JSON?

HTTP stays 200; content comes from a fallback model. Audit JSON should record request model and actual inference model so billing and evals stay clean.

Summary

Fable 5.2, launched 9 November 2026, changes capability, price, and effort—not the Messages JSON envelope. claude-fable-5-2 lists at $8/$40; fast mode at $16/$80; push effort to xhigh/max for heavy tasks; keep Fable 5.1 thinking and tool_choice limits. Before launch, diff request JSON locally, parse content by type, and store the full usage object.