Claude Mythos 5.1 for Product Managers: Cache Pricing, Benchmarks, and What Changed

Anthropic released Claude Mythos 5.1 on September 1, 2026, alongside Claude Fable 5.1. Both are point releases to the Mythos-class frontier model family introduced in June 2026. The version increment signals meaningful improvements rather than a full generational upgrade.

The two changes that matter most for product teams are cache pricing and benchmark performance. The core specs are unchanged: a 1 million token context window, 128,000 token maximum output, and $10 per million input tokens plus $50 per million output tokens. Adaptive thinking remains always on. Access to Mythos 5.1 still requires approval through Project Glasswing, Anthropic's restricted partner program.

According to VentureBeat's coverage of the launch, Anthropic measures the 75% cache pricing reduction as roughly 25% lower cost on typical workloads and up to 45% on agentic ones, where prompt caching does more of the heavy lifting.

Prompt caching stores a model's intermediate computation state for a fixed portion of a prompt, typically a large system prompt or a shared document corpus. Subsequent calls that reuse the cached prefix pay the cache read rate rather than the full input token rate. For workloads with stable system prompts and repeated context, caching cuts the effective input cost by a substantial fraction.

The economics matter most for agentic pipelines. An agent that executes a 10-step task passes the same base context (system prompt, tool definitions, document corpus) on every step. Without caching, each step pays full input token rates. With caching, only the new content in each step is billed at full rate. The cache read rate covers everything that is reused.

The practical implication: agentic pipelines that were cost-prohibitive at Mythos 5 pricing become viable at 5.1 rates. If your team evaluated Glasswing access and deferred because of inference economics, the 5.1 pricing revision changes the calculation for high-volume research and security workflows.

Terminal-Bench-Science is a benchmark suite that evaluates frontier models on their ability to complete long-horizon scientific and software engineering tasks using terminal access. Tasks require models to autonomously navigate file systems, run code, call scientific libraries, and validate results across multi-step sequences, with no human assistance.

Mythos 5.1 scores 52.6% on this benchmark. The benchmark design intentionally targets the boundary of what current frontier models can do: tasks are selected so that no model scores above 60%, which makes the 52.6% result a meaningful leadership position rather than a saturated metric.

The 5.1 release updates the cost profile for the top two tiers of the Claude model family. The pricing below reflects September 2026 rates, including the 5.1 cache pricing changes. Fable 5.1 carries the same 75% cache read reduction as Mythos 5.1 and is generally available without Glasswing approval.

For existing Glasswing partners already running Mythos 5, the migration to 5.1 is straightforward. The architecture is unchanged: same context window, same output limit, same input and output pricing, same adaptive thinking behavior. The only breaking surface is the model identifier.

The decision framework for Mythos 5.1 differs depending on where you are starting from.

The model ID update and cache pricing change are the full scope of what is new in Mythos 5.1. The restricted access model, use case profile, and benchmark leadership in cybersecurity and life sciences remain as documented for Mythos 5. The AWS Bedrock model card for Mythos 5.1 confirms the 1M context window and 128K output limit are unchanged.