Claude Fable 5 API Example: Retries, Fallbacks, OpenRouter & Bedrock

The API call is the easy part. A reliable Claude Fable 5 integration needs a provider decision, a cost ceiling, bounded retries, and an explicit fallback route. Those four choices matter more in production than a copy-pasted request body.

Start with a provider decision

Fable 5 can appear through a direct API, AWS Bedrock, or an aggregator such as OpenRouter. Before writing code, write down which account owns the invoice and which region or provider is the production source of truth.

Do not hardcode a model ID from a blog post. Confirm it in the provider’s current documentation at deploy time.

A safe request shape

Keep model selection and budget policy outside the prompt. The following is provider-neutral pseudocode, not a replacement for the current SDK documentation:

const result = await runWithPolicy({
  primaryModel: 'claude-fable-5',
  fallbackModel: 'claude-opus-4-8',
  maxAttempts: 2,
  timeoutMs: 90_000,
  budget: { maxInputTokens: 120_000, maxOutputTokens: 8_000 },
  task: { kind: 'repository-refactor', id: jobId },
  messages,
});

The policy should return which provider and model actually completed the task. That one field prevents weeks of confusing billing and quality analysis later.

Retry only the failures that can recover

Retry transient failures such as timeouts, short-lived rate limits, and provider 5xx errors with exponential backoff and a small retry budget. Do not blindly retry authentication errors, invalid input, policy blocks, or a job that has already exceeded its spend ceiling.

For long-running agents, save a checkpoint after each completed step. A retry should resume from a durable state, not rerun an entire expensive task.

Design fallback as a product decision

The useful fallback is rarely “try every model.” Instead, define a narrow rule:

• Use Fable 5 for hard, long-horizon work where retries are costly.
• Use Opus 4.8 for routine work, capacity issues, or a constrained fallback lane.
• Stop and surface an error for tasks that require the primary model’s specific capability or a human approval.

Compare the two models by task in Fable 5 vs Opus 4.8, and keep a portable system prompt with the prompt generator. A fallback model is only useful if it can understand the same task contract.

OpenRouter: use it for routing, not obscurity

OpenRouter can reduce integration work when your product needs more than one model. It is a good fit for controlled experiments and a provider-agnostic routing layer. It is a poor fit if you have not decided who owns observability, privacy review, outage handling, and the final invoice. Log the resolved provider and model on every request; otherwise a “Fable 5” label is not enough to explain cost or latency.

Bedrock: plan for region and access variation

With Bedrock, availability and approval can differ by region and account. Confirm access in the exact deployment region and keep an alternative model route ready. Our Bedrock access guide covers the operational checks; this article focuses on application-level retries and routing.

Production checklist

• Confirm current provider model ID and availability before deployment.
• Enforce token and dollar ceilings before the request leaves your service.
• Retry only transient failures, with a bounded budget.
• Persist agent checkpoints so a retry does not restart paid work.
• Log requested model, resolved model, provider, token use, latency, and outcome.
• Test the fallback with real prompts before an outage makes it your default.

For token-cost planning, use the pricing calculator. For a source-aware reading of model scores, see How to read Fable 5 benchmarks.