Anthropic released Claude Fable 5.1 on September 1, and the number that travelled is 52.6% on Terminal-Bench-Science, more than double Fable 5's 24.7% and well ahead of GPT-5.6 Sol's 22.4%. It is a real jump. It is also, for most people who build on these models, the third most important thing in the release.
The two bigger ones are quieter: cache reads now cost 75% less, and the same model ships behind two different safeguard layers with a measurable gap between them. Both change how you should read the benchmark table.
The cache cut is the release, if you run agents
Base pricing is unchanged at $10 per million input tokens and $50 per million output. What changed is the line item most agent workloads are actually made of: cached input drops from $1.00 to $0.25 per million tokens.
Anthropic's own framing is that typical workloads get roughly 25% cheaper and context-heavy agentic ones save up to 45%, and the arithmetic checks out. An agent loop re-reads its system prompt, its tool schemas and its conversation history on every turn; in a long session, cached tokens outnumber fresh ones many times over. Our own pipelines are a fair example: a nightly writer batch that holds a large measured-facts brief in context pays for that brief again on every generation step. At the old rate the cache was a quarter of the bill; at the new rate it is closer to a rounding error.
Put numbers on it. Take an agent session that reads 2 million fresh input tokens, re-reads 20 million cached tokens across its turns, and writes 500,000 tokens of output. Yesterday that session cost $20 fresh + $20 cache + $25 output, so $65. Today the middle line is $5, and the session costs $50: a 23% cut with nothing about the workload changed. Stretch the cache-to-fresh ratio the way long research or coding sessions actually do and the saving climbs toward Anthropic's 45% figure honestly, without marketing arithmetic.
That is why we would call this a pricing release wearing a benchmark release's clothes. A 45% cost cut on agentic workloads moves more budgets than a Terminal-Bench score does. And it lands the day after Anthropic reportedly leased $35 billion of compute from Lambda, infrastructure to be housed in a data centre under construction in Nueces County, Texas. Cheaper tokens for customers, secured capacity to serve them: the two announcements are one strategy.
One model, two names, and a measurable safety tax
The stranger half of the release is Claude Mythos 5.1. Anthropic describes Fable 5.1 and Mythos 5.1 as one model behind two safeguard layers. Mythos, the less restricted variant, stays limited to vetted US organizations inside Project Glasswing. Everyone else gets Fable. Project Glasswing itself remains thinly documented in public: what is known is that admission involves vetting, that it is limited to US organizations, and that Anthropic treats the less-wrapped variant as the thing requiring gatekeeping rather than the base capability.
What makes this more than naming is that Anthropic published scores for both, so the cost of the safeguard layer is visible for once: on Terminal-Bench 4.0, Fable 5.1 scores 55.8% and Mythos 5.1 scores 60.9%. Five points of capability, spent on safety wrapping, in public. Whatever you think of the trade, it is the first time we can recall a lab printing the price tag.
The benchmark table, with the caveats attached
The verified numbers, against the models that matter:
- Terminal-Bench-Science 0.1: 52.6%; Fable 5 scored 24.7%, Opus 5 scored 29.0%, GPT-5.6 Sol scored 22.4%
- Terminal-Bench 2.1: 85.02%, up from 80.52%, still behind GPT-5.6 Sol's 85.77%
- LiveCodeBench: 90.52%, first place
- ProofBench v1.1: a flat 100.00%, up from 95.00%
- Vals Index: 67.87%, ahead of Claude Opus 5 (67.21%) and Fable 5 (66.04%)
- Humanity's Last Exam: 60.9% without tools, 65.0% with
- OSWorld 2.0: 41.7% under strict scoring
Two caveats belong next to that list. First, the reported standard error runs 3.5 to 4.5 points per model, which means the Vals Index gap over Opus 5, all of 0.66 points, is noise wearing a ranking. Second, the headline doubling comes on Terminal-Bench-Science 0.1, a benchmark young enough to carry a version number starting with zero; new benchmarks tend to reward whoever tuned for them first. The coding numbers on the older, harder-worn suites moved less: a 4.5-point gain on Terminal-Bench 2.1, and second place on it.
The spec sheet is easier to state: a 1M-token context window, 128K max output, model ID claude-fable-5-1, available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.
Three breaking changes builders should read twice
The API notes carry three changes that will break real code. Forced tool use is gone. Thinking blocks are now model-bound, meaning earlier models cannot read Fable 5.1's reasoning traces. And editing conversation history invalidates thinking blocks outright.
The second and third matter to anyone running multi-model pipelines: a cheap model triaging before an expensive one finishes, or a session resumed by a different model than the one that started it. That pattern now needs thinking blocks stripped at the boundary. It is a small migration, but it is the kind that fails silently in production rather than loudly in testing, because nothing errors: the downstream model simply loses the reasoning it used to see.
We run Claude models across our own pipelines daily (the review writers, the growth agent, this site's entire editorial toolchain), so the upgrade path here is not academic for us. The cache change alone reprices every long-context job we run, and the thinking-block rule touches any flow where one model hands work to another.
Where this sits in the season
September opens with the same pattern August closed on: capability announcements that are really price announcements. GLM-5.3-Flash arrived with half-price launch tokens and a half-price launch window; OpenAI spent the last week of August pulling its models out of Cursor in a dispute about who may resell access to what. Fable 5.1 fits the moment exactly — the frontier labs are competing on the cost of running agents at scale, because that is where the volume is going.
For developers, the practical read is short. If you run agentic workloads on Claude, the upgrade likely pays for itself in cache savings before you ever look at a benchmark. If you run multi-model pipelines, read the thinking-block changes before you switch anything. And if you are choosing between labs on the strength of Terminal-Bench-Science, wait for the 1.0 of the benchmark before you move. Version-zero leaderboards reward whoever arrived first; the suites that have survived a few release cycles are the ones that still mean something a quarter later.