Anthropic Releases Claude Fable 5.1 With 75% Cache Cost Cut

Anthropic Releases Fable 5.1 and Mythos 5.1

On 1 September 2026, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. Both are configurations of the same underlying model, differentiated by safeguard levels rather than architecture. Fable 5.1 is generally available to all API users. Mythos 5.1 is restricted to Anthropic's trusted access programmes and is available to organisations working in cybersecurity and the life sciences. The dual-configuration design allows Anthropic to offer frontier capability with more permissive constraints for professional research contexts, while the standard-access model continues under conventional safety guardrails.

Benchmark Improvements

Fable 5.1's headline improvement is in scientific reasoning. On Terminal-Bench-Science 0.1, the model scores 52.6%, more than doubling Fable 5's 24.7% on the same evaluation. Claude Opus 5, Anthropic's prior top-tier model, scores 29.0% on Terminal-Bench-Science 0.1, meaning Fable 5.1 surpasses both its immediate predecessor and the previous flagship by a wide margin. GPT-5.6 Sol, OpenAI's comparable model on this benchmark, scores 22.4%, placing Fable 5.1 more than 30 percentage points ahead. On Terminal-Bench 4.0, which measures performance on full software engineering tasks completed end-to-end in a terminal environment, Fable 5.1 scores 55.8% against Fable 5's 42.0% and Opus 5's 52.3%.

What Terminal-Bench Measures

Terminal-Bench evaluates a model's ability to complete multi-step software engineering tasks in a real command-line environment: writing and running code, reading failure output, and iterating to a working solution. A score of 55.8% means the model completes 55.8% of tasks successfully from start to finish. This is a substantially harder evaluation than benchmarks that score whether a model produces the correct function in isolation, and gains on Terminal-Bench tend to correspond to measurable improvements in production agentic coding workflows.

Pricing: The 75% Cache Cut

Fable 5.1 keeps Fable 5's input and output list prices: $10 per million input tokens and $50 per million output tokens. The significant change is in cache read pricing, which dropped 75% from $1 per million tokens to $0.25. Agentic applications repeatedly call the model with a large shared context — a system prompt, tool definitions, a document corpus, or a long conversation history — and cache reads make up a substantial fraction of total token cost in those workflows. Anthropic estimates the new cache pricing makes Fable 5.1 approximately 25% cheaper than Fable 5 for typical API workloads and up to 45% cheaper for heavily agentic ones.

Mythos 5.1 and Enterprise Frontier Safeguards

Mythos 5.1 ships with calibrated relaxed safeguards for professional research environments. The new models flag benign cybersecurity requests approximately 60% less often than prior Claude generations and reduce fallback refusals on basic biology and medical questions by approximately 85%. These reductions address a documented friction point for security researchers and life science practitioners: earlier Claude versions would frequently refuse or heavily qualify responses that a qualified professional would consider routine.

Alongside the model release, Anthropic announced Enterprise Frontier Safeguards, abbreviated EFS, a new security architecture that lets enterprise customers keep model interaction data and monitoring logs inside their own cloud infrastructure. Organisations running EFS get privacy equivalent to Anthropic's zero data retention offering, but without any interaction data leaving the customer's environment. EFS is scheduled to roll out in phases beginning in autumn 2026, targeting regulated industries — financial services, healthcare, and defence — where data sovereignty requirements make conventional cloud API use difficult to justify at an enterprise compliance level.

Breaking Change: Context Editing Removed

One API change in Fable 5.1 requires developer attention. Accounts created on or after 1 September 2026 can no longer manually edit Claude's prior responses in a multi-turn conversation while simultaneously preserving the transcript of the model's internal thinking. Anthropic describes this as closing a publicly documented distillation technique — a method that allowed developers to extract and use the model's reasoning traces to train smaller models, effectively creating a Fable-quality reasoning distillate without paying for full API access. The change does not affect standard agentic workflows, which do not rely on editing prior turns, but it does affect developers who had built custom context-manipulation logic around the extended thinking transcript.

What This Means for Software Teams in India

For development teams in India using Claude via the Anthropic API, the most actionable change is the cache pricing reduction. Agentic pipelines — document analysis, code review agents, retrieval-augmented generation systems — routinely cache large system prompts and call the model repeatedly within the same session. A 75% drop in cache read pricing translates directly to reduced monthly API costs with no code changes required. For teams building complex agentic workflows, the Terminal-Bench coding gains mean the model will complete more multi-step engineering tasks autonomously, reducing the need for explicit error-handling and retry logic. For organisations in the banking, financial services, and healthcare sectors with data residency requirements, the planned Enterprise Frontier Safeguards rollout will be worth monitoring closely as it could open frontier model access that cross-border API calls currently preclude.

The Bottom Line

Anthropic released Claude Fable 5.1 and Mythos 5.1 on 1 September 2026. Fable 5.1 scores 55.8% on Terminal-Bench 4.0, up from 42.0% for Fable 5, and 52.6% on Terminal-Bench-Science 0.1, more than double Fable 5's 24.7% and ahead of GPT-5.6 Sol's 22.4%. Cache reads dropped 75% to $0.25 per million tokens, cutting agentic workload costs by up to 45%. Mythos 5.1 ships with relaxed safeguards for professional research, reducing false refusals on cybersecurity by 60% and on biology and medicine by 85%. Enterprise Frontier Safeguards will bring data-sovereignty-compatible API access to regulated industries in autumn 2026. Accounts created from 1 September can no longer edit prior context while preserving thinking transcripts.