The new models shipped September 1. The benchmark bump is not the point.
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 that day. The press kit talks about benchmark gains, as usual. The detail worth watching this time is the bill: the flagship was split into two versions, the sticker price did not move, but the cost of using it dropped.
The two are the same underlying weights in two configurations. Fable 5.1 is generally available on every major channel. Mythos 5.1 is reserved for vetted cybersecurity and life-sciences organizations, and for now only US institutions. The difference is not capability. It is how strict the safety guardrails are.
The price sheet did not change. One line dropped 75%.
Input stays at 10 US dollars per million tokens, output at 50, identical to Fable 5. What moved is cache reads: from 1.00 to 0.25 US dollars per million tokens, a 75% cut.
Why this line matters: an agent re-reads its system prompt, tool definitions and accumulated notes on every single step. On Fable 5 that repeated reading was a large share of the cost. On 5.1 it is a quarter of the price. So in Anthropic's own estimate, typical workloads cost about 25% less, and highly agentic workloads up to 45% less. The more a job looks like running for hours without stopping, the more it saves.
The benchmarks rose. Where they rose is the real signal.
Terminal-Bench-Science 0.1, the agentic-research measure, more than doubled from 24.7% to 52.6%. Terminal-Bench 4.0, agentic coding, went from 42% to 55.8%, and the looser-guarded Mythos 5.1 reaches 60.9%. Humanity's Last Exam with tools scored 65%. CursorBench came in at 73.4%.
But multidisciplinary reasoning barely moved, from 63.8% to 65.0%. That distribution is the tell. This release was tuned for long-horizon tool use, where a model grinds through a task step by step inside a terminal or a lab pipeline, not for answering a single prompt. For chat and writing, the change is small.
Enterprise privacy moved too
The new architecture is called EFS, Enterprise Frontier Safeguards. Data lives in cloud infrastructure the customer controls, not Anthropic, giving effective zero-data-retention while keeping abuse guardrails. It rolls out in phases starting this fall. Until then, eligible customers can run Fable 5.1 in zero-retention mode.
Two safety numbers: cybersecurity false positives fell 60%, and biology guardrails now misfire 85% less on benign requests. Mythos 5.1 can flag likely vulnerabilities in reviewed code but will not produce working exploits. That line is drawn hard.
What this means for you
If you run long agentic pipelines, re-run the math. Things you left off because Fable 5 was too expensive to keep running are worth re-checking at 5.1's numbers. Measure how much of your spend is cache reads before you decide to switch.
For daily chat, writing and ordinary coding help, almost nothing changes. Fable is the top of the line at top-of-the-line pricing; Sonnet or Opus usually cover those jobs more cheaply. And the Mythos door is not open to domestic enterprises yet. It is US institutions only, for now.
One line: the cut is real, but it cuts the cost of agentic marathons, not your everyday chat.
Run the actual numbers
Say you run a coding agent with a 50,000-token context that re-reads it 100 times in one task. On Fable 5 those 5 million cache-read tokens cost 5 dollars at 1.00 per million. On 5.1 the same 5 million cost 1.25 dollars. Across a debugging cycle of a few hundred turns, the saving lands in the tens of dollars per run.
That is before counting the fewer-tokens effect. Anthropic says 5.1 reaches the same answer with fewer tokens. Stack the two and the cost structure of long-running work genuinely changes. But if your daily usage is a few dozen short questions, cache reads were never a big share of the bill, and the saving is hard to see.
Read it next to the others
Horizontally, this cut follows the same rhythm as OpenAI and Google: nudge the flagship's capability up, shave its effective unit price down. The difference is Anthropic aimed the knife at cache reads, the cost item specific to agents, rather than touching the list price. Teams still on Sonnet 5 or Opus 5 for daily work do not need to rush. Fable's list price did not move. The saving is someone else's money.
What Mythos actually unlocks
Mythos 5.1 is the more interesting half for a narrow set of buyers. Same weights, fewer refusals, gated to cybersecurity and life-sciences shops that pass verification. For a defensive security team that kept hitting false positives on legitimate analysis, 60% fewer false positives is the difference between a usable tool and a frustrated one. For biotech, the Life Sciences Verification Program, built with the US government, is the door. Outside those two groups and outside the US, it is simply not available, and that is the point.
The watermark deserves a line too. Every Fable 5.1 output carries an invisible text watermark that detection-API holders can statistically identify, in line with the EU AI Act Code of Practice. If you ship generated content at scale, that is now part of the deal, not an afterthought.