Anthropic cut the only price that matters and left the sticker alone

Most price cuts are performed loudly on the number nobody's bill is actually driven by. Anthropic did the opposite on 1 September: it left $10/$50 in place, said little about pricing in the headline, and quietly removed three quarters of the cost of the thing agentic workloads spend most of their tokens on.

Run an agent through a codebase for an hour and count what you paid for. It is not novel input. It is the same system prompt, the same file context, the same accumulating history, re-sent every single turn and billed at the cache-read rate. For long-horizon work that line item dwarfs everything else, which is why cutting it 75% matters more than a headline discount would.

Anthropic puts the effect at roughly 25% lower cost for typical workloads and up to about 45% for highly agentic work. Read the provenance before you budget on it: those numbers come from Anthropic's own usage measurement over four weeks in August 2026. They are a vendor estimate of a vendor's workload mix. If your agent re-sends less context, your saving is smaller; if it re-sends more, it may be larger.

Two days later OpenAI launched GPT-6 Astra at $10 input and $50 output — the identical sticker. But Astra did not arrive with a matching cache-discount lever, which leaves it at a relative disadvantage precisely where enterprise agentic spend concentrates. Anthropic did not need to undercut the sticker; it undercut the invoice.

That is a sharper piece of competitive positioning than the launch-day coverage gave it credit for, because it is legible only to buyers who have actually run the workload. Casual comparison sites list $10/$50 against $10/$50 and call it parity. Anyone with three months of agent billing in front of them sees a different table.

The second lever is the effort ladder. Anthropic says Low and Medium effort now match or beat Fable 5's results at much lower cost, while High effort goes well beyond it — and it has set defaults accordingly, High in Claude Code, Medium in Claude Cowork and on Claude.ai.

There is an implicit admission in that structure worth naming: a meaningful share of what you were paying for in the previous generation was compute you did not need for the task at hand. Effort dials are the industry finally acknowledging that 'the best model' and 'the right amount of thinking' are separate decisions, and that charging one price for both was leaving money on the table for everyone.

Fable and Mythos are the same weights with different permissions. Mythos loosens safeguards for vetted individuals and organisations — cybersecurity and life sciences — through two trusted access programmes; under Fable, requests tripping classifiers on cyber, bio/chem or distillation get handled by a less capable Opus instead.

This is now the shape of frontier safety across the industry: not one refusal boundary for everyone, but capability tiers gated by who you are. OpenAI's Daybreak programme does the same for cyber. It is defensible and probably necessary. It also means the phrase 'the model refuses that' is no longer a fact about the model — it is a fact about your access tier, and benchmark comparisons that ignore which tier was measured are comparing different products. Anthropic, to its credit, says exactly that: production safeguards can affect scores.

The capability claim worth taking seriously is not on any coding leaderboard. Anthropic reports that on three targets, the model's protein binding affinities came in ten times higher than the best designs submitted to Adaptyv Bio's protein design competitions. That is an externally anchored result against real human submissions in a domain where the answer is physical rather than scored by a rubric — a far better piece of evidence than another percentage point on a saturated benchmark, and a hint at where the next round of competitive claims will be fought.

The rest of the release is housekeeping done well: generally available across the Claude API, AWS, Google Cloud and Azure on day one, callable as claude-fable-5-1, live in Claude Code and Cursor immediately. No staged rollout theatre, no waitlist. For a lab whose customers increasingly build production systems on these models, shipping everywhere at once is its own form of competitive statement.