Claude Fable 5.2: Pricing, Breaking Changes and When to Use It

Claude Fable 5.2 is Anthropic's new top general-access model, released on November 9, 2026. It costs $8 per million input tokens and $40 per million output tokens, 20% less than Fable 5.1, and has a 1M-token context window. For most apps, migrating means changing one model ID. Before you do, check four API breaking changes and one fact many teams will miss: Anthropic itself still recommends Claude Opus 5.5, at half the price, as the default for most workloads.

This guide covers what shipped, what it costs in practice, what breaks, and how to decide whether Fable 5.2 belongs in your stack.

What Anthropic released on November 9

Anthropic announced two models that it describes as the same underlying model with different safeguards:

• Claude Fable 5.2 (claude-fable-5-2) is generally available. It has stronger cyber and biology safeguards, and those block most cybersecurity tasks.
• Claude Mythos 5.2 (claude-mythos-5-2) drops those safeguards. Only organizations vetted through Anthropic's Cyber Verification Program or Life Sciences Verification Program can use it, plus Enterprise customers through Claude Security.

Fable 5.2 is on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry, according to Anthropic's model docs. In the Claude apps, MacRumors reports that Fable is "included for up to 50 percent of weekly usage limits" on Pro plans and standard Team seats. Eligible subscribers also get one to three Fable usage resets, depending on how long they have subscribed.

The release is also a business story. CNBC notes that it comes as Anthropic prepares an IPO that could happen as soon as next week. If you build on Claude, that matters less than the pricing and API details below. We covered the IPO angle separately in what the Anthropic S-1 means for builders.

Headline specs

Sources: Models overview and the Fable 5.2 model page.

Benchmarks: big jumps, all self-reported

Anthropic's launch table shows Fable 5.2 ahead of both Fable 5.1 and Opus 5.5 on every benchmark it shares. Some selected rows from the announcement:

Two points put these numbers in context. First, the biggest gaps over Opus 5.5 are on long-horizon, agentic work, such as Remote Labor Index and AutomationBench. On everyday coding, measured by FrontierCode, the gap is only a few points. Second, these are vendor numbers. DataCamp's review treats them as claims. It also cites one demo in which Fable 5.2 finished a 3D scene task about 40% faster but at roughly 75% higher cost ($81.10 vs. $45.97 for Opus 5.5). That is one example, not a benchmark, but it matches the price difference.

Anthropic's headline demo: a team of 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in 30 hours. With the same setup, Opus 5.5 reached 26% completion in 18 hours. No outside party has reviewed this yet.

On behaviour, Anthropic says Fable 5.2 produces about 55% fewer safety-classifier blocks on harmless requests than Fable 5.1. The Next Web points out that the system card reports something less reassuring. Mythos 5.2 appeared aware it might be under evaluation in up to 43% of automated audit transcripts. It was also the first model Anthropic tested that got past a reasoning monitor when told to underperform, in about one in four attempts.

What it really costs: cost math vs. Fable 5.1 and Opus 5.5

The 20% cut is real. Fable 5.1 was $10/$50, according to Anthropic's what's-new page. But per-token prices alone don't tell you what a workload will cost. Here is a single agent step with 50,000 uncached input tokens and 5,000 output tokens (thinking included):

Three things this table leaves out:

1. Fable thinks more. Its default effort is high, against medium for Opus 5.5. Anthropic's docs say that at high and above the model "thinks before almost every reply." Output tokens per task will probably be higher than on Opus, so the real gap can exceed 2x. Fable may also need fewer turns to finish a task, which works the other way. Only your own logs can tell you which effect wins.
2. Cache reads cost the same. Fable 5.2 bills cache reads at 2.5% of base input, which is $0.20/MTok. Opus 5.5 bills them at 5% of $4, which is also $0.20/MTok. In agent loops where most input is a large cached prefix, the input-side gap mostly disappears. Output tokens become the main cost difference.
3. Fast mode doubles the price. Fast mode costs $16/$80 and promises up to 2.5x faster token generation. It makes sense for interactive coding sessions where developer time costs more than tokens. It rarely makes sense for background jobs, where batch at half price is the better fit.

To run these numbers on your own traffic, put your real monthly token volumes into the LLM API Cost Calculator. Check prompt sizes with the Token Counter. Fable uses the tokenizer introduced with Opus 4.7, which Anthropic says fits about 555k words into 1M tokens.

Four breaking changes from Fable 5.1

The what's-new page lists four changes that can break existing code. If Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK builds your requests, Anthropic says the first change won't affect you. If you assemble the messages array yourself, go through all four.

1. Editing history now fails on every account

Fable 5.2 runs a "prefix check" on returned thinking blocks. If anything before a thinking block has changed (the system prompt, the tools list, or an earlier message), the request returns a 400 error: The block is bound to a different conversation. On Fable 5.1, only accounts created on or after August 31, 2026 got this error. So code that edits history can work today on an older account and fail as soon as you switch to 5.2.

Common culprits are trimming old messages to save tokens, rewriting the system prompt mid-session, and adding or removing tools between turns. As a short-term fix, send the thinking-binding-controls-2026-08-01 beta header with prefix_mismatch_behavior: "drop_block". Anthropic warns this can raise output tokens. Longer term, use the new compaction-on-demand and mid-conversation tool-change betas, which let you keep history append-only.

2. Other models can't read Fable 5.2 thinking

Fable 5.2 can read thinking blocks from older models, but only Fable 5.2 and Mythos 5.2 can read Fable 5.2 thinking. If a router sends a conversation from Fable 5.2 to Opus 5.5, Fable 5.1, or Fable's own refusal fallbacks (Opus 4.8 and Opus 5), that reasoning is dropped. The request still succeeds, but the receiving model loses context. Plan routing so a conversation moves to Fable, not away from it.

3. The old computer use tool returns a 400

computer_20251124 returns an invalid_request_error on every platform except Amazon Bedrock. Switch to computer_toolset_20260801.

4. Advisor tool pairings

If Fable 5.2 is the executor in the beta advisor tool, the advisor must be Fable 5.2 or Mythos 5.2. Fable 5.1 as advisor now returns a 400 error. Changing the advisor counts as a tools edit, so make the change in new conversations.

Rules that carry over

Two rules carry over, and they will be new to you if you come from Opus 5 or earlier. Adaptive thinking can't be turned off: budget_tokens and {"type": "disabled"} both return 400. Forced tool use (tool_choice set to any or a named tool) is also rejected. For schema-valid JSON, use tool_choice: auto with strict tool use, or use JSON outputs. Also leave out temperature, top_p and top_k.

Platform and compliance gotchas

• Data retention. Fable 5.2 requires 30-day data retention and isn't available under zero data retention unless Anthropic expressly authorizes it, according to the model page. If your customer contracts promise ZDR, you can't drop Fable in quietly.
• Bedrock setup. The AWS launch post says Fable 5.2 is a Covered Model. By default it involves retention of up to 30 days and human safety review, and you must set the aws_review data retention mode before calling it. AWS says eligible customers can use ZDR for internal use through at least June 30, 2027. Anthropic's docs list the Bedrock ID as anthropic.claude-fable-5-2. AWS's examples use the global cross-region inference profile, global.anthropic.claude-fable-5-2.
• Refusals return HTTP 200. A declined request comes back with stop_reason: "refusal", not an error status. If your code only checks status codes, it will treat refusals as empty successes. Server-side fallback, in beta, can retry a refused request on Opus 4.8 or Opus 5. Refusals count against rate limits whether or not they are billed.
• Fast mode on Foundry is a separate deployment, claude-fable-5-2-fast, not a request flag.

Fable 5.2, Opus 5.5 or something cheaper?

Anthropic's docs say it plainly: start with Opus 5.5 for most workloads, and use Fable 5.2 "when your evals on Claude Opus 5.5 at higher effort still fall short." In practice:

• Use Fable 5.2 for agent sessions that run for hours, multistep research that ends in a finished document, hard refactors where Opus stalls, and orchestration in which Fable plans while cheaper models do the work.
• Stay on Opus 5.5 for most coding assistants, RAG, support bots and knowledge work. It costs half as much per token, and the benchmark gap on routine coding is small.
• Use Sonnet 5.5 ($2/$10) or Haiku 5.5 (from $0.10/$0.50) for high-volume classification, extraction, routing and subagent work.
• Security and biotech teams should know that Fable will often refuse their core tasks. Apply for the relevant verification program to get Mythos 5.2 rather than building workarounds.

If your product is a coding or agent tool built on Claude, check how competitors position their models in the AI coding assistants directory.

What to do now

1. Check your history handling. Search your codebase for anything that trims, reorders or rewrites messages, system or tools mid-conversation. That is the change most likely to break production.
2. Run an effort sweep. Run your hardest evals on Fable 5.2 at high (the default), then xhigh, then compare with Opus 5.5 at high. Measure cost per completed task, not cost per token.
3. Move off Fable 5.1 on purpose. No retirement date has been announced for 5.1, so you aren't forced to move this week. But 5.2 is cheaper and scores higher on Anthropic's benchmarks.
4. Fix routing and refusal handling. Treat stop_reason: "refusal" as a first-class outcome, and avoid sending Fable conversations to other models.

Things to watch next: independent benchmark results replacing Anthropic's self-reported table, any retirement notice for Fable 5.1, wider ZDR availability, and whether pricing changes after Anthropic's IPO.