Anthropic Cut Cache Reads 75% and Left the Sticker Alone

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1. The headline price did not move: $10 per million input tokens, $50 per million output, same as Fable 5.

The number that moved is the one most release coverage buries. Cached input reads dropped to $0.25 per million — a 75% cut. Anthropic says that brings typical workload costs down roughly 25%, and highly agentic workloads down by as much as 45%.

If you are running agents, the second number is your number, and it is the release.

Why cache is the real price

A single-turn chat call is mostly fresh tokens. An agent loop is not.

An agent working a repository re-sends the same system prompt, the same tool definitions, the same file context, and a growing transcript on every turn. Turn forty is nearly identical to turn thirty-nine plus a diff. In a long-running coding or research loop, cached reads routinely dominate total input volume — which means the cache rate, not the input rate, is what the invoice is actually computed from.

Cutting that line by three-quarters is a structural change to what is economically viable, not a discount. Loops that were too expensive to leave running now aren't. Context you were trimming to save money, you can stop trimming. The 45% figure for agentic workloads is not marketing generosity; it is arithmetic on a workload whose shape Anthropic understands because it ships Claude Code.

That is the strategic read. Anthropic is not competing on frontier price per token — it held that line at $10/$50, the identical sticker OpenAI attached to GPT-6 Astra two days later. It is competing on the price of the thing it wants you to build.

One model, two doors

Fable 5.1 and Mythos 5.1 are the same underlying model.

Fable 5.1 is the generally available version, carrying Anthropic's production safeguards. It is live across Claude, Claude Code, the Claude platform, and Cursor.

Mythos 5.1 is the same weights reached through restricted-access programs, available to vetted cybersecurity and life-sciences organizations that need capabilities the general safeguards constrain.

Flux has tracked this structure through a rough year: the government suspension of Mythos, the export review, and the Commerce decision that restored Anthropic to trusted-partner status. What was an exception handled under pressure is now simply how the product ships. The twin release is the default shape.

And it is no longer Anthropic's alone. OpenAI's Daybreak, announced with Astra on September 3, is functionally the same architecture — safeguarded public model, less-restricted tier for vetted defenders. Two labs, converging within seventy-two hours on the conclusion that frontier capability gets rationed by access tier rather than by capability ceiling.

Fewer false positives

The other change Anthropic flags is a reduction in false-positive refusals — cases where safeguards block work that was never dangerous.

This is unglamorous and it matters more than it sounds. A refusal in a chat window costs you a rephrase. A refusal on turn thirty of an autonomous loop costs you the run: the agent stalls, the state is half-mutated, and a human has to reconstruct what happened. As loops get longer, the expected number of spurious blocks per run compounds, and reliability becomes a function of refusal precision rather than model intelligence.

Anthropic shipping this alongside a cache cut is coherent. Both changes target the same failure mode: agentic runs that are too expensive or too brittle to leave unattended.

Internal testing also reports 5.1 outperforming Fable 5 on coding, knowledge work, and problem-solving, with similar results at low and medium effort settings at substantially lower cost, and a 52.6% score on Terminal-Bench-Science. As always, internal testing is internal testing — the independent numbers will land over the next few weeks.

What it means for the tooling layer

The immediate beneficiaries are the products whose entire cost structure is long context repeated many times.

Coding agents get the largest single line-item reduction they have had in a year. Research agents that maintain a large working set across dozens of tool calls get the same benefit. Anything that was architected around aggressive context pruning can revisit that decision, because pruning was often a cost optimization wearing a quality justification.

The less obvious effect is on model choice. A shop deciding between frontier tiers now has to compare not the headline rate — those have converged at $10/$50 — but the effective rate under its own cache-hit profile. That is a harder comparison, it is workload-specific, and it favors whoever ships the tooling that generates the workload. Fable 5.1 being generally available inside Cursor on day one is not incidental to that.

The pattern underneath

Two frontier releases landed in seventy-two hours, at the same sticker price, with the same two-tier access structure, both optimizing for agents rather than chat.

Neither lab is trying to win on being cheaper per token. That competition has moved to the open-weight tier, where Chinese labs set the floor. At the frontier, the price is settled and the contest is over what the token buys: how long a loop can run, how much context it can hold, how often it stops for no reason, and — increasingly — whether you are cleared for the version that does not refuse.

Anthropic's answer this week was to make the repeated part nearly free. It is the most consequential pricing move any lab has made this year, and it arrived with no change to the number on the front of the box.