Claude Fable 5.1 Release: Context and Pricing

Breaking Down the Claude Fable 5.1 Release: 1M Context, 128K Output, and 75% Cheaper Cache Reads

Claude Fable 5.1 launched on September 1, 2026. The rollout began with Claude Code sightings and an official model-list entry before Anthropic's developer announcement confirmed availability through Claude Code and the Claude Platform.

TLDR Claude Fable 5.1 is generally available with a confirmed 1M-token context window, a 128K maximum output, adjustable thinking effort, and stronger long-running agent behavior. The price is $10 per million input tokens, $50 per million output tokens, and $0.25 per million cached input tokens. Cache reads are 75% cheaper than Fable 5. Anthropic's launch materials also report substantial benchmark gains, including 52.6% on Terminal-Bench-Science 0.1 and 55.8% on Terminal-Bench 4.0. The parameter count, latency, rate limits, and detailed evaluation methodology remain open.

Key Takeaways

• Claude Fable 5.1 launched on September 1, 2026, after first being observed rolling out in Claude Code at 17:33 UTC.
• The confirmed public model ID is claude-fable-5-1.
• The model has a 1M-token context window and 128K tokens of maximum output.
• Pricing is $10 per 1M input tokens, $50 per 1M output tokens, and $0.25 per 1M cached input tokens.
• Cache reads are 75% cheaper than Fable 5. Anthropic estimates token-billed workloads will cost about 25% less typically and up to 45% less for highly agentic work.
• Published results include 52.6% on Terminal-Bench-Science 0.1, 55.8% on Terminal-Bench 4.0, and 73.4% on CursorBench 3.2 at maximum effort.

What Was Actually Shipped

The first public rollout signal arrived at 17:33 UTC. LuminaBench reported that “Fable 5.1 is rolling out in Claude Code right now.” The initial rollout sighting was the first timestamp in the supplied launch sequence, rather than the complete release record.

Evidence accumulated quickly. At 17:56 UTC, a German user said Fable 5.1 was live in their account. At 18:02 UTC, Alvaro Cintas posted a screenshot and wrote that the model “looks” released. Those reports showed that access was reaching users, although they did not by themselves establish a complete regional rollout.

The strongest metadata claim arrived at 18:05 UTC. LuminaBench said Fable 5.1 had appeared on Anthropic’s official model list under the first-party identifier claude-fable-5-1. The model-list report includes an image of the relevant entry.

Source: @LuminaBench

At 18:13 UTC, ClaudeDevs stated that Fable 5.1 was live in Claude Code and the Claude Platform. The same post instructed developers to select it from the model picker or use claude-fable-5-1. That developer announcement is the clearest developer-facing availability signal in the bundle.

The launch was subsequently reflected across additional surfaces. Fable 5.1 became generally available through the API and Claude Platform, was available in Claude and Claude Code, and was offered through Amazon Bedrock. Platform announcements also placed it in Cursor, OpenRouter, GitHub Copilot, and eligible Microsoft Copilot products.

The shape of the event matters. The rollout moved from a Claude Code observation to first-party model metadata, a developer-facing announcement, official launch materials, and published capability information within the same release window. The release record now includes a model ID, context and output limits, pricing, benchmark results, and access guidance.

What this tells us: Fable 5.1 launched on September 1, a public model identifier is available, and developer-facing access is active across multiple surfaces. What it doesn’t: the supplied evidence still does not establish every regional endpoint, rate-limit rule, latency figure, or model-size detail.

Why This Release Matters for AI Engineers

Fable 5.1 is positioned around sustained work rather than short conversational exchanges. The release materials mention agentic coding, research, documents, spreadsheets, slides, computer use, and long-horizon tasks. That list covers the workflows where context retention and recovery from intermediate failures can matter more than a single-turn answer.

The 1M-token context window is significant for repository-scale and document-heavy tasks. It allows an agent to retain requirements, tool outputs, source material, and prior decisions in one working session. Context size alone does not prove retrieval quality or task reliability. It establishes the amount of information the model can accept.

The 128K maximum output is also relevant to agent orchestration. A long output limit can support large patches, detailed research reports, or multi-stage planning. It can also increase latency and token consumption when prompts fail to constrain the response. The practical value depends on how often the model uses that ceiling and how accurately it stops.

The cost signals are now confirmed rather than speculative. ClaudeDevs described Fable 5.1 as having the same base price as Fable 5, with API cache reads 75% cheaper. Anthropic's launch materials estimate that typical token-billed workloads will cost about 25% less and that highly agentic workloads can save up to approximately 45%. The pricing comparison report supplies screenshots from the release period but does not define the workload or accounting method behind the range.

This creates a useful distinction for budgeting. A 75% reduction in cache-read cost can materially change long-running workflows with repeated context. It may have less effect on workloads dominated by fresh input and generated output. The 25% to 45% estimates apply where usage is billed by token; they do not represent a reduction to Claude subscription prices or limits. Claude Code subscriptions remain unchanged, and Fable 5.1 requires extra usage credits on Claude.

What We Can Reasonably Expect

The release is a point update within the Fable 5 line. The 5.1 label follows the 5 predecessor, so the public naming indicates an incremental refinement rather than a separately named model family. The name alone does not reveal whether training, inference, safety, or routing changed.

Long-Horizon Work is the most consistent capability theme. In this article, the term means completing multi-stage tasks while preserving goals, intermediate results, and tool state for longer periods. ClaudeDevs says Fable 5.1 gets further into a long task before it needs user input and is better at saying when it is stuck. Anthropic's launch materials position it for work that can run for hours across software projects, research, and enterprise workflows.

Adaptive Thinking refers to the adjustable effort setting. Fable 5.1 has always-on adaptive thinking with adjustable effort. The available levels are low, medium, high, xhigh, and max, with no option to turn reasoning off entirely in the reported interface. The signal set does not establish a complete latency or token-use profile for each level.

Stuck-State Signaling describes the model's ability to tell users when it is stuck. This is a valuable agent behavior because a system that announces failure can be easier to supervise than one that silently loops. The claim is part of the developer release description, but it should still be tested against stalled tool calls, contradictory requirements, and incomplete repositories.

Natural Writing Style captures the developer report that Fable 5.1's writing feels more natural and concise. This may matter for research summaries, business documents, spreadsheets, and slides. A style preference cannot substitute for factuality, citation quality, or reliable tool execution.

Cache-Read Economics is the clearest operational change in the release. Cache reads fall from $1 to $0.25 per million tokens, a 75% reduction. Cache behavior varies by prompt structure and application architecture, so teams should measure cache utilization rather than apply the percentage to every request.

Safety and Security are also part of the release. Fable 5.1 can look for vulnerabilities in a user's own source code, while cyber-related fallbacks to Opus are reported to be down about 40% from Fable 5 today and 55% from where Fable 5 started. Fable 5.1 remains the generally accessible configuration with additional safeguards, while Mythos 5.1 retains broader cyber and biology capabilities under restricted access.

On the current evidence, Claude Fable 5.1 targets longer autonomous workflows more directly than ordinary chat and has published benchmark results showing large gains over Fable 5 on several tasks. The size and generality of those gains still depend on evaluation conditions and real-world workload testing.

Claude Fable 5.1 vs Claude Mythos 5.1: What the Signal Says

Claude Mythos 5.1 launched alongside Fable 5.1. Anthropic describes the two as the same model with different levels of safeguards, and the supplied release signals describe them as sharing identical model weights. Fable 5.1 is generally available. Mythos 5.1 is available through trusted-access programs intended for cybersecurity and life-sciences work.

The available comparison is therefore about positioning and safeguards rather than a separate model architecture:

Dimension	Claude Fable 5.1	Claude Mythos 5.1
Context window	1M tokens confirmed	1M tokens for the same underlying model
Maximum output	128K tokens confirmed	128K tokens for the same underlying model
Availability	Generally available to API customers and across multiple products	Limited to trusted-access programs for cybersecurity and life-sciences work
Main task emphasis	Agentic coding, research, documents, spreadsheets, slides, and computer use	Cybersecurity and life-sciences research with broader safeguards
Official benchmarks	52.6% Terminal-Bench-Science; 55.8% Terminal-Bench 4.0; 73.4% CursorBench at maximum effort	No separate benchmark score confirmed in the signal set

The distinction is access and safety scope, not a published benchmark hierarchy. Any claim that Mythos 5.1 is faster, smarter, or more capable than Fable 5.1 would exceed the available evidence. The two configurations share weights, but the bundle does not provide a controlled head-to-head test.

How to Evaluate Fable 5.1 Without Overreading a Benchmark

The correct evaluation target is not a single leaderboard score. It is the behavior of the model inside the workflows named in the release materials, with the published scores used as reference points rather than as a substitute for local testing.

For coding, use a repository task with explicit requirements, tests, and a defined stopping condition. Record whether the model asks for help, reports a stuck state, edits the right files, and leaves the test suite in a usable state. Compare those observations with Fable 5 under the same instructions. The official Terminal-Bench 4.0 result is 55.8% for Fable 5.1 versus 42.0% for Fable 5; that establishes a measured comparison, but not how the same gap will appear in every repository.

For research and knowledge work, test a fixed packet containing documents, spreadsheets, and slide material. Track factual errors, omitted requirements, unsupported claims, and the number of human interventions. The 1M-token context window should be evaluated with long inputs, but context capacity should be separated from context recall.

For cost, measure fresh input tokens, cached input tokens, output tokens, and total elapsed time separately. The prices are $10 per 1M input tokens, $50 per 1M output tokens, and $0.25 per 1M cached input tokens. Those units make a workload-level comparison possible without assuming that every prompt receives the 75% cache-read discount.

The published benchmark set gives teams several additional anchors. Fable 5.1 scores 52.6% versus 24.7% for Fable 5 on Terminal-Bench-Science 0.1, 31.4% versus 17.1% on AutomationBench, and 73.4% versus 70.5% on CursorBench 3.2 according to the supplied benchmark reports. Cursor reports the 73.4% result at maximum effort and highlights the model's ability to verify its own work.

Teams that need a stable comparison point can use the separately cataloged Claude Fable 5 model as a predecessor baseline, while keeping model access and evaluation conditions consistent.

The remaining benchmark gap is methodological rather than numerical. The supplied evidence does not provide complete test configurations, sample sizes, evaluation dates, or enough detail to independently reproduce every reported score.

What We Know vs. What We Don't

The following separation is deliberately strict. “Known” means directly reported in the supplied launch signals or official release materials. It does not mean every claim has been independently audited.

What we know

• Claude Fable 5.1 launched on September 1, 2026. It was first publicly observed rolling out in Claude Code at 17:33 UTC. The earliest rollout report provides the first timestamp in the supplied sequence.
• The identifier claude-fable-5-1 appeared in Anthropic’s first-party public model metadata and is the model ID used on the Claude Platform.
• Fable 5.1 is generally available through the Claude Platform and API, Claude, Claude Code, and Amazon Bedrock. It is also available through reported integrations including Cursor, OpenRouter, GitHub Copilot, and eligible Microsoft Copilot products.
• The model has a 1M-token context window and a 128K maximum output. It supports always-on adaptive thinking with adjustable effort.
• The release focuses on agentic coding, scientific research, documents, spreadsheets, slides, computer use, and long-horizon work. Developer descriptions say it gets further into long tasks, is better at identifying when it is stuck, and has a more natural writing style.
• The usage-based price is $10 per 1M input tokens, $50 per 1M output tokens, and $0.25 per 1M cached input tokens. Cache reads are 75% cheaper than Fable 5.
• Anthropic estimates token-billed costs will be about 25% lower for typical workloads and up to approximately 45% lower for highly agentic workloads. These estimates do not reduce Claude subscription prices or limits.
• Published benchmark results include 52.6% versus 24.7% for Fable 5 on Terminal-Bench-Science 0.1, 55.8% versus 42.0% on Terminal-Bench 4.0, 31.4% versus 17.1% on AutomationBench, and 73.4% versus 70.5% on CursorBench 3.2.
• Claude Fable 5.1 and Claude Mythos 5.1 are configurations of the same model with identical model weights. Fable 5.1 has the broader public safeguards; Mythos 5.1 is limited to trusted-access programs for cybersecurity and life-sciences work.
• For new API accounts using Fable 5.1, the Messages API preserves the context associated with thinking blocks. Developers cannot edit the preceding system prompt, tools, or messages without triggering a mismatch; non-strict handling can remove the affected thinking blocks so a request can continue.
• Anthropic has introduced Enterprise Frontier Safeguards for eligible customers, with enterprise-controlled cloud infrastructure and phased availability beginning later in fall 2026. Until EFS is available, eligible customers can use Fable 5.1 with zero data retention.

What we don’t know

• The architecture is not fully disclosed. The supplied evidence confirms shared weights with Mythos 5.1 but does not provide a parameter count, training-data size, attention design, optimizer, mixture-of-experts structure, or inference-routing detail.
• Latency, throughput, and complete rate-limit rules remain unspecified in the supplied evidence.
• The benchmark methodology remains incomplete. The supplied posts do not establish every test configuration, sample size, evaluation date, or independent replication.
• The rollout is broad, but the evidence does not establish a complete geographic availability matrix or the behavior of every regional endpoint, account type, and product surface.
• Usage-limit complaints from subscription users remain anecdotal. The launch confirms unchanged subscription limits and extra usage-credit requirements, but the supplied evidence does not establish a complete rate-limit policy or typical consumption pattern.
• A purported Fable 5.1 system-prompt leak remains unresolved as a document-size claim. One post described more than 270,000 characters, while another analysis in the bundle measured a different length; the supplied evidence does not independently authenticate a definitive character count.
• The launch materials do not provide a separate complete specification sheet or separate benchmark set for Claude Mythos 5.1.

Frequently Asked Questions

Where did Claude Fable 5.1 first appear?

Claude Fable 5.1 first appeared publicly in a Claude Code rollout observed at 17:33 UTC on September 1, 2026. Anthropic subsequently announced the model and listed it under the identifier claude-fable-5-1.

What model ID is associated with Claude Fable 5.1?

The confirmed model ID for Claude Fable 5.1 is claude-fable-5-1.

Where is Claude Fable 5.1 available?

Claude Fable 5.1 is generally available through the Claude Platform and API, Claude, Claude Code, and AWS services including Amazon Bedrock. It is also available through reported integrations including Cursor, OpenRouter, GitHub Copilot, and eligible Microsoft Copilot products.

What context and output limits does Claude Fable 5.1 have?

Claude Fable 5.1 has a 1M-token context window and a maximum output of 128K tokens.

What are Claude Fable 5.1's prices?

Usage-based pricing is $10 per million input tokens, $50 per million output tokens, and $0.25 per million cached input tokens. Cache reads are 75% cheaper than Fable 5, and Anthropic estimates token-billed workloads will cost about 25% less typically and up to 45% less for highly agentic workloads.

Has Anthropic published Claude Fable 5.1 benchmark results?

Yes. The launch materials report 52.6% on Terminal-Bench-Science 0.1 and 55.8% on Terminal-Bench 4.0. Cursor separately reports 73.4% on CursorBench 3.2 at maximum effort.

What is known about Claude Fable 5.1's architecture?

Claude Fable 5.1 and Claude Mythos 5.1 are two configurations of the same model and share identical model weights. The supplied evidence does not confirm the parameter count, training-data size, attention design, optimizer, or inference-routing details.

How does Claude Fable 5.1 compare with Claude Fable 5 on benchmarks?

Fable 5.1 scores 52.6% versus 24.7% for Fable 5 on Terminal-Bench-Science 0.1, 55.8% versus 42.0% on Terminal-Bench 4.0, and 31.4% versus 17.1% on AutomationBench. Cursor reports 73.4% versus 70.5% on CursorBench 3.2.

Is Claude Fable 5.1's rollout global?

Claude Fable 5.1 is generally available, and access has been reported in Germany and through multiple platforms. The supplied evidence does not establish a complete geographic rollout or regional endpoint matrix.

What is known about Claude Mythos 5.1's access and specifications?

Claude Mythos 5.1 shares Fable 5.1's underlying model weights but has different safeguards. It is available through trusted-access programs for cybersecurity and life-sciences work, while the supplied evidence does not provide a separate complete specification sheet.

How should the different Claude Fable 5.1 pricing claims be interpreted?

The $10 input and $50 output prices remain unchanged from Fable 5, while cache reads fell from $1 to $0.25 per million tokens. Anthropic's 25% typical and up-to-45% highly agentic savings estimates apply where usage is billed by token, not as a reduction to Claude subscription prices or limits.

What Builders Should Do Today

First, use the confirmed model ID in an explicit, logged integration. The identifier claude-fable-5-1 is available in the named developer surfaces, but the supplied evidence does not document every compatibility detail. Pin the model explicitly, log responses, and keep a fallback to Fable 5 during the initial rollout.

Second, test intervention behavior rather than only answer quality. The most consequential capability claims concern long-horizon work and stuck-state signaling. Give the model tasks that require tool use, intermediate decisions, and recovery from incomplete information. Measure how often a human must intervene and whether the model identifies failure before wasting additional tokens.

Third, separate cache savings from total savings. Build a token ledger for fresh input, cached input, and output. The $0.25 per 1M cached-input figure could matter greatly for repeated agent context, while the $50 per 1M output figure could dominate verbose workflows. A single blended cost figure would hide that difference.

A fourth practical step is to preserve the old prompt and tool contract. Fable 5.1 may respond better to the prompting guide and updated claude-api skill mentioned in the developer announcement, but changing the model, prompt, tools, and evaluation set at the same time would make attribution difficult. The developer migration note points to those updated resources without supplying their full contents in the signal bundle.

Finally, account for subscription behavior separately from API economics. The cache-read reduction applies to token-billed usage, while Claude subscription prices and limits remain unchanged and Fable 5.1 requires extra usage credits. API cost projections should therefore not be used as a direct estimate of Claude Code or Claude subscription capacity.

The Week Ahead

The next useful evidence should reduce uncertainty around the parts of the launch that remain open. Watch for fuller documentation of latency, rate limits, regional endpoint behavior, and the evaluation configurations behind the published benchmark scores. Run a controlled coding and document evaluation against Fable 5 before assuming that benchmark gains will transfer directly to a particular production workflow.

Also watch whether independent builders publish reproducible results. A benchmark with task definitions, prompts, tool settings, token counts, and failure cases would be more valuable than another “better across the board” post. The current evidence supports a launched model, broad agentic-work positioning, lower cache-read pricing, and clear measured gains on several reported evaluations. It does not yet provide a complete capability, efficiency, or model-architecture profile.

Building similar long-horizon chat agents? On kie.ai you can try Claude Opus 5.5, Claude Sonnet 5.5, and Claude Fable 5.