Claude Fable 5.2: pricing, benchmarks and what breaks when you migrate from 5.1
Anthropic launched Claude Fable 5.2 on November 9, 2026. It costs 20% less than 5.1, but four changes can make your existing code fail.
On November 9, 2026, Anthropic released Claude Fable 5.2, its most capable model available to everyone, alongside Claude Mythos 5.2, the same model restricted to verified organizations. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, 20% less than 5.1. Moving to it takes more than swapping the model ID: four changes can make your code return 400 errors.
In sintesi
- Claude Fable 5.2 (
claude-fable-5-2) costs $8 / $40 per million tokens, against $10 / $50 for 5.1. It has a 1M-token context window and 128K max output. - Anthropic reports 45.3% on FrontierCode 2.0 at
higheffort, almost matching Opus 5.5 atmax(45.5%), at about half the cost per task. - Four changes break code: the thinking-block prefix check on every account, thinking blocks readable only by Fable and Mythos 5.2, the
computer_20251124tool rejected, and the advisor pairing. - Fable 5.2 requires 30-day data retention and is not available under zero data retention without Anthropic's express authorization.
- Anthropic itself recommends starting with Opus 5.5 and moving up to Fable only if your evals at higher effort still fall short.
What shipped on November 9
Claude Fable 5.2 and Claude Mythos 5.2 are the same model with two access tiers. The official Anthropic announcement says Fable has strong cyber and biology safeguards and is open to all users, while Mythos is limited to organizations approved through verification programs such as the Cyber Verification Program.
Compared with 5.1, Anthropic claims better coding, reasoning and vision, output up to 30% faster, and about 55% fewer safeguard blocks (about 60% fewer in cybersecurity, 40% fewer in biology). It also claims stronger anti-distillation protections that cut the data harvested by distillers by 90-99%. These are all vendor figures, not independently verified.
The model page in the documentation confirms the technical facts: released November 9, 2026, 1M-token context window, 128K max output, adaptive thinking always on, knowledge cutoff June 2026, retirement not sooner than November 9, 2027. It is available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry.
Pricing: how it compares with 5.1 and Opus 5.5
Fable 5.2 is cheaper than 5.1 and still twice the price of Opus 5.5. The prices come from Anthropic's pricing page and the What's new page.
| Model | Input / MTok | Output / MTok | Cache read / MTok | Batch (input / output) |
|---|---|---|---|---|
| Claude Fable 5.2 | $8 | $40 | $0.20 | $4 / $20 |
| Claude Fable 5.1 | $10 | $50 | $0.25 | n/a |
| Claude Opus 5.5 | $4 | $20 | n/a | n/a |
| Claude Sonnet 5.5 | $2 | $10 | n/a | n/a |
The optional fast mode generates tokens up to 2.5 times faster and costs $16 per million input tokens and $80 per million output tokens. On Mythos 5.2, fast mode is limited to organizations in the verification programs.
A worked example
Take a hypothetical workload of 50 million input tokens and 10 million output tokens per month, with no caching. This is an example we built from the official prices, not a measurement.
| Scenario | Calculation | Monthly cost |
|---|---|---|
| Fable 5.1 | 50 × $10 + 10 × $50 | $1,000 |
| Fable 5.2 | 50 × $8 + 10 × $40 | $800 |
| Fable 5.2 in batch | 50 × $4 + 10 × $20 | $400 |
| Fable 5.2 in fast mode | 50 × $16 + 10 × $80 | $1,600 |
| Opus 5.5 | 50 × $4 + 10 × $20 | $400 |
The 20% saving over 5.1 holds only if token consumption stays the same. Anthropic says that at medium effort 5.2 finishes tasks with fewer tokens than 5.1, so the real saving may be larger, but you have to measure it on your own prompts.
Benchmarks: what Anthropic reports and what to read with care
In the announcement Anthropic compares Fable 5.2 with Fable 5.1, Opus 5.5, GPT-6 Astra and Gemini 4 Carbon. These are vendor-published results. The ones below come from the announcement page.
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| FrontierCode 2.0 (agentic coding, Xhigh) | 49.1% | 32.4% | 45.5% |
| FrontierSWE v2 (agentic coding) | 69.9% | 56.3% | 62.3% |
| AutomationBench (business workflows) | 51.4% | 31.4% | 42.5% |
| Humanity's Last Exam, with tools | 74.2% | 65.6% | 67.7% |
The number that matters most for whoever pays the bill is a different one. According to Anthropic, on FrontierCode at high effort Fable 5.2 scores 45.3%, essentially Opus 5.5 at max (45.5%), at about half the cost per task. On WANDR (agentic search) at medium effort it matches Opus 5.5's best score at about 40% of the cost per task.
The four changes that break code
According to the What's new page, four changes from 5.1 can make requests fail. If you use Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK, you normally need to change nothing beyond the model ID.
- The thinking-block prefix check now applies to every account. The API checks whether anything before a thinking block (the system prompt,
toolsor an earlier message) changed after the model produced it. On Fable 5.2, a request with an invalid block andprefix_mismatch_behaviorunset gets a 400 with the messageThe block is bound to a different conversation, whatever the age of the account. On 5.1 the error only reaches accounts created on or after August 31, 2026, so you can have code that works today and fails after the switch. - Thinking blocks are bound to the model that produced them. Only Fable 5.2 and Mythos 5.2 read Fable 5.2's blocks. If a router or fallback moves the conversation to Opus 5.5, Fable 5.1 or the fallback models (Opus 4.8 and Opus 5), the reasoning from those turns is lost: the API drops the blocks, the request succeeds, and dropped blocks are not billed.
- The
computer_20251124tool is rejected. Any request that declares it gets a 400invalid_request_error, on every platform except Amazon Bedrock. You need thecomputer_toolset_20260801toolset instead. - The advisor must be paired correctly. With the advisor tool (in beta), a Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor. Using 5.1 as the advisor returns a 400.
Two more rules apply to every Fable 5.2 request. Adaptive thinking is always on: thinking: {"type": "enabled"} with budget_tokens and thinking: {"type": "disabled"} both return 400. Forced tool use is also unsupported: tool_choice set to {"type": "any"} or {"type": "tool", "name": "..."} returns 400. To get schema-valid JSON, the documentation points to tool_choice: {"type": "auto"} with strict: true, or to JSON outputs.
How to migrate without surprises
The base change is one line, as in the documentation example:
model = "claude-fable-5-1" # Before
model = "claude-fable-5-2" # After
If your code builds the messages array itself, the risk is in the prefix check. Your code is safe if system and tools stay the same for the whole conversation, if you only append messages without changing, reordering or removing earlier ones, and if you send each assistant turn back exactly as the API returned it.
As a stopgap, Anthropic documents the beta header thinking-binding-controls-2026-08-01 with thinking.block_binding.prefix_mismatch_behavior set to "drop_block": invalid blocks are dropped and the response lists them in an input_transformations array. The documentation warns that dropping blocks can raise output tokens, so treat it as a temporary fix while you correct the history edit.
To keep the history append-only, 5.2 supports compaction on demand (header compact-2026-09-04) and mid-conversation tool changes (header inline-tools-2026-09-15), both in beta.
If you want Claude Code to migrate the project for you, the migration guide suggests the command /claude-api migrate this project to claude-fable-5-2.
Behavior and effort: what changes in your prompts
The documentation lists behavior differences that show up without any code change. The ones that weigh most on cost:
- A prompt line asking the model to answer without deliberating does not stop it from thinking at
mediumeffort and above. Athighand above, the model thinks before almost every reply. For simple requests, use a lower effort level. - On long tasks, a turn can end before the work is done, with a list of what remains. This happens more often at
lowandmedium. - Subagents can account for a large part of a session's cost if your harness invites delegation.
The API default effort is high, and the documentation recommends starting there because most of the capability gain is already present. Move to xhigh only for your longest agentic jobs where you have measured a gain, and keep max for measuring the model's ceiling.
What it means for a company in the EU
For many European companies the heaviest constraint is data retention. Fable 5.2 and Mythos 5.2 require 30-day data retention and are not available under zero data retention unless Anthropic expressly authorizes it. If you process personal or confidential data, this belongs in your GDPR assessment and your record of processing activities before you move a workflow to Fable.
The rest is a model-selection question. Rayo's view: Fable 5.2 makes sense for long agentic tasks where Opus 5.5 at high effort has already failed your evals. As a default model it costs too much. Anthropic gives the same guidance: start with Opus 5.5, move up to Fable or down to Sonnet 5.5 and Haiku 5.5 depending on the task. To keep overall consumption in check you also need the techniques in how to save tokens with Claude Code.
What to do now
- Search your code for anything that modifies
messages,systemortoolsafter a conversation has started, and fix it before the switch. - Check whether you use
computer_20251124, forcedtool_choice,thinkingwithbudget_tokens, or the advisor tool. - Review your routers and fallbacks: every hop to a different model drops Fable 5.2's thinking blocks.
- Run your evals with an effort sweep that starts at
high, and compare cost per task, not just price per token. - Ask your data protection officer whether 30-day retention is compatible with the workflow.
Sources
- Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2 (November 9, 2026)
- Claude docs: Claude Fable 5.2
- Claude docs: What's new in Claude Fable 5.2
- Claude docs: Fable 5.2 migration guide
- Claude docs: Pricing
- Claude docs: Models overview
FAQ
How much does Claude Fable 5.2 cost? It costs $8 per million input tokens and $40 per million output tokens, against $10 and $50 for 5.1. Cache reads cost $0.20 per million tokens and the Batch API halves input and output. Fast mode costs $16 and $80.
Can I use Fable 5.2 with zero data retention? Not by default. Fable 5.2 and Mythos 5.2 require 30-day data retention and are available under zero data retention only with Anthropic's express authorization. If your data is sensitive, talk to your account team before migrating.
What do I need to change when moving from Fable 5.1 to 5.2?
Change the model ID from claude-fable-5-1 to claude-fable-5-2, then run four checks: edits to message history, mid-conversation model switches, the computer_20251124 tool, and the executor and advisor pairing. If you use Claude Code or the Agent SDK, you normally need nothing else.
Is Fable 5.2 better than Opus 5.5? In Anthropic's published benchmarks it is, for example 49.1% against 45.5% on FrontierCode 2.0 at Xhigh effort, but it costs twice as much per token. Anthropic recommends starting with Opus 5.5 and moving to Fable only if your evals at higher effort fall short.
What is the difference between Fable 5.2 and Mythos 5.2? They have the same capabilities, specs and pricing. Mythos 5.2 is available only to verified organizations, for example through the Cyber Verification Program, and has fewer cyber and biology safeguards than Fable.
