# Anthropic's Fable 5.1 More Than Doubles Its Terminal-Bench-Science Score and Cuts Cache Reads by 75 Percent

The model holds Fable 5's prices at ten dollars per million input tokens and fifty per million output, while dropping cached reads to twenty-five cents, which changes the economics of long-running agents more than the headline benchmarks do.

- Published: 2026-09-13T06:28:59.111Z
- Canonical: https://polylog.news/ai/2026-09-13/anthropic-s-fable-5-1-more-than-doubles-its-terminal-bench-s
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Anthropic News](https://www.anthropic.com/claude-fable-and-mythos-5-1), [Vellum](https://www.vellum.ai/blog/claude-fable-5-1-mythos-5-1-benchmarks-explained)

Anthropic released [Claude Fable 5.1 and Claude Mythos 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1) at the start of September, with Fable 5.1 generally available on the Claude application programming interface (API) and on Amazon Web Services, Google Cloud and Microsoft Azure. Mythos 5.1 is the same model with lighter safeguards, restricted to organizations vetted through Anthropic's Cyber Verification and Life Sciences Verification programs, and currently limited to the United States.

The benchmark gains are uneven, which is the interesting part. On [Terminal-Bench-Science the score moves from 24.7 percent to 52.6 percent](https://www.vellum.ai/blog/claude-fable-5-1-mythos-5-1-benchmarks-explained), and on Terminal-Bench 4.0 coding from 42.0 percent to 55.8 percent. SWE-bench Verified sits at 95.0 percent, effectively unchanged from Fable 5, which is what you expect from a benchmark approaching saturation. On Anthropic's own knowledge-work benchmark GDPval-AA v2, Fable 5.1 scores 1853 against 1824 for Opus 5 and 1723 for Fable 5. The model also reports 100 percent on ProofBench v1.1 for formal mathematical proofs, a result that says more about the benchmark's difficulty ceiling than about the model.

Treat the framing with care. Terminal-Bench-Science and GDPval-AA are the evaluations where the gain is largest, and Anthropic is a substantial contributor to how knowledge-work evaluation is defined. A doubling on an agentic terminal benchmark is consistent with better tool-use scaffolding and longer effective task horizons rather than a new reasoning capability, and no independent reproduction of the science-terminal figure has been published.

The pricing is the part engineers can act on immediately. Input stays at ten dollars per million tokens and output at fifty, unchanged from Fable 5, while cached reads fall to twenty-five cents per million, a 75 percent reduction. The context window is one million tokens with flat per-token pricing across the full window and a maximum output of 128,000 tokens. For agent loops that replay a large fixed context on every step, the cache price is the dominant term in the bill, not the headline rate.

## What this means

The cache-read cut, not the benchmark movement, is what changes deployment decisions, because long-horizon coding agents re-read the same repository context hundreds of times per task and that read cost scales with steps rather than with output. Anthropic is defending the coding-agent segment against Meta's Muse Spark at one dollar twenty-five per million input tokens and against OpenAI's Astra, and it is doing so by cutting the price of the specific token class agents consume most rather than by cutting headline rates. Teams running agent frameworks on top of Claude see the largest saving. Teams running single-turn chat see almost none.

## What to watch

- Whether an independent evaluation group reproduces the Terminal-Bench-Science jump, since a doubling reported only by the vendor on its own preferred benchmark is the pattern that most often fails replication.
- Whether OpenAI or Google respond with comparable cache-read pricing, which would confirm that the competitive axis has moved from headline token price to the cost structure of agent loops.
- How many organizations actually clear the Cyber Verification and Life Sciences Verification gates for Mythos 5.1, because a capability behind a vetting queue reaches a much smaller market than a capability behind a simple sign-up.