Anthropic model
Claude Fable 5.1
Anthropic’s high-end model for demanding reasoning and long-horizon agentic work, with a 1M-token context window, 128K output, always-on adaptive thinking, and stronger coding, research, and knowledge-work performance than Fable 5.
- Context window
- 1M
- tokens
- Max output
- 128K
- tokens
- Input
- $10.00
- per 1M tokens
- Cache read
- $0.25
- per 1M tokens
- Output
- $50.00
- per 1M tokens
01 / Overview
What Claude Fable 5.1 Is
Claude Fable 5.1 is Anthropic’s specialist model for demanding reasoning and long-horizon agentic work, positioned above the normal starting point for most production workloads.
Not the Default Claude — the Escalation Tier
Anthropic explicitly recommends starting most workloads with Claude Opus 5. Fable 5.1 is intended for cases where a difficult task still falls short even after evaluating Opus 5 at higher effort.
That makes its role unusually clear. Fable 5.1 is not primarily a cheaper, faster, or more general replacement for the rest of the Claude lineup. It is the model to evaluate when the application needs stronger long-running agentic coding, deeper multistep research, or more demanding document and knowledge work.
The model has a 1,000,000-token context window, supports up to 128,000 output tokens, accepts text and images, and uses adaptive thinking on every request. Its default effort level is high and Anthropic categorizes comparative latency as slower.
- Designed for demanding reasoning and long-horizon agents.
- 1M-token context window with 128K maximum output.
- Text and image input with text output.
- Adaptive thinking is always on.
- General availability through the Claude API and supported cloud platforms.
- Provider
- Anthropic
- Family
- Claude Fable 5.1
- Model ID
- claude-fable-5-1
- Knowledge cutoff
- Jun 2026
- Input modalities
- Text, Image
- Output modality
- Text
- Thinking
- Adaptive · always on
02 / Agentic work
Claude Fable 5.1 for Long-Horizon Agentic Work
Anthropic highlights stronger long-running agentic coding as one of the central improvements in Fable 5.1 over Fable 5.
Evaluate the Whole Trajectory, Not One Answer
Agentic workloads are different from single-turn prompting. A coding agent may inspect a repository, formulate a plan, call tools repeatedly, edit files, recover from failures, re-check its work, and continue across a long sequence of decisions.
The useful metric is therefore not only whether one response looks intelligent. You need to measure whether the model remains coherent across the entire trajectory: does it preserve the goal, choose appropriate tools, recover from mistakes, avoid unnecessary loops, and complete the task without human rescue?
Fable 5.1 is specifically positioned for that kind of sustained work. Its large context window gives an agent room for source files, conversation history, tool results, plans, and intermediate state, while the 128K output ceiling supports unusually large generated artifacts when the task requires them.
Because the model is expensive and comparatively slower, it is most rational as an escalation tier for difficult agent tasks rather than the first model called by every workflow.
- 01
Repository-scale coding
Work across multiple files, dependencies, tests, tool calls, and implementation steps while retaining a large project context.
- 02
Long-running autonomous tasks
Maintain a goal across extended sequences of planning, execution, observation, correction, and verification.
- 03
Hard debugging
Trace failures through code, logs, configuration, and tool output when the root cause is not obvious from one local symptom.
- 04
Escalation workflows
Route only the hardest tasks to Fable 5.1 after a cheaper or faster model fails an evaluation or confidence threshold.
03 / Knowledge work
Multistep Research, Documents, Spreadsheets, and Slides
Fable 5.1 extends the Fable line beyond coding by improving multistep research and complex work across documents, spreadsheets, and presentation materials.
Long Context Matters Most When the Work Has State
Research-heavy tasks often accumulate evidence over time. The model may need to inspect many sources, compare conflicting details, preserve intermediate findings, update a working hypothesis, and finally synthesize a coherent result.
The same applies to knowledge-work artifacts. A spreadsheet analysis may require understanding formulas, assumptions, tables, and narrative context. A presentation task may depend on source documents, prior slides, and a consistent argument across many pages.
Fable 5.1’s 1M-token context window is especially relevant when the model needs to keep a large amount of working material available without repeatedly discarding or compressing earlier state.
However, capacity is not evidence of quality by itself. Evaluate whether the model actually retrieves the right facts from the supplied context, maintains consistency across long tasks, and produces artifacts that survive downstream review.
- 01
Multistep research
Accumulate evidence, compare sources, revise hypotheses, and synthesize findings across a long workflow.
- 02
Document analysis
Work across long reports, policies, specifications, and collections of reference material.
- 03
Spreadsheet reasoning
Analyze structured business data, assumptions, tables, and calculation-heavy workflows.
- 04
Presentation work
Transform source material into coherent slide narratives while preserving structure and supporting evidence.
04 / Pricing
Claude Fable 5.1 Pricing
Fable 5.1 costs $10 per million input tokens and $50 per million output tokens—the same base token rates as Fable 5—but dramatically reduces prompt-cache read cost.
Cache Economics Are the Major Pricing Change
The direct input and output prices did not change from Fable 5. The meaningful cost improvement is cache reading: Fable 5.1 cache reads cost $0.25 per million tokens, one quarter of Fable 5’s $1 per million rate.
That matters for long-running agents because they often reuse substantial stable context: system instructions, repository summaries, policies, long documents, or previous material that can be cached rather than reprocessed at full input price.
Anthropic lists 5-minute cache writes at $12.50 per million tokens and 1-hour cache writes at $20 per million tokens. Batch API usage receives a 50% discount on standard input and output pricing.
The right economic unit is therefore not simply one uncached request. For an agent, measure the full trajectory: uncached input, cache writes, cache reads, output generation, number of turns, and whether the task reaches a useful completion.
- Standard input: $10.00 per 1M tokens.
- Output: $50.00 per 1M tokens.
- 5-minute cache write: $12.50 per 1M tokens.
- 1-hour cache write: $20.00 per 1M tokens.
- Cache read: $0.25 per 1M tokens.
- Batch API: 50% discount on input and output.
1M tokens · USD
- Input
- $10.00
- 5m cache write
- $12.50
- 1h cache write
- $20.00
- Cache read
- $0.25
- Output
- $50.00
Example: 20K input + 4K output
- Input cost
- $0.2000
- Output cost
- $0.2000
- Estimated uncached total
- $0.4000
05 / Context
A 1M Context Window With 128K Maximum Output
Claude Fable 5.1 provides a 1,000,000-token context window by default and can generate up to 128,000 output tokens in a single request.
Built for Tasks That Accumulate Large Working State
A million-token context window can hold extensive source code, long research material, large document sets, prior conversation turns, tool results, and detailed system instructions at the same time.
For long-running agents, this can reduce the frequency of aggressive summarization or context eviction. For research, it can make more evidence available in one working session. For document and artifact workflows, it can keep source material closer to the generation step.
Anthropic now exposes the 1M window as the default for Fable 5.1 rather than requiring a special beta header, and the documentation states that long-context requests use standard model pricing.
Even so, a larger context window is not a reason to indiscriminately send everything. Relevant-context selection, caching, and measuring retrieval accuracy still matter for both quality and cost.
- Context window: 1,000,000 tokens.
- Maximum output: 128,000 tokens.
- 1M context is available by default.
- Text and image inputs can share the working context.
Context window
1,000,000
Max output
128,000
A long-running agent may use the context for system instructions, source files, images, previous turns, tool results, research evidence, and generated output.
06 / Thinking & effort
Adaptive Thinking Is Always On
Fable 5.1 uses adaptive thinking on every request and defaults to high effort, giving Anthropic’s highest-demand workloads more inference budget without requiring a separate extended-thinking mode.
Effort Is a Workload Control, Not a Quality Guarantee
Anthropic’s model comparison lists Fable 5.1 as adaptive thinking, always on, with high as the default effort level.
Fable 5.1 also introduces per-message effort as a beta capability. This is important for long workflows where not every turn deserves the same computational intensity. An agent might use more effort for architecture decisions or difficult debugging and less for routine intermediate steps.
The best setting should come from evaluation. Higher effort may improve difficult reasoning but can also change latency and economics. Measure whether the additional compute improves the outcome that matters: fewer failed trajectories, better code, stronger research synthesis, or fewer human interventions.
Routine agent step
Use lower effort where the next action is mechanical and evaluation shows no meaningful quality loss.
Hard decision point
Use higher effort for architecture, difficult debugging, deep synthesis, or other steps where reasoning quality determines the trajectory.
07 / Fable 5 → 5.1
What Changed From Claude Fable 5 to Fable 5.1
Fable 5.1 keeps the same $10/$50 base token pricing and 1M/128K capacity as Fable 5, but changes agent behavior, cache economics, thinking compatibility, and several API details.
The Version Upgrade Is Not Fully Drop-In
The most visible economic change is cache-read pricing: $0.25 per million tokens instead of $1 per million on Fable 5.
Anthropic also documents stronger long-running agentic coding, multistep research, and document, spreadsheet, and slide work. Additive API features include beta per-message effort, beta turn-scoped system messages, readable progress updates between tool calls using display: "updates", and content provenance.
There are also breaking changes. Forcing a tool call with tool_choice values such as any or a named tool returns a 400 error. Fable 5.1 can read thinking blocks from earlier supported Claude models, but those earlier models cannot read Fable 5.1 thinking blocks. Editing earlier turns can invalidate thinking blocks.
That means migration testing should include conversation persistence and tool orchestration—not only answer quality.
- 01
Cheaper cache reads
Prompt-cache reads fall from $1/MTok on Fable 5 to $0.25/MTok on Fable 5.1.
- 02
Stronger long-running work
Anthropic highlights improvements in agentic coding, multistep research, and office-document workflows.
- 03
New beta controls
Per-message effort, turn-scoped system messages, readable tool-call progress updates, and content provenance are additive features.
- 04
Breaking tool and thinking changes
Forced tool use errors, thinking-block compatibility is asymmetric, and editing earlier turns can invalidate thinking state.
08 / Capabilities
Claude Fable 5.1 Capabilities
Fable 5.1 combines multimodal input, adaptive reasoning, large context, prompt caching, tool-oriented agent workflows, and unusually long output capacity in one premium model.
Designed Around Long Tasks Rather Than Short Chat Alone
Text and image input allow the model to combine written instructions with screenshots, diagrams, document images, and other visual evidence.
Tool use is central to its agentic positioning, but Fable 5.1 changes how tool selection works: automatic and no-tool modes are supported, while forced use of a specific tool or any tool is not.
Prompt caching is particularly relevant because long-running tasks often reuse large stable prefixes. Anthropic’s reduced cache-read rate makes repeated access to that state materially cheaper than on Fable 5.
The model is available through the Claude API as claude-fable-5-1 and is also listed for Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS.
- Supported
Text
Accept text input and generate text output.
- Supported
Image input
Analyze supported images together with text and long-context material.
- Supported
Adaptive thinking
Thinking is always on, with high effort as the default.
- Supported
Tool use
Supports agentic tool workflows, with important restrictions on forced tool choice.
- Supported
Prompt caching
Cache reads cost $0.25/MTok, significantly lower than on Fable 5.
- Supported
Batch API
Batch processing receives a 50% discount on standard input and output pricing.
- Supported
128K output
Generate unusually large code, research, and artifact outputs when the workload requires them.
09 / Evaluation
Claude Fable 5.1 Strengths and Limitations
Fable 5.1 is a specialist premium model: it makes the most sense when difficult long-running work produces enough value to justify slower latency and substantially higher token cost than Opus or Sonnet tiers.
Strengths
Long-horizon agentic capability
Anthropic positions Fable 5.1 for demanding coding and agent workflows that need to remain coherent across many steps.
1M context and 128K output
Large working memory and output capacity suit repository-scale, research-heavy, and artifact-generation tasks.
Stronger research and knowledge work
Fable 5.1 improves multistep research and complex work with documents, spreadsheets, and slides.
Much cheaper cache reads than Fable 5
$0.25/MTok cache reads improve the economics of agents that repeatedly reuse large stable context.
What to consider
Premium token pricing
$10 input and $50 output per million tokens make indiscriminate routing expensive.
Slower comparative latency
Anthropic classifies Fable 5.1 as slower, so it is not the natural choice for latency-sensitive routine requests.
Migration has breaking changes
Forced tool selection and thinking-block compatibility require explicit testing when moving from Fable 5 or other Claude models.
Retention requirements
Anthropic documents a 30-day data-retention requirement unless an exception is expressly authorized, which may affect regulated or ZDR deployments.
Use the expensive tier only when the workload earns it
Evaluate Claude Fable 5.1 on your hardest real tasks
Compare long-running agent behavior, coding reliability, research depth, document work, token usage, cache economics, and total cost against Opus and your current production model.
Start FreeAnthropic positions Fable 5.1 for demanding workloads rather than as the default model for every request. Test it where your existing frontier model still misses the acceptance threshold.
Common Questions
What is Claude Fable 5.1?
Claude Fable 5.1 is Anthropic’s premium model for demanding reasoning and long-horizon agentic work. Anthropic highlights stronger long-running coding, multistep research, and complex document, spreadsheet, and slide workflows.
How much does Claude Fable 5.1 cost?
Standard pricing is $10 per 1M input tokens and $50 per 1M output tokens. Five-minute cache writes cost $12.50/MTok, one-hour cache writes cost $20/MTok, and cache reads cost $0.25/MTok. Batch API input and output receive a 50% discount.
What is the Claude Fable 5.1 context window?
Claude Fable 5.1 has a 1,000,000-token context window by default and supports up to 128,000 output tokens per request.
What is Claude Fable 5.1’s knowledge cutoff?
Anthropic lists June 2026 as both the reliable knowledge cutoff and training-data cutoff for Claude Fable 5.1.
Does Claude Fable 5.1 support images?
Yes. Claude Fable 5.1 accepts text and image input and produces text output.
Does Claude Fable 5.1 use adaptive thinking?
Yes. Anthropic lists adaptive thinking as always on for Fable 5.1, with high as the default effort level. Per-message effort is also available as a beta feature.
How is Claude Fable 5.1 different from Claude Fable 5?
The two models share $10/$50 base token pricing, a 1M context window, and 128K output. Fable 5.1 adds stronger long-running coding, research, and knowledge-work capabilities, reduces cache-read pricing from $1 to $0.25 per million tokens, adds several beta controls, and introduces tool-choice and thinking-block migration changes.
Should I choose Claude Fable 5.1 or Claude Opus 5?
Anthropic recommends starting most workloads with Claude Opus 5. Fable 5.1 is intended for demanding reasoning and long-horizon agentic work, especially when evaluations on Opus 5 at higher effort still fall short.
Can Claude Fable 5.1 force a specific tool call?
No. Anthropic’s migration guide states that automatic tool choice and no-tool mode are supported, but forcing any tool or a named tool returns a 400 error.
Does Claude Fable 5.1 require special data retention?
Anthropic documents a 30-day data-retention requirement for Fable 5.1 unless an exception is expressly authorized. Organizations with zero-data-retention arrangements should verify compatibility with Anthropic before adoption.
What is Claude Mythos 5.1?
Claude Mythos 5.1 shares Fable 5.1’s capabilities, specifications, and pricing but is available only by invitation through Anthropic’s Project Glasswing trusted-access program.
Model information
Last updated
Specifications, pricing, context, thinking behavior, migration changes, and capability positioning on this page are based on Anthropic’s official Claude Fable 5.1 model documentation, migration guide, and September 2026 launch materials. Provider behavior and beta features may change over time.