Claude Fable 5.1 API Specs: 1M Context, $10/M Input, and Tool Use Limitations

Seed story: "Introducing Claude Fable 5.1 and Claude Mythos 5.1" (Anthropic) · search original Written from facts verified across 3 report(s) — original explainer, not a copy or translation. Sources at the end.

Anthropic has released Claude Fable 5.1, a model featuring a 1M token context window and 128K maximum output that targets long-horizon agentic workflows with a reported 25% cost reduction over its predecessor. While the new pricing structure slashes cache read costs to $0.25 per million tokens, developers must navigate a significant API constraint: the model rejects forced tool use requests, returning a 400 error if tool_choice is explicitly set.

Release Overview and Availability

Claude Fable 5.1 and Claude Mythos 5.1 launched on September 1, 2026, sharing identical specifications and pricing structures. While Fable 5.1 is generally available, Mythos 5.1 remains restricted to Project Glasswing participants. This dual-track release allows developers to access the new capabilities immediately through Fable 5.1, with a knowledge cutoff of June 2026.

Current platform availability includes:

Claude API
Amazon Bedrock
Google Cloud
Microsoft Foundry

For engineering teams, this broad distribution means you can integrate the new model into existing CI/CD pipelines or cloud-native applications without migrating providers. The general availability of Fable 5.1 ensures that the 1M token context and updated pricing are accessible to all, not just a select group, facilitating immediate adoption in production environments.

Context Window and Output Limits

The 1M token context window in Claude Fable 5.1 provides substantial headroom for complex, long-horizon agentic workflows. This capacity allows developers to feed extensive documentation, codebases, or conversation histories into a single session without aggressive summarization. Such breadth is critical for agents that must maintain coherent state across numerous steps, reducing the risk of context drift during extended autonomous tasks.

While the input capacity is vast, the maximum output is capped at 128K tokens. This limit ensures that responses remain manageable for downstream processing and latency constraints. Developers should design their prompts to elicit concise, structured outputs rather than expecting the model to generate entire codebases in one go.

1M token input context
128K token maximum output
Supports long-horizon agentic tasks

Pricing Structure and Cache Economics

The headline figure is the $10 per million token input price, paired with a $50 per million token output rate. While the output cost remains steep for high-volume generation, the input tier positions the model competitively for tasks requiring massive context ingestion. The real economic shift, however, lies in the cache layer.

Anthropic has set cache reads at just $0.25 per million tokens. This rate is reportedly a quarter of what previous models charged, fundamentally altering the unit economics of stateful applications. For developers running multi-turn agents or RAG pipelines, this discount means that repeated access to shared context becomes significantly cheaper than fresh inference.

Input: $10 per million tokens
Output: $50 per million tokens
Cache Read: $0.25 per million tokens

According to Anthropic, these adjustments make Fable 5.1 approximately 25% less expensive than its predecessor for typical workloads. This reduction directly impacts total operational spend, allowing teams to scale context-heavy features without proportional budget increases.

Tool Use Constraints and Error Handling

Developers integrating Claude Fable 5.1 must account for a significant shift in tool invocation mechanics. The model no longer supports forced tool use, meaning you cannot mandate specific function calls. If your application logic relies on deterministic tool execution, this change will break existing workflows.

Specifically, the API now rejects requests that attempt to enforce tool selection. You will encounter immediate failures if you configure your requests with the following parameters:

Setting tool_choice to 'any'
Specifying a particular tool name in tool_choice

In both scenarios, the API returns a 400 error. This limitation requires developers to update their orchestration layers to rely on the model's native decision-making for tool selection rather than explicit directives.

Thinking Block Compatibility

The compatibility of extended thinking blocks in Claude Fable 5.1 is notably asymmetric. While the new model can successfully parse and utilize thinking blocks generated by earlier Claude versions, the reverse is not true. Earlier models are unable to read the thinking output produced by Fable 5.1. This one-way compatibility means that while you can migrate existing reasoning chains into the new model, you cannot downgrade those specific reasoning artifacts to older systems.

For developers managing multi-model pipelines, this limitation requires careful workflow design:

Forward migration: Existing thinking blocks from older models remain readable by Fable 5.1.
No backward compatibility: Fable 5.1 thinking blocks are opaque to previous model versions.
Pipeline isolation: Ensure reasoning chains generated by Fable 5.1 do not flow into legacy model stages.

Consequently, teams should avoid hybrid architectures that mix Fable 5.1 reasoning with older model inference steps. If your workflow relies on passing intermediate reasoning between different model versions, you must keep Fable 5.1 isolated or re-derive the reasoning in a compatible format before passing it to older systems.

Implementation Implications for Developers

The 25% estimated cost reduction for typical workloads primarily stems from the new cache read pricing of $0.25 per million tokens. This economic shift directly impacts the architecture of existing agentic pipelines, making long-running, stateful agents significantly more viable. Developers can now maintain larger context windows across multiple steps without the previous financial penalty, encouraging designs that rely heavily on persistent memory rather than frequent summarization.

However, specific API limitations require structural adjustments to your orchestration logic. Key architectural considerations include:

Tool Selection: You must remove any code relying on forced tool use, as setting tool_choice to 'any' or a specific name returns a 400 error.
Model Interoperability: Ensure your pipeline does not attempt to pass Fable 5.1 thinking blocks to earlier Claude models, as they cannot read them.
Platform Consistency: Since Fable 5.1 shares specs with Mythos 5.1, your infrastructure must handle the availability gap, as Mythos is restricted to Project Glasswing participants.

These constraints mean that while the cost floor drops, your error handling and model routing logic must be updated to accommodate the new strictness in tool invocation and cross-model compatibility.

FAQ

What are the context window and pricing details for Claude Fable 5.1?

Claude Fable 5.1 features a 1M token context window with a maximum output of 128K tokens. The pricing is set at $10 per million input tokens and $50 per million output tokens, with cache reads costing $0.25 per million tokens.

Does Claude Fable 5.1 support forced tool use via the tool_choice parameter?

No, Claude Fable 5.1 does not support forced tool use. If you set the tool_choice parameter to 'any' or a specific tool name, the API will return a 400 error.

How does Claude Fable 5.1 handle thinking blocks compared to earlier models?

Claude Fable 5.1 can read thinking blocks generated by earlier models, but earlier Claude models cannot read the thinking blocks produced by Fable 5.1. This creates a one-way compatibility for these specific data blocks.