Claude Fable 5.2: What Changed, Pricing and How to Migrate

Last updated: November 9, 2026

Anthropic shipped Claude Fable 5.2 (claude-fable-5-2) on November 9, 2026, at $8 per million input tokens and $40 per million output tokens, 20 percent below Claude Fable 5.1's $10 and $50, with a 1M token context window and 128k max output tokens. The catch is in the request handling, not the price: four changes from Fable 5.1 can break working code, and the biggest one, the prefix check on every account, can fail code that never failed before.

Everything below comes from the November 9 release notes, What's new in Claude Fable 5.2 and the migration guide, read on launch day. The code was type-checked against anthropic 2.7.0 and @anthropic-ai/sdk 1.7.0, both published the same day. I did not have an API key in this environment, so none of it was run against the live API.

The release in one table#

Fable 5.2 has a 1M token context window (default and maximum), 128k max output tokens and fast mode on the Claude API (speed: "fast", $16 / $80 per MTok, no beta header). Claude Mythos 5.2 (claude-mythos-5-2) has the same capabilities and prices but is only available to organizations verified through Anthropic's programs, such as the Cyber Verification Program. Both models require 30-day data retention and are not available under zero data retention unless Anthropic expressly authorizes it. Batch processing is $4 / $20 per MTok, 5-minute cache writes are $10 and 1-hour cache writes are $16.

Anthropic's own guidance on when to use it is worth quoting: "For most workloads, start with the latest Claude Opus model ... and move to Claude Fable 5.2 when your evals on a Claude Opus model at higher effort still don't meet your requirements." If you are choosing between models rather than upgrading, our Fable 5 effort levels guide covers the effort trade-off that decision rests on.

The four breaking changes#

1. The prefix check now applies to every account#

When you send back a thinking block, the API checks that nothing before it has changed since the model produced it: the system prompt, the tools or an earlier message. On Fable 5.1 a failed check returned a 400 only on accounts created on or after August 31, 2026. On Fable 5.2 and Mythos 5.2 it returns a 400 "whatever the account's age", with a message containing The block is bound to a different conversation.

The docs spell out the trap: "code on Claude Fable 5.1 that edits its history can run without errors on an older account before the switch and fail after it."

You are not affected if Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK builds your requests, or if your own code keeps system and tools fixed for the whole conversation, only appends to messages, and sends each assistant turn back exactly as the API returned it. Otherwise, test first. The migration guide's recommended test is to run a session with the thinking-binding-controls-2026-08-01 beta header and prefix_mismatch_behavior: "drop_block", log input_transformations, and fix every prefix_binding_mismatch entry. In production, leave the setting at its default, "error", if a mismatch can only mean a bug, and treat "drop_block" as a stopgap, because dropping blocks can raise output tokens.

Two newer features exist to help you avoid edits: compaction on demand (beta, compact-2026-09-04) and mid-conversation tool changes (beta, inline-tools-2026-09-15 on the Claude API), which add or remove tools in a role: "system" message instead of editing tools.

%% title: What happens to a thinking block you send back on Claude Fable 5.2
flowchart TD
  A[Request sends back an earlier thinking block] --> B{Was it made by a model Fable 5.2 can read?}
  B -->|No| C[Block dropped, request succeeds, not billed]
  B -->|Yes| D{system, tools and earlier messages unchanged?}
  D -->|Yes| E[Block used, reasoning preserved]
  D -->|No| F{prefix_mismatch_behavior}
  F -->|error, the default| G[400: The block is bound to a different conversation]
  F -->|drop_block, beta header| H[Block dropped, listed in input_transformations]

2. Only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks#

Moving a conversation onto Fable 5.2 keeps its reasoning: on the Claude API and Claude Models on Google Cloud (Google managed), Fable 5.2 reads thinking blocks from Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1, Mythos 5.1 and earlier Opus, Sonnet and Haiku models. Moving the other way loses it. Every model except Mythos 5.2 drops Fable 5.2's blocks, including Opus 5.5, Fable 5.1, and Fable 5.2's own fallback models, Claude Opus 4.8 and Claude Opus 5. The request still succeeds and dropped blocks are not billed, but a router that bounces long agent sessions between Fable 5.2 and a cheaper model will quietly lose the expensive model's reasoning on every bounce.

Fable 5.2 thinking blocks are also bound to the account that produced them, or an account linked to it. Another account's blocks are dropped with reason: "organization_binding_mismatch".

3. computer_20251124 returns a 400 outside Bedrock#

Fable 5.1 accepted both the computer_toolset_20260801 toolset and, with a beta header, the older computer_20251124 tool. Fable 5.2 and Mythos 5.2 return a 400 invalid_request_error for the older tool "on every platform except Amazon Bedrock". Move to the toolset before you switch models.

4. A Fable 5.2 executor needs a Fable 5.2 or Mythos 5.2 advisor#

With the advisor tool (beta), the advisor must be at least as capable as the executor. A Fable 5.2 executor with any other advisor, Fable 5.1 included, gets a 400. Change the advisor's model to claude-fable-5-2 or claude-mythos-5-2 in the same change, and do it in new conversations: changing the advisor's model is an edit to tools, which fails the prefix check above. The reverse is easy: every executor model accepts Fable 5.2 as its advisor, and a Fable 5.2 advisor returns its advice encrypted.

Rules that are not new, but bite if you skip Fable 5.1#

If you are coming from Fable 5.1, Mythos 5.1 or Opus 5.5, these are unchanged. From Opus 5 or older, both are breaking:

• Thinking is always on. thinking: {"type": "disabled"} and {"type": "enabled", "budget_tokens": N} both return a 400. Omit thinking or send {"type": "adaptive"}, and use effort.
• No forced tool use. tool_choice of {"type": "any"} or {"type": "tool", ...} returns a 400. For schema-valid JSON, keep tool_choice: {"type": "auto"} with strict: true, or use JSON outputs. In organizations with customer-managed encryption keys, structured outputs are not available on Fable 5.2 at all.

Coming from Opus 5.5, also note that Fable 5.2's default effort is high where Opus 5.5's is medium, so an unchanged request runs one level higher. Set effort explicitly.

How to switch#

Change the model ID, set effort explicitly, and turn on the audit while you test:

import anthropic

client = anthropic.Anthropic()

# 1. A plain Fable 5.2 request. Set effort explicitly: the default is "high".
reply = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=16000,
    output_config={"effort": "high"},
    messages=[{"role": "user", "content": "Summarize the tradeoffs of event sourcing."}],
)
for block in reply.content:
    if block.type == "text":
        print(block.text)

# 2. Fast mode on the Claude API: no beta header, $16/$80 per MTok.
client.messages.create(
    model="claude-fable-5-2",
    max_tokens=4000,
    speed="fast",
    messages=[{"role": "user", "content": "One-line status update, please."}],
)

# 3. Find history edits before they become 400s: drop failing blocks and log them.
audited = client.beta.messages.create(
    model="claude-fable-5-2",
    max_tokens=16000,
    betas=["thinking-binding-controls-2026-08-01"],
    thinking={"type": "adaptive", "block_binding": {"prefix_mismatch_behavior": "drop_block"}},
    messages=[{"role": "user", "content": "Continue the refactor plan."}],
)
for t in audited.input_transformations or []:
    if t.type == "thinking_dropped":
        print(t.path, t.reason)

Pyright in strict mode reports 0 errors against anthropic 2.7.0, and misspelling "drop_block" as "drop_blocks" fails the check. The same three calls in TypeScript (client.messages.create, speed: "fast", client.beta.messages.create with betas and thinking.block_binding) pass tsc --strict against @anthropic-ai/sdk 1.7.0. In the SDK types, reason on a thinking_dropped entry is one of model_binding_mismatch, prefix_binding_mismatch, organization_binding_mismatch or end_user_binding_mismatch. After a deliberate model switch, model_binding_mismatch entries are expected; prefix_binding_mismatch entries are the edits to fix.

Two practical notes from the migration guide. Thinking tokens count toward max_tokens, so a small max_tokens can stop with stop_reason: "max_tokens" after the thinking block and before any text. And if code reads content[0].text, select blocks by type instead, because thinking comes first.

For fast mode elsewhere: it uses the same speed: "fast" on Claude Models on Google Cloud (both Anthropic and Google managed), while on Microsoft Foundry it is a separate claude-fable-5-2-fast deployment with speed left out. Remove speed: "fast" in Claude Managed Agents. If you are weighing fast mode at all, our Claude Code fast mode breakdown covers when the 2x price pays off.

Behavior to re-check in your prompts#

From the "Behavior differences" section, the items most likely to show up in an existing agent:

• "Answer directly" lines stop working. Fable 5.1 usually follows a line like "answer directly without deliberating". In Anthropic's testing, Fable 5.2 still thought at medium effort and above, and at high it "thinks before almost every reply." Remove those lines and lower effort for easy turns.
• Turns can end early. On long multistep tasks the model sometimes stops with a list of what remains or an offer to continue, more often at low and medium.
• It may review itself with subagents. At xhigh and max in an agentic coding product, Anthropic saw Fable 5.2 start subagents to review its own work, which Fable 5.1 did not do. "Subagents can account for a large part of what a session costs."
• Code ownership in the prompt matters. Without it, the model tended to add to the existing structure and report nearby problems instead of fixing them.

Refusals work as before: HTTP 200 with stop_reason: "refusal" and stop_details. The permitted fallback targets for Fable 5.2 are Opus 4.8 and Opus 5, which, per change 2, cannot read its thinking blocks. If you already handle refusals, our Fable 5 refusal handling guide still applies.

In Claude Code#

Claude Code 2.1.321 adds claude-fable-5-2 as "now the default Fable model". Saved Fable 5.1 picks move to fable: on Pro and Max plans at once, and on Team and Enterprise plans once your organization allows it. To stay on the older model, run /model claude-fable-5-1. Because Claude Code builds its own requests, the prefix check change does not affect it. The same release also fixes fast mode retrying for minutes when fast serving is rate-limited: it now tries standard speed on the same model first. How the previous default switch played out is in our Claude Code Fable 5.1 default post.

Who should move now#

• You use Claude Code, Managed Agents or the Agent SDK: nothing to change beyond re-checking cost, since the price per token drops.
• You build messages yourself and never edit history: switch, set effort explicitly, re-run evals.
• You edit history, route between models, or use the advisor tool or computer_20251124: run the drop_block audit first, fix the edits, then switch. If you skip this on an older account, Fable 5.2 can turn a working integration into a stream of 400s.

For the previous release's API details, see our Fable 5.1 and Mythos 5.1 API guide.

FAQ#

How much does Claude Fable 5.2 cost?#

$8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Claude Fable 5.1. Cache reads are $0.20 per MTok, 5-minute cache writes $10 and 1-hour cache writes $16. Batch is $4/$20, and fast mode on the Claude API is $16/$80.

What breaks when I switch from Fable 5.1 to Fable 5.2?#

Four things: the prefix check returns a 400 on every account if your code edits system, tools or an earlier message and sends back a later thinking block; only Fable 5.2 and Mythos 5.2 read Fable 5.2 thinking blocks; the computer_20251124 tool returns a 400 outside Amazon Bedrock; and a Fable 5.2 executor needs a Fable 5.2 or Mythos 5.2 advisor.

Does Claude Code use Fable 5.2 automatically?#

Claude Code 2.1.321 makes claude-fable-5-2 the default Fable model and moves saved Fable 5.1 picks to fable: at once on Pro and Max plans, and on Team and Enterprise plans once the organization allows it. Run /model claude-fable-5-1 to switch back.

Can I turn thinking off on Claude Fable 5.2?#

No. Adaptive thinking is always on, and thinking: {"type": "disabled"} or {"type": "enabled", "budget_tokens": N} returns a 400 error. Use the effort parameter instead; the default is high.

Is fast mode available for Claude Fable 5.2?#

Yes, in Messages API requests on the Claude API with speed: "fast" and no beta header, at $16 per MTok input and $80 output. It is also on Claude Models on Google Cloud, and on Microsoft Foundry as a separate claude-fable-5-2-fast deployment.

Sources#

• Claude Platform release notes, November 9, 2026
• What's new in Claude Fable 5.2
• Claude Fable 5.2 and Claude Mythos 5.2 migration guide
• Preserved thinking: the prefix check
• Prompting Claude Fable 5.2
• Claude Code CHANGELOG, 2.1.321
• anthropic on PyPI and @anthropic-ai/sdk on npm - the versions type-checked here

Continue Reading#

• Claude Fable 5.1 and Mythos 5.1 API Guide - the release this one replaces
• Claude Code Fable 5.1 Default Model - how the last default switch worked in Claude Code
• Fable 5 Effort Levels Explained - the setting that now controls Fable 5.2's thinking
• Claude PCB Design: Fable 5.2 Built Three Working Prototypes - Anthropic's electronics design research with Fable 5.2
• AI API Pricing Comparison - per-token input and output prices side by side
• Where to Access AI Models (2026) - official APIs, clouds and routers compared