Claude Fable 5.2: 20% Price Cut, Four Breaking Changes

Anthropic dropped Claude Fable 5.2 and Mythos 5.2 on November 9. The models are 20% cheaper, 30% faster, and they come with four breaking changes. The most urgent: if you are using the computer_20251124 computer-use tool anywhere outside Amazon Bedrock, your integration is returning 400 errors right now.

Four Things That Break

Anthropic lists four breaking changes from Fable 5.1. Two are likely to catch teams off guard.

1. computer_20251124 Is Gone (Except on Bedrock)

The old computer-use tool declaration now returns 400 invalid_request_error on every platform except Amazon Bedrock, where both old and new tools still work. The fix is a one-line change: replace computer_20251124 with computer_toolset_20260801 in your tool declarations and drop the computer-use-2025-11-24 beta header. The migration guide has the step-by-step.

2. Prefix Check Is Now Universal

On Fable 5.1, the thinking-block prefix check only applied to API accounts created after August 31, 2026. On Fable 5.2, it applies to everyone. If your code modifies system, tools, or any earlier message after a thinking block is in the conversation, the API returns 400: The block is bound to a different conversation. The temporary workaround is the thinking-binding-controls-2026-08-01 beta header with prefix_mismatch_behavior: "drop_block" — but dropping blocks raises output tokens, so treat it as a stopgap while you fix the underlying history mutation.

3. Adaptive Thinking Is Always On

Sending thinking: {"type": "enabled"} or thinking: {"type": "disabled"} both return a 400. Remove the thinking parameter entirely and use the effort parameter instead. The five levels are low, medium, high (default), xhigh, and max. Start at high and only step up after measuring a real gain.

4. Forced tool_choice Is Out

Setting tool_choice to {"type": "any"} or {"type": "tool", "name": "..."} returns a 400. For structured output, use strict tool use (strict: true) or JSON output mode. The advisor tool also requires an upgrade: a Fable 5.2 executor must have a Fable 5.2 or Mythos 5.2 advisor. Pointing it at Fable 5.1 returns a 400.

Pricing: Down 20% Across the Board

Fable 5.2 and Mythos 5.2 share the same price table. Everything dropped 20% from 5.1 — input, output, batch, and cache reads.

• Input: $8/MTok (was $10)
• Output: $40/MTok (was $50)
• Batch: $4/$20 per MTok
• Cache reads: $0.20/MTok (was $0.25)
• Fast mode: $16/$80 per MTok — 2.5x faster for latency-sensitive workloads

The cut also undercuts GPT-6 Astra Standard, which still runs at $10/$50. At the task level, Fable 5.2 at high effort runs about $6.72 per FrontierCode task — essentially matching Astra at max effort ($6.68) while leading it on the benchmark by 6.3 percentage points.

Performance: Where It Actually Gains

The headline numbers are real, but the effort-level nuance matters for your wallet. On FrontierCode 2.0, xhigh reaches 49.1% at $14.19 per task. max scores 45.7% at $44.70. That is a 3x cost increase for a worse result. The takeaway: start at high, move to xhigh for hard coding or agentic work where you have measured the gain, and treat max as a ceiling test rather than a production setting.

Other notable jumps vs Fable 5.1: FrontierSWE v2 goes from 56.3% to 69.9% (also ahead of GPT-6 Astra’s 65.5%), AutomationBench from 31.4% to 51.4%, and RoboDojo robotics roughly doubles to 64.9%. On the agentic coordination side, Anthropic demonstrated 1,000+ agents building a working OS prototype over eight days and 1,000 agents formalizing Fermat’s Last Theorem in 30 hours with minimal human input.

Migration Steps

If you are running Fable 5.1 in production, here is what to do before switching the model ID:

1. Migrate computer use. Replace computer_20251124 with computer_toolset_20260801. Remove the computer-use-2025-11-24 beta header.
2. Check history mutation. If your code ever modifies system, tools, or any earlier message in an active conversation, fix it before switching. Add the thinking-binding-controls-2026-08-01 beta header temporarily to catch violations in staging.
3. Remove thinking/tool_choice parameters. Delete any explicit thinking type settings. Replace forced tool_choice with strict tool use or JSON output mode.
4. Update the advisor. If using the advisor tool, set its model to claude-fable-5-2 or claude-mythos-5-2 in new conversations.
5. Re-run your evals at high effort. Then step up to xhigh for the specific tasks that justify it. Do not default to max.

The official Fable 5.2 release notes and migration guide cover each path, including routes from Opus models.

Mythos 5.2: Same Model, Restricted Access

Mythos 5.2 is the same underlying model as Fable 5.2 with certain cybersecurity and biology safeguards removed. It requires verification through Anthropic’s Cyber or Life Sciences programs. For teams doing authorized security work, Mythos 5.2 hits 89.6% on CVE discovery benchmarks, up from 81.3% on Opus 5.5. Fast mode is also available on Mythos 5.2 at $16/$80 per MTok.