Claude Mythos Preview Retires June 30: Five Code Changes Required Before the Deadline

The Anthropic deprecation documentation lists claude-mythos-preview as retiring on June 30, 2026. After that date, any API call specifying that model ID will fail with no grace period. The replacement is claude-mythos-5, which has been available since June 9, 2026, when Anthropic released Claude Fable 5 and Claude Mythos 5 together. Claude Mythos Preview was the limited-access model that Project Glasswing partners used prior to that launch. It supported manual extended thinking configuration, assistant message prefill, and raw chain-of-thought output. None of those three capabilities exist in Mythos 5. Anthropic notified customers with at least 60 days notice, consistent with their stated deprecation policy. The June 30 deadline is firm.

Anthropic's official migration guide identifies five specific patterns that will return 400 errors on Mythos 5. First, any thinking block configured with type 'enabled' and a budget_tokens value will be rejected. Second, thinking configured with type 'disabled' will also return an error. Third, budget_tokens as a standalone parameter must be removed from all inference calls entirely. Fourth, assistant message prefill, the pattern where you begin a turn by specifying a partially written assistant response, is not supported and will return an error. Fifth, the raw chain of thought is never returned on Mythos 5, even for teams that previously relied on it for interpretability or compliance audit workflows. Each of these five patterns requires a targeted code change, not just a model name update. Teams that built extended thinking workflows on Mythos Preview will need to restructure their approach to inference control before the migration is complete.

Mythos 5 uses adaptive thinking by default. The model decides how deeply to reason based on the complexity of each task, with no budget_tokens ceiling to configure. If you need to influence reasoning depth, use the effort parameter instead. Setting effort to 'low' prioritizes speed over depth and suits routine extraction or summarization tasks. Setting effort to 'high' directs the model toward thorough analysis, complex code generation, or multi-step reasoning. To receive readable summaries of the model's reasoning process, set thinking.display to 'summarized'. The default is 'omitted', meaning thinking output is suppressed unless explicitly requested. If your workflow previously relied on thinking blocks for interpretability or compliance documentation, the summarized display setting is the closest available equivalent in Mythos 5. The raw chain of thought has been permanently removed from Mythos 5 outputs. Anthropic has not indicated whether a future model will restore it.

The migration is not purely subtractive. Mythos 5 is the same underlying model as Claude Fable 5, trained together and differentiated only by which safety classifiers are active. The pricing on Mythos 5 is $10 per million input tokens and $50 per million output tokens, which is less than half the cost of Claude Mythos Preview at $25 per million input tokens and $125 per million output tokens. Token counts are roughly unchanged because both models share the same tokenizer. Latency on routine inference is meaningfully lower as well, since the model no longer waits for a manually configured extended thinking budget to exhaust before responding. Organizations that complete the migration should see faster, cheaper responses on comparable task types.

Anthropic's migration guide provides eight specific checklist items for teams moving from Mythos Preview to Mythos 5. Update the model name string from claude-mythos-preview to claude-mythos-5. Remove any thinking configuration that uses type 'enabled' with budget_tokens. Remove any thinking configuration that uses type 'disabled'. Remove budget_tokens from all inference calls. Verify that any code handling the thinking field treats it as display text only, not as raw reasoning output. When continuing conversations with prior history, pass thinking blocks back unchanged as context. If your system replays conversation history on a different model, strip thinking and redacted_thinking blocks first to avoid format errors. Finally, re-baseline your token count and cost assumptions on representative workloads, since adaptive thinking changes throughput characteristics compared to manually configured extended thinking budgets.

Separately from the Mythos Preview retirement, Anthropic has extended sampling parameter restrictions to cover Claude Opus 4.8. Setting temperature, top_p, or top_k to any non-default value on Claude Opus 4.7 or Opus 4.8 now returns a 400 error. This restriction was originally applied only to Opus 4.7 and has since been extended to Opus 4.8 as well. The recommended approach is to remove these parameters from API calls and use prompt engineering to guide model behavior instead. Any integration that sets these parameters explicitly and tests against an older model will pass in staging but break in production when pointing at Opus 4.7 or 4.8.

The timeline is immediate. If your organization holds a Project Glasswing credential or has been using the claude-mythos-preview model ID in any production workload, migration must be completed before June 30. Start by searching your codebase for the string claude-mythos-preview and for patterns using thinking type 'enabled', budget_tokens, and assistant message prefill. Migrate those configurations using the checklist above. Run your workload against claude-mythos-5 in a staging environment today and monitor for 400 errors at each of the five failure points identified in Anthropic's migration guide. If you are uncertain whether your organization qualifies for Mythos 5 access under current export restrictions, note that the partial restoration announced June 27 covered organizations classified under critical infrastructure designations. Contact Anthropic through the Project Glasswing access portal to verify your credentials before the June 30 deadline.