Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- Claude Fable 5.1 was released on September 1, 2026, with Claude API model ID claude-fable-5-1. Mythos 5.1 is available only to approved organizations.
- Fable 5.1 does not support forced tool_choice with any or a named tool. Continuing to send these returns 400 invalid_request_error.
- Thinking is always adaptive. disabled, enabled with budget_tokens, and a final assistant prefill message all fail.
- Fable 5.1 and Mythos 5.1 require 30-day data retention by default. Requests from ZDR organizations without explicit authorization return 400.
- Establish an old-model baseline before gradually rolling out the new model. Stop if tool calls, thinking blocks, refusals, costs, or long-task recovery behave incorrectly.
What changed: why replacing only the model ID can still produce 400
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. Fable 5.1 serves general API customers with model ID claude-fable-5-1. Mythos 5.1 uses claude-mythos-5-1 but remains limited to approved Project Glasswing customers. Neither is an alias guaranteed to accept an unchanged old request body.
A typical scenario is that the model name has been updated, yet production requests repeatedly return 400. Changing proxies first or retrying indefinitely is the wrong instinct. The actual risks usually involve forced tools, manual thinking, assistant prefill, or data-retention requirements. First save the error body and request ID, then statically inventory request parameters.
Choose between Fable 5.1 and Mythos 5.1 first
For ordinary coding, knowledge work, and long-running agent tasks, evaluate Fable 5.1 first because it does not require Mythos access approval. Mythos 5.1 serves approved cyberdefense and life-sciences organizations. Without approval, do not treat model not found as a network failure or repeatedly change accounts to bypass access controls.
Fable 5.1 and Mythos 5.1 use the same underlying capabilities but differ in safety classifiers, availability, and approval paths. Official guidance also says Fable requests may be routed to other models because of safety classification. Your application must record the actual response, stop_reason, and stop_details, not just the model name in the request.
Four request configurations that must be cleaned up
The first is tool_choice: auto and none are available, while any or a specified tool returns 400. If the goal is merely to force structured JSON, use an output format with a strict schema. If the application must call a particular tool, state that requirement in the instructions and verify the actual tool_use; do not pretend a server-side enforcement guarantee still exists.
The second is thinking: do not send disabled or enabled with budget_tokens. The third is a final assistant prefill message; replace it with system instructions or structured output. The fourth is old sampling controls such as temperature, top_p, and top_k. During a model-generation migration, remove these according to official model guidance rather than continuing to try values.
Cross-model thinking blocks can break session continuation
Fable 5.1 can read thinking blocks from some older models, but older models cannot read blocks produced by Fable 5.1. If gradual-rollout routing switches models back and forth within a session, directly reusing all assistant content can break rollback.
Before launch, test all three paths separately: old to new, new to new, and new back to old. During rollback, rebuild context from user-visible text, tool results, and application state first. Do not assume protected thinking content can make a round trip between models.
Data-retention, pricing, and availability boundaries
The official migration guide states that both 5.1 models require 30-day data retention by default. ZDR organizations or workspaces without explicit authorization receive 400 invalid_request_error. Have security, legal, and account owners confirm workspace eligibility first; do not change organization-wide data policy on the fly after a production error.
Fable 5.1 has the same input and output unit prices as Fable 5, with a lower cache-read unit price. Actual per-task cost still depends on cache hits, tool loops, output length, and fallback models. The pricing page and billing records are authoritative; do not turn average savings in official examples into a promise for your own workload.
Seven migration steps from inventory to gradual rollout
Proceed in order: search for every model ID; export a request-parameter matrix; remove configurations that trigger 400; fix a set of text, tool, long-context, and refusal samples; replay them in a test workspace; roll out to a small traffic share; and only then update the default model and background tasks. At each step, record the owner, samples, error rate, latency, tokens, cache, and cost.
Unattended agents also need checks for timeouts, cancellation, tool idempotency, session recovery, and notifications. One successful synchronous text request is not enough to approve scheduled tasks or long-running agents. Use the AI API configuration guide to check the separate endpoint, key, and model boundaries.
When to stop, roll back, and troubleshoot
Stop expanding traffic if 400 errors persist, forced-tool semantics cannot be replaced, cross-model sessions cannot recover, ZDR eligibility is unclear, refusal classifications are unhandled, or cost or latency exceeds thresholds. Retain old-model routing and a separate session baseline. Do not change prompts, the SDK, model, and network together during an incident.
Troubleshoot in this order: error body and request ID, model availability, request parameters, workspace retention policy, tool and session state, and finally the connection layer. Move to the proxy connection troubleshooting checklist only when there is evidence of DNS, TLS, connection-timeout, or proxy-authentication problems.
Sources
Frequently Asked Questions
What is Claude Fable 5.1's API model ID?
The official migration guide specifies claude-fable-5-1. Mythos 5.1 uses claude-mythos-5-1 but is available only to approved organizations.
Why does tool_choice return 400 after migration?
Fable 5.1 does not support forcing any or a specified tool. Use auto or none, replacing the old forced behavior with strict structured output or explicit instructions.
Can I keep using budget_tokens?
No. Thinking in Fable 5.1 and Mythos 5.1 is always adaptive. Both manual enabled with budget_tokens and disabled return 400.
Can a ZDR workspace call Fable 5.1 directly?
Not by default. Official guidance requires 30-day data retention unless Anthropic explicitly authorizes an exception. Requests from ineligible organizations or workspaces return 400.
Can Fable 5.1 directly continue a session from an older model?
Test it first. It can read thinking blocks from some older models, but older models cannot read Fable 5.1 blocks. This is especially important to verify for rollback during a gradual rollout.
When should I roll back to the old model?
If tool calls, session recovery, data retention, refusal handling, cost, or error rate fail preset thresholds, stop expanding traffic and return to the verified old route.