Anthropic Launches Claude Fable 5.1 With Lower Agent Costs and a Restricted Mythos Variant

Anthropic released Claude Fable 5.1 on September 1 as its most capable generally available model for long-running coding, research, and knowledge-work agents. The company simultaneously introduced Claude Mythos 5.1, an invitation-only version of the same underlying model with more permissive safeguards for approved cybersecurity and life-sciences organizations.

Fable 5.1 retains the previous generation’s base API prices of $10 per million input tokens and $50 per million output tokens. The consequential pricing change is a 75% reduction in cache-read costs, from $1 to $0.25 per million tokens. Anthropic estimates that this lowers the total cost of typical Fable workloads by about 25% and highly agentic workloads by as much as 45%.

The release also revises the safeguards that made the original Fable 5 difficult to use for some legitimate biology and security tasks. Fable 5.1 can now identify vulnerabilities in source code, while penetration testing, exploit generation, binary vulnerability scanning, and advanced life-sciences research remain subject to restrictions or routing to less capable models.

1. Fable 5.1 Targets Long-Horizon Agent Work

Fable 5.1 succeeds Fable 5, which Anthropic introduced on June 9, 2026 for complex tasks that could run asynchronously for days. The new model preserves that emphasis but improves performance on coding, computer use, scientific research, business workflows, and multidisciplinary reasoning.

The API model identifier is claude-fable-5-1. It accepts text and images and produces text, with a one-million-token context window and a maximum output of 128,000 tokens. Adaptive thinking is always enabled, and the model’s documented reliable-knowledge and training-data cutoffs are June 2026.

Anthropic describes Fable 5.1 as its option for demanding reasoning and long-horizon agents, rather than the default choice for every application. Its own model-selection guidance recommends starting most workloads with the cheaper Claude Opus 5 and moving to Fable 5.1 when evaluations at higher Opus effort levels remain insufficient.

The model is available through Claude’s Pro, Max, Team, and Enterprise plans. Developers can access it through the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Anthropic also offers US-only inference at 1.1 times the standard input and output prices.

Fable 5.1 defaults to high effort in Claude Code but medium effort in Claude Cowork and Claude.ai. Anthropic says lower effort settings can approach or exceed Fable 5’s performance at a lower cost, giving developers another control over latency and token consumption.

2. Anthropic Reports Broad Benchmark Gains

Anthropic’s published evaluations show Fable 5.1 improving over Fable 5 across every benchmark in its launch comparison. These are vendor-reported results, and several use recently revised tests or Anthropic’s own evaluation configuration.

On Terminal-Bench-Science 0.1, which measures agentic scientific work in a terminal environment, Fable 5.1 scored 52.6%, compared with 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol. Anthropic reports a standard error of between 3.5 and 4.5 percentage points per model.

Fable 5.1 scored 55.8% on Terminal-Bench 4.0 for agentic coding. Mythos 5.1 reached 60.9% on the same test, while Fable 5 scored 42.0%, Opus 5 scored 52.3%, and GPT-5.6 Sol scored 37.3%. Because Fable and Mythos use the same underlying model, Anthropic attributes the gap between their results to tasks where Fable’s security safeguards intervened.

On CursorBench 3.2.0, Fable 5.1 scored 73.4%, ahead of Fable 5 at 70.5%, Opus 5 at 70.0%, and GPT-5.6 Sol at 67.2%. Its AutomationBench score rose to 31.4%, from 17.1% for Fable 5. The model also reached 65.0% on Humanity’s Last Exam when tools were enabled, compared with 63.8% for Fable 5 and 63.6% for Opus 5.

Anthropic evaluated Fable 5.1 with its production safeguards active. It assigned zero credit on certain OSWorld tasks when those safeguards intervened, while some other flagged cyber and biology tasks were completed by fallback Opus models. The company therefore cautions that the published results reflect both the underlying model and its deployed routing system.

3. Fable and Mythos Share Intelligence but Not Access

Claude Mythos 5.1 is not a more powerful base model than Fable 5.1. Anthropic says the two are identical at the model level; the distinction is that Mythos applies more permissive safeguards for vetted organizations whose professional work would otherwise trigger Fable’s restrictions.

Mythos 5.1 has the same one-million-token context window, 128,000-token maximum output, adaptive thinking, and $10-per-million-input and $50-per-million-output pricing. Its API identifier is claude-mythos-5-1, but access is invitation-only and currently limited to selected US organizations.

Life-sciences access is being organized through the Life Sciences Verification Program, developed with the US government. Anthropic has enrolled an initial group and says broader enrollment is planned. The program relaxes biology restrictions for approved research and development while retaining the model’s other safeguards.

The Cyber Verification Program currently gives qualified defenders access to certain Opus- and Sonnet-class models with reduced cyber restrictions. Anthropic says Mythos-class access will be added in the near future. Claude Security, the company’s codebase-scanning product, is already running on Mythos 5.1.

This two-tier deployment lets Anthropic distribute the model’s general coding and reasoning capabilities broadly while keeping its least restricted cyber and biological capabilities behind institutional verification. It also means benchmark results for Mythos cannot automatically be treated as results available to ordinary Fable users.

4. Safeguards Now Produce Fewer False Positives

Fable 5 introduced unusually restrictive classifiers for biology and cybersecurity. When a request triggered them, Claude could switch to an Opus model rather than complete the task with Fable. Anthropic acknowledges that the original implementation blocked many benign medical, educational, and defensive-security requests.

The updated biology safeguards intervene about 85% less often on benign elementary biology and medical prompts than the classifiers shipped with Fable 5. Everyday health questions, clinical assistance, and biology education should therefore encounter fewer model substitutions. Research involving areas such as virology, toxicology, or molecular design can still be routed to Opus unless the user qualifies for Mythos access.

In cybersecurity, Claude Code users are expected to experience approximately 60% fewer safeguard interventions per session. Fable 5.1 may inspect source code to identify vulnerabilities, but it still redirects penetration testing, exploit development, and binary-based vulnerability scanning.

Axios independently reported that these interventions had become a practical customer concern, with some developers seeking to reduce their dependence on Claude because legitimate work was being interrupted. The changes therefore affect more than refusal rates: fewer unexpected fallbacks should make coding-agent behavior and cost estimates more predictable.

Anthropic has also added anti-distillation restrictions. New API accounts can no longer edit Claude’s earlier context while preserving the model’s previous thinking transcript. Existing accounts are initially exempt, but Anthropic says the restriction will apply to all users with future model releases, potentially requiring changes to integrations that rewrite conversation history.

5. Cache Pricing and Enterprise Data Controls Change Deployment Economics

The standard input and output rates have not fallen, but cache reads are particularly important for agents that repeatedly reuse long instructions, repository context, tool results, or accumulated task history. Cutting that price to $0.25 per million tokens disproportionately benefits long sessions in which cached context accounts for most token consumption.

Anthropic calculated its 25% typical-workload and 45% highly agentic savings estimates from four weeks of real usage during August 2026. Actual savings will depend on how frequently an application reuses cached context; workloads dominated by new input or generated output will see a smaller reduction.

Data retention remains another deployment constraint. Fable and Mythos require 30-day retention for safety monitoring by default. Anthropic says this monitoring is intended to detect misuse patterns spread across multiple sessions or accounts, not to train models on enterprise data without permission.

A new Enterprise Frontier Safeguards system is intended to reconcile that monitoring with zero-data-retention requirements. Under EFS, activity data can remain in cloud infrastructure controlled by the customer, using the customer’s encryption keys, access policies, and audit logs. Automated monitoring can flag suspicious patterns, while human review is performed by the customer by default rather than Anthropic.

EFS was designed with more than 100 organizations and is scheduled for a phased rollout beginning in fall 2026. Eligible customers may use Fable 5 and Fable 5.1 under zero-data-retention terms until the system is available. Anthropic does not charge separately for EFS, although customers remain responsible for cloud storage and data-transfer charges.

Fable 5.1 and Mythos 5.1 also introduce invisible text watermarking to comply with the European Union’s transparency rules for models released after August 2, 2026. Anthropic says the watermark changes token-selection randomness without adding hidden characters, extra tokens, or user-identifying data. Detection is initially available through a private-preview API for eligible organizations.

Frequently Asked Questions

Is Claude Fable 5.1 generally available?

Yes. It is available through paid Claude plans, the Claude API, and supported cloud platforms. The API model identifier is claude-fable-5-1.

Is Claude Mythos 5.1 a different model?

No. Anthropic says Fable 5.1 and Mythos 5.1 use the same underlying model. Mythos has more permissive safeguards and is restricted to vetted organizations.

How much does Fable 5.1 cost through the API?

Standard pricing is $10 per million input tokens and $50 per million output tokens. Cache reads cost $0.25 per million tokens.

What are the context and output limits?

Fable 5.1 supports a one-million-token context window and up to 128,000 output tokens per request.

Can Fable 5.1 perform penetration testing?

Its safeguards still redirect penetration testing, exploit generation, and binary-based vulnerability scanning. It can identify vulnerabilities in source code for defensive purposes.

Sources

• Original announcement post by Mike Krieger on X
• Anthropic: Introducing Claude Fable 5.1 and Claude Mythos 5.1
• Claude Platform documentation: Claude Fable 5.1
• Anthropic: Claude Fable product and availability details
• Anthropic: Claude Mythos access and safeguard details
• Anthropic: Developing Enterprise Frontier Safeguards with customers
• Anthropic: How Claude’s text watermarking works
• Axios: Anthropic releases new models, cost structures and safeguards