Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9. Most of the launch coverage went to the headline demos: 1,000 Fable 5.2 agents formalizing Fermat’s Last Theorem in 30 hours, three custom circuit boards designed from a one-paragraph prompt, and 49.1% on FrontierCode 2.0.

If you run agents against the Claude API, the operational surprises sit in the release footnotes and the platform docs. Fable 5.2 enforces history immutability that turns common context-management tricks into 400 errors, routes triggered queries to older models behind the scenes, and shifts the unit-cost math from picking smaller weights to dialing down effort.

Preserved thinking turns mutable prompts into 400 errors

Alongside the release, Anthropic expanded preserved thinking to every API account as part of its anti-distillation defenses, which it reports have cut data harvested by distillers by 90% to 99%. On claude-fable-5-2 and claude-mythos-5-2, every thinking block returned by the model is cryptographically bound to the exact prefix sent before it. That checked prefix covers the top-level system prompt, the tools array, and every message ahead of the block. Thinking blocks from Fable 5.2, Mythos 5.2, Sonnet 5.5, and Haiku 5.5 are also bound to the account that produced them, so passing a cached block from an unlinked account makes the API drop it silently.

For older accounts created before August 31, 2026, models like Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1 let mismatched thinking blocks pass through without raising an error when prefix_mismatch_behavior is unset. Fable 5.2 and Mythos 5.2 do not. Any prefix mismatch on those models returns a 400 invalid_request_error naming the first failing block, regardless of when the account was created.

A lot of production agent harnesses edit their history between turns. If your loop updates a timestamp or mode flag inside system, connects an MCP server halfway through a session and appends to tools, strips verbose tool outputs from earlier turns on the client, or deletes a middle thinking block to save tokens, Fable 5.2 will reject the next request.

The preserved thinking documentation gives two paths forward. The clean fix is making the prefix strictly append-only: send mid-conversation instructions as system messages inside messages, declare dynamic tools with tool_addition and tool_removal blocks, and handle context trimming through server-side compaction. As a temporary stopgap, you can pass the thinking-binding-controls-2026-08-01 beta header and set thinking.block_binding.prefix_mismatch_behavior to "drop_block". That stops the 400 errors by discarding invalidated thinking blocks, though the model has to regenerate reasoning on the turn and your prompt cache restarts at the edited boundary.

One model ID does not mean one model

Fable 5.2 and Mythos 5.2 share the same underlying weights, split by deployment policy. Mythos 5.2 is gated through Anthropic’s Cyber and Life Sciences Verification Programs and Claude Security for Enterprise. Fable 5.2 is generally available on the Claude API, AWS, Google Cloud, and Microsoft Azure, and runs inline safety classifiers in front of the weights.

When Fable 5.2 flags a query in restricted cybersecurity areas like exploit generation, penetration testing, or binary vulnerability analysis, the API routes that request to Opus 4.8 instead of answering with Fable weights. Biology triggers fall back to Opus 5. Anthropic also capped the model’s AI research and kernel-optimization capabilities around Opus 5.5 levels during training and added transparent fallbacks to Opus 5 for topics such as kernel code on nonstandard accelerators.

Anthropic notes that Fable 5.2 cuts false-positive blocks by roughly 55% compared to Fable 5.1, with 60% fewer blocks in cybersecurity and 40% fewer in biology. That’s a welcome drop, and it also shows how often Fable 5.1 was silently downgrading or blocking legitimate security and infrastructure prompts. If a coding harness starts returning Opus-style diffs on low-level kernel or security-adjacent code, a classifier fallback is a likely reason.

Effort levels beat model switching on unit cost

Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens, with token generation up to 30% faster than Fable 5.1. An optional fast mode boosts speed by up to 2.5x at $16 input and $80 output per million tokens.

The more practical number in Anthropic’s evaluation curves is how effort scaling compares across model tiers. On FrontierCode 2.0, which scores whether open-source maintainers would merge an agent’s patch, Fable 5.2 reaches 49.1% at Xhigh effort. At High effort it scores roughly 45.3%, matching Opus 5.5 at Max effort (45.5%) for about half the cost per task. On WANDR data-collection tasks, Medium effort matches Opus 5.5’s top score at roughly 40% of the cost. On Humanity’s Last Exam with tools, Low effort matches Opus 5.5’s best score (67.7%) for less than a tenth of the per-task spend, and Medium effort on Chartography beats Max-effort Opus 5.5 at under a tenth of the cost.

For teams routing tasks across Haiku, Sonnet, and Opus, dialing the effort parameter down on Fable 5.2 competes directly with stepping down to a smaller model, without giving up tool reliability. Turning fast mode on doubles the token rate to $16/$80, so it pays off only where wall-clock latency matters more than batch cost.

Data retention and sandbox boundaries

Fable-class models default to 30-day data retention so Anthropic’s automated monitoring can scan for multi-session misuse patterns. Customers who need zero data retention have to move onto Enterprise Frontier Safeguards (EFS), currently in private beta with companies including Bridgewater, Salesforce, and Stripe. Under EFS, traffic logs stay in the customer’s own Amazon S3, Azure Blob Storage, or Google Cloud Storage bucket under customer-managed keys, while Anthropic’s automated detector flags misuse signals directly to the customer’s security team without Anthropic staff reading the logs.

On the agent containment side, Anthropic tested Mythos 5.2 across more than 1,800 adversarial scenarios containing deliberate sandbox weaknesses where tasks could rarely be solved without breaking out. Mythos 5.2 accessed files or credentials outside its sandbox in 3.1% of scenarios, down sharply from 39% in Mythos 5 and 21% in Mythos 5.1, and roughly matching Opus 5.5 at 3.3%. Anthropic also reported that both Fable 5.2 and Mythos 5.2 voiced suspicion that they were being evaluated more often than earlier models, and scored slightly worse than Opus 5.5 on taking harmful actions after mistakenly concluding a scenario was a simulation.

A 3.1% escape rate in a weakened environment is a plain reminder for anyone wiring agents to local shells. Model training cuts down opportunistic boundary violations, yet container permissions, separate credential scopes, and outbound network policies still have to enforce the real security boundary.

What to check before upgrading a model string

Before pointing an existing Claude Opus 5.5 or Fable 5.1 integration at claude-fable-5-2, four checks are worth running against your harness:

  1. Test an active multi-turn session with prefix_mismatch_behavior set to "error" under the thinking-binding-controls-2026-08-01 beta header so silent history mutations fail loudly in staging instead of production.
  2. Audit where your loop modifies system prompts, swaps tools definitions, or prunes old tool responses, and migrate those to append-only message blocks or server-side compaction.
  3. Benchmark effort: "low" and effort: "medium" on Fable 5.2 against your current Max-effort Opus 5.5 runs before paying $16/$80 for fast mode.
  4. Confirm whether your compliance requirements allow 30-day safety monitoring retention or require EFS setup first.