Most complex production work
Opus 5.5 costs $4 input / $20 output per million tokens and is Anthropic’s recommended starting point for most coding, analysis and agentic workloads.
Open Opus 5.5 reference ↗MODEL INTELLIGENCE
Claude Fable 5.1
model reference.
Claude Fable 5.1 reference: 1M context, 128K output, $10/$50 API pricing, adaptive thinking, effort levels, coding and agent capabilities, migration, benchmarks and availability.
1,000,000context tokens
128,000max output
September 1, 2026released
MODEL REFERENCE
Claude Fable 5.1 at a glance.
Claude Fable 5.1 is Anthropic’s highest-capability model generally available to all customers. It is designed for demanding reasoning, long-horizon agentic coding, multistep research and complex knowledge work where a model must remain coherent across many tool calls, documents or hours of execution.
The important positioning detail is that Anthropic does not recommend Fable 5.1 as the default for every workload. Its current guidance says most applications should start with Claude Opus 5.5, then move to Fable 5.1 when higher-effort Opus evaluations still fall short. That makes Fable 5.1 a quality-first escalation tier rather than a universal replacement for cheaper Claude models.
Model ID: claude-fable-5-1 · Released September 1, 2026 · Active / latest Fable · Retirement not sooner than September 1, 2027 · Last verified 2026-09-26
MODEL IDclaude-fable-5-1
STATUSActive / latest
CONTEXT WINDOW1,000,000tokens
MAX OUTPUT128,000tokens
STANDARD INPUT$10/ 1M tokens
STANDARD OUTPUT$50/ 1M tokens
QUICK DECISION
Fable 5.1 vs Opus 5.5: when should you pay more?
The most useful way to understand Fable 5.1 is against the model Anthropic recommends for most workloads: Claude Opus 5.5.
When quality still falls short
Use Fable when higher-effort Opus 5.5 still misses the acceptance bar on difficult reasoning, long-running coding, multistep research or complex artifact work.
When latency and economics matter
Fable is slower and 2.5× the headline input/output token price of Opus 5.5. If task success is similar, Opus is usually the more efficient production choice.
When one hard task is worth more than token cost
Fable makes sense when solving a difficult issue correctly is materially more valuable than minimizing inference spend—for example root-cause analysis, deep research or high-value engineering work.
CLAUDE LINEUP
Where Fable 5.1 sits in the current Claude family.
Anthropic separates its current lineup by capability, speed and cost rather than presenting one model for every task.
Claude Fable 5.1
Demanding reasoning and long-horizon agentic work. Use when cheaper Claude tiers fail your evaluations.
1M context · 128K outputAdaptive thinking always on · default effort high
$10input / 1M$50output / 1M
Claude Opus 5.5
Long-running agentic coding and knowledge work with a stronger price-performance balance for most complex tasks.
1M context · 128K outputAdaptive thinking always on · default effort medium
$4input / 1M$20output / 1M
Open reference ↗
Claude Sonnet 5
Everyday coding, agents, analysis and enterprise workflows where lower latency and cost matter more.
1M context · 128K outputAdaptive thinking · default effort high
$2input / 1M$10output / 1M
Claude Haiku 4.5 remains the fastest lower-cost tier at $1 input / $5 output per million tokens with a 200K context window and 64K max output.
WHAT CHANGED
Fable 5.1 improves capability without raising base token prices.
Anthropic kept Fable 5.1’s standard input and output prices at the same $10 / $50 per million tokens as Fable 5, but cut prompt-cache reads from $1 to $0.25 per million tokens. That matters disproportionately for long-running agents because they repeatedly reuse large prefixes such as repository context, tool definitions, instructions and accumulated task state.
Anthropic estimates that the new model costs about 25% less than Fable 5 for typical token-billed workloads and can save up to roughly 45% on highly agentic work. Those are vendor measurements, not universal savings: actual economics depend on cache hit rate, effort, prompt size, tool loops and how quickly the model completes the task.
The capability update also targets a specific failure mode of long-running AI work: models that look strong on short benchmarks but lose coherence over many steps. Anthropic’s release emphasizes long-horizon coding, multistep research, complete document/spreadsheet/slide work and extended unattended tasks.
CAPABILITIES
What Claude Fable 5.1 is built to do.
Hours-long engineering work
Anthropic positions Fable 5.1 for demanding software engineering that spans large repositories, repeated tool calls, verification loops and long execution windows.
Multistep investigation
The model is designed to gather evidence, revise hypotheses and continue through research workflows rather than stopping after a shallow first-pass synthesis.
Finished artifacts
Anthropic highlights complex documents, spreadsheets and slide decks where the model must carry analysis through to a complete deliverable instead of only generating notes.
Research-oriented problem solving
Anthropic reports gains on agentic scientific research and describes experiments in computational analysis, software optimization and multidisciplinary reasoning.
Text and image input
Fable 5.1 accepts text and images and returns text, with the same broad Claude vision and tool-use architecture used across current high-end Claude models.
1M-token context
The context window supports large repositories and document corpora, but long context remains an engineering budget: retrieval quality, caching and context discipline still affect results and cost.
ADAPTIVE THINKING
Thinking is always on; effort controls depth.
Claude Fable 5.1 uses adaptive thinking on every request. The model decides when and how much internal reasoning to use, while the effort parameter controls the overall capability, latency and cost tradeoff.
Fable 5.1 supports low, medium, high, xhigh and max effort. Anthropic recommends starting at high, which is also the API default, then evaluating other levels against the workload. For routine or latency-sensitive work, medium or low can reduce cost. For the most capability-sensitive coding and agent tasks, xhigh or max may improve results.
One important operational detail is that effort labels are not directly comparable across model generations. Anthropic recommends re-running the effort sweep even if you already optimized Fable 5, because medium on Fable 5.1 can approximately match older Fable performance on some workloads at lower cost.
PRICING
Claude Fable 5.1 API pricing.
Standard token prices are unchanged from Fable 5, but prompt caching is materially cheaper.
Cache economics are unusually important
At $0.25 per million tokens, Fable 5.1 cache reads are 2.5% of its standard input price. Long-running agents that reuse stable context can therefore have a very different effective cost profile from one-shot requests that repeatedly pay full input price.
Thinking tokens count as output
Always-on adaptive thinking means a difficult task can consume substantial output-token budget even when the visible final answer is short. Set max output high enough for higher effort, then measure total task cost rather than visible-response length.
BENCHMARK EVIDENCE
Anthropic reports large gains on long-horizon and agentic tasks.
These are vendor-reported benchmark results from Anthropic’s Fable 5.1 launch material. They are useful evidence about where to test the model, not a substitute for independent or workload-specific evaluation.
52.6%
Anthropic reports 52.6% for agentic scientific research, compared with 24.7% for Fable 5 and 29.0% for Opus 5 in its reproduced setup.
55.8%
Anthropic reports 55.8% for Fable 5.1 on agentic terminal coding, with Mythos 5.1 higher in the same table under different safeguards.
73.4%
The launch page reports 73.4% for Fable 5.1 versus 70.5% for Fable 5, 70.0% for Opus 5 and 67.2% for GPT-5.6 Sol in Anthropic’s comparison.
31.4%
Anthropic reports 31.4% on business workflows, compared with 17.1% for Fable 5 and 26.9% for Opus 5 in the same evaluation table.
65.0% with tools
Anthropic reports 60.9% without tools and 65.0% with tools for multidisciplinary reasoning under its evaluation setup.
77.9% partial
Anthropic reports 77.9% partial and 41.7% strict on computer use, with production safeguards enabled and explicit caveats about safeguard interventions.
How to use these numbers
Do not turn a benchmark table into a universal ranking. Re-run your own repository, documents, tools and acceptance criteria at the effort levels you can afford. For agent systems, measure task completion, retries, tool errors, latency, human correction and total cost.
AGENT OPERATIONS
Fable 5.1 adds controls for long-running workflows.
Several of the most useful changes are not benchmark features; they are controls for keeping long conversations and tool loops efficient and understandable.
PER-MESSAGE EFFORTBeta
Change effort later in a conversation without invalidating the prompt cache, useful when only some stages need maximum reasoning.
TURN-SCOPED SYSTEM MESSAGESBeta
Add guidance for one turn without rewriting earlier conversation history, helping preserve both cache validity and thinking state.
PROGRESS UPDATESReadable between tool calls
With the appropriate beta display mode, applications can surface concise model progress during long tool chains instead of leaving users with a silent agent.
CONTENT PROVENANCENew in 5.1
Anthropic lists content provenance among the additive Fable 5.1 changes, supporting stronger traceability in long agent workflows.
PARALLEL TOOLSPrompt for independence
Anthropic’s prompting guidance recommends explicitly asking the model to request independent tool inputs together when coding or computer-use loops become unnecessarily serial.
APPEND-ONLY HISTORYImportant for thinking state
Long conversations should avoid rewriting earlier turns because edits can invalidate cached context and later thinking blocks.
MIGRATION
Moving from Claude Fable 5 to Fable 5.1.
The model uses the same Messages API family, but several behavior changes are breaking enough to deserve regression tests.
01Update the model ID.
Use claude-fable-5-1 on the Claude API. Cloud platform identifiers follow each provider’s documented naming convention.
02Do not force tool use with unsupported tool_choice modes.
Anthropic states that tool_choice types any and tool return a 400 error on Fable 5.1. Use auto tool selection plus strict tool use or structured outputs when you need schema-conformant calls.
03Preserve thinking blocks exactly.
Earlier Claude models cannot read Fable 5.1 thinking blocks, and editing prior conversation state can invalidate them. Pass live assistant turns back unchanged in tool loops.
04Keep conversation history append-only.
Changing the system prompt, tools or earlier messages after thinking blocks were produced can invalidate later thinking state and restart caching.
05Re-run effort evaluations.
Fable 5.1’s effort levels do not correspond one-to-one with Fable 5’s internal reasoning behavior. Test medium, high and higher levels against representative work.
06Update UX for long tool chains.
If users need visible progress, opt into readable progress-update thinking blocks and render them as status updates rather than exposing internal reasoning.
FABLE VS MYTHOS
Fable 5.1 and Mythos 5.1 share the model, not the access policy.
Anthropic describes Fable 5.1 and Mythos 5.1 as the same underlying model with different safeguards and availability.
Generally available
Designed for general coding, knowledge work and research with production safeguards suitable for broad customer access.
Trusted access only
Available through Project Glasswing for approved organizations whose cybersecurity or life-sciences work requires a different safeguard profile.
1M context · 128K output
Anthropic states that Mythos 5.1 shares Fable 5.1’s underlying capabilities, specifications and token pricing.
Policy and access differ
For production planning, model capability is only one dimension. Eligibility, safeguards and deployment controls differ even when the base model is the same.
AVAILABILITY
Where Claude Fable 5.1 is available.
claude-fable-5-1
Anthropic’s native API uses the fixed model ID claude-fable-5-1.
anthropic.claude-fable-5-1
Available through Amazon Bedrock with provider-specific identifiers and deployment controls.
Claude on Google Cloud
Anthropic lists Fable 5.1 as available on Google Cloud with the same core model specification.
Enterprise cloud access
Anthropic lists Microsoft Foundry among the supported deployment platforms for Fable 5.1.
SAFETY & ENTERPRISE
Capability and safeguards are both part of the deployment decision.
Anthropic reports that Fable 5.1 uses updated cybersecurity and biology safeguards intended to reduce false positives while retaining restrictions around higher-risk use. The company also reports stronger performance on an external prompt-injection benchmark for the underlying 5.1 model family, while acknowledging that very long-context and multi-agent settings remain areas where evaluation coverage is less complete.
For enterprise privacy, Anthropic announced Enterprise Frontier Safeguards, a system intended to let customers retain data in their own cloud infrastructure while enabling misuse detection. The rollout is phased. Organizations with strict retention requirements should verify current eligibility and terms directly with Anthropic or their cloud provider rather than relying on a static model page.
For agents, the operational rule remains simple: higher model capability should not automatically grant broader permissions. Tool scopes, credentials, execution environments, approval gates and audit trails belong outside the model.
WHEN TO USE FABLE 5.1
Use it when task value justifies the premium.
Long-running coding agents
Use Fable when an agent must work across repositories and tools for hours and your Opus 5.5 evaluation shows meaningful quality gaps.
Deep multistep research
Use it when a workflow must gather evidence, revisit assumptions and deliver a complete research artifact rather than a quick summary.
High-value professional analysis
Use it when the cost of an incomplete or incorrect answer is materially higher than the model’s inference premium.
Routine production traffic
For everyday coding, extraction, standard analysis or high-volume traffic, Opus 5.5, Sonnet 5 or another lower-cost model may deliver better economics.
LIMITS & CAVEATS
What to keep in mind.
01It is intentionally expensive.
At $10/$50 per million tokens, Fable 5.1 should be routed to tasks where its added capability produces measurable business or engineering value.
02It is slower than Opus 5.5 and Sonnet 5.
Anthropic labels Fable 5.1’s comparative latency as slower. Interactive workflows should test whether the quality gain justifies the wait.
03Always-on thinking affects cost.
Thinking tokens are part of the output budget. Higher effort and longer tasks can consume significantly more output tokens than the visible answer suggests.
04Long context is not free accuracy.
A 1M-token window is capacity, not guaranteed retrieval. Use caching, retrieval and context selection rather than blindly filling the window.
05Vendor benchmarks need local validation.
Anthropic’s launch results are informative but use specific harnesses, safeguards and effort settings. Production routing should be based on your own acceptance tests.
06Migration changes can break agent loops.
Forced tools, thinking-block preservation and conversation-history rules need regression tests before moving a long-running agent to 5.1.
OFFICIAL SOURCES
Anthropic · Introducing Claude Fable 5.1 and Mythos 5.1↗Claude Platform · Fable 5.1 model overview↗Claude Platform · Choosing the right model↗Claude Platform · Effort↗Claude Platform · Prompting Fable 5.1↗Claude Platform · Migration guides↗Claude Platform · Mid-conversation system messages↗Anthropic · Model system cards↗QUICK ANSWERS
What is Claude Fable 5.1?
Claude Fable 5.1 is Anthropic’s highest-capability model generally available to all customers. It is designed for demanding reasoning, long-horizon agentic work, coding, research and complex knowledge work.
What is the Claude Fable 5.1 model ID?
The Claude API model ID is claude-fable-5-1. Amazon Bedrock uses anthropic.claude-fable-5-1, while Google Cloud, Microsoft Foundry and Claude Platform on AWS use the claude-fable-5-1 name.
What is the Claude Fable 5.1 context window?
Claude Fable 5.1 has a 1 million-token context window and supports up to 128,000 output tokens.
How much does Claude Fable 5.1 cost?
Claude Fable 5.1 costs $10 per million input tokens and $50 per million output tokens. Five-minute cache writes cost $12.50 per million tokens, one-hour cache writes cost $20, and cache reads cost $0.25.
Does Claude Fable 5.1 use adaptive thinking?
Yes. Adaptive thinking is always on. The effort parameter is the primary control for trading off capability, latency and cost. Fable 5.1 supports low, medium, high, xhigh and max effort and defaults to high.
Is Claude Fable 5.1 better than Claude Opus 5.5?
Anthropic positions Fable 5.1 as the higher-capability model, but recommends starting with Opus 5.5 for most workloads. Fable 5.1 is intended for demanding reasoning and long-horizon agentic work when Opus 5.5 at higher effort still does not meet the required quality.
What is Claude Fable 5.1 best for?
Anthropic positions Fable 5.1 for long-running agentic coding, multistep research, difficult knowledge work, scientific research and complex document, spreadsheet and presentation tasks.
What is Claude Mythos 5.1?
Claude Mythos 5.1 uses the same underlying model and specifications as Fable 5.1 but is access-restricted through Project Glasswing and uses different safeguards for approved cybersecurity and life-sciences work.
Does Claude Fable 5.1 support prompt caching?
Yes. Cache reads cost $0.25 per million tokens, one quarter of Fable 5’s cache-read price according to Anthropic. This can materially reduce the cost of long-running agents that repeatedly reuse large context.
What changes when migrating from Claude Fable 5?
Key changes include unsupported forced tool choice types, stricter handling of thinking blocks and conversation-history edits, lower cache-read pricing, and new beta features such as per-message effort, turn-scoped system messages and readable progress updates.