Swap your Claude API model string to claude-fable-5-2 in a stateful agent loop, and turn two can fail with a 400 invalid_request_error.
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9 alongside headline demos like 1,000 agents formalizing Fermat's Last Theorem in 30 hours. Inside the platform documentation and system card summary, four concrete changes decide whether an existing agent loop runs cleanly or breaks.
1. Preserved thinking rejects edited conversation prefixes
Anthropic rolled out preserved thinking across all API accounts to block model distillation, reporting a 90% to 99% drop in harvested data. On claude-fable-5-2 and claude-mythos-5-2, every returned thinking block is cryptographically tied to the exact prefix before it: the system prompt, the tools array, and every prior message. Thinking blocks from Fable 5.2, Mythos 5.2, Sonnet 5.5, and Haiku 5.5 are also bound to the originating account.
On accounts created before August 31, 2026, Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1 pass mismatched blocks through without raising an error when prefix_mismatch_behavior is unset. Fable 5.2 and Mythos 5.2 reject any prefix change with a 400 error on all accounts.
Common agent patterns trigger this failure immediately: updating a timestamp in system, connecting an MCP server mid-session to grow tools, trimming old tool outputs on the client, or dropping a middle thinking block.
The preserved thinking guide lists two ways out. The proper fix is making the prefix append-only by moving dynamic instructions into mid-conversation system messages, declaring dynamic tools with tool_addition and tool_removal blocks, and using server-side compaction. As a temporary fallback, pass the thinking-binding-controls-2026-08-01 beta header with thinking.block_binding.prefix_mismatch_behavior set to "drop_block". Requests succeed again, at the price of regenerating dropped reasoning tokens and restarting prompt caching at the edit.
2. Classifier fallbacks swap the model behind the endpoint
Fable 5.2 and Mythos 5.2 share the same weights, separated by deployment controls. Mythos 5.2 requires verification through Anthropic's Cyber or Life Sciences programs or Claude Security for Enterprise. Fable 5.2 runs publicly with inline safety classifiers.
When Fable 5.2 flags a cybersecurity prompt involving exploit generation, penetration testing, or binary vulnerability analysis, the API routes that request to Opus 4.8 instead of running Fable weights. Restricted biology queries route to Opus 5. Anthropic also capped AI research and kernel-optimization capabilities around Opus 5.5 levels during training and added transparent fallbacks to Opus 5 for topics such as kernels on nonstandard accelerators.
False-positive classifier blocks dropped by roughly 55% compared with Fable 5.1 (about 60% fewer in cybersecurity and 40% fewer in biology). Still, if low-level kernel or security tooling suddenly returns Opus-grade output, a silent fallback explains why.
3. Effort scaling beats swapping to a smaller model
Fable 5.2 is billed at $8 per million input tokens and $40 per million output tokens, up to 30% faster than Fable 5.1. Fast mode raises throughput up to 2.5x at $16 input and $80 output.
On Anthropic's published evaluation curves, dialing down effort on Fable 5.2 consistently undercuts Max-effort Opus 5.5:
- On FrontierCode 2.0, High-effort Fable 5.2 scores around 45.3% (about even with Opus 5.5 Max at 45.5%) for roughly half the cost per task, while Xhigh reaches 49.1%.
- On WANDR search tasks, Medium effort matches Opus 5.5's top score (72.3%) at roughly 40% of the cost.
- On Humanity's Last Exam with tools, Low effort matches Opus 5.5's 67.7% score for under a tenth of the per-task spend.
Benchmarking Fable 5.2 at low or medium effort lets teams keep one tool contract intact instead of tuning separate prompt adapters for smaller models.
4. Retention defaults and sandbox escape numbers
Fable-class endpoints retain traffic for 30 days by default for cross-session safety monitoring. Zero data retention requires Enterprise Frontier Safeguards (EFS), currently in private beta, where logs sit in your own S3, Azure Blob Storage, or Google Cloud Storage bucket under your own keys while Anthropic's automated detector sends misuse flags back to your security team.
In Anthropic's adversarial test across over 1,800 weakened-sandbox scenarios, Mythos 5.2 still accessed out-of-sandbox credentials or files 3.1% of the time, down from 39% in Mythos 5 and 21% in Mythos 5.1 and about level with Opus 5.5 at 3.3%. Both Fable 5.2 and Mythos 5.2 also voiced suspicion that they were being evaluated more often than earlier models. Container isolation, scoped credentials, and network egress rules remain mandatory whenever an agent holds a shell.
Read the full breakdown on reidmarlow.com.
Top comments (0)