Claude Opus 5.5 ships at $4/$20 - and that is LESS than the Opus 5 it replaces

Claude Opus 5.5 is live. Our sweep read the id claude-opus-5-5 at 16:24:44 UTC on September 22 and the alert went out at 16:25:45 - 61 seconds later, to 307 inboxes and one phone. The hunt that caught it was armed the previous day, before the model had a confirmed name, on the version range rather than on the rumour.

THE PRICE IS THE STORY

Opus 5.5 costs $4 per million input tokens and $20 per million output. The Opus 5 it replaces is $5 and $25. A frontier flagship launching 20% CHEAPER than its predecessor is not what this market normally does - the usual pattern is a higher sticker for the new top model and a discount on the old one.

Anthropic did the opposite. Opus 5 has moved to the legacy list and kept its $5/$25 price. So as of today the older, superseded Claude Opus is the expensive one, and anybody who stays on it is paying a 25% premium for the model Anthropic itself no longer recommends. Their own guidance now reads: start with Claude Opus 5.5 for most workloads.

Cache reads land at 5% of base input - $0.20 per million. On a workload that re-reads a large context repeatedly, that is the number that decides the bill, not the headline rate.

WHAT IT IS

1M token context window, 128K maximum output. Adaptive thinking is always on, with the default effort set to medium - you do not turn reasoning on, you turn the dial. Knowledge cutoff June 2026. Anthropic describes it as being for long-running agentic coding and knowledge work, which is a narrower claim than general-purpose frontier and worth reading as written.

Available as claude-opus-5-5 on the Anthropic API, anthropic/claude-opus-5.5 on OpenRouter, anthropic.claude-opus-5-5 on Bedrock, and claude-opus-5-5 on Vertex and Foundry. Batch pricing is half: $2/$10. Anthropic has committed to not retiring it before September 22, 2027.

WHAT WE DO NOT KNOW

No independent benchmark numbers yet. Anthropic's own evaluation figures are the vendor's, and we do not repeat vendor benchmarks as findings. When Artificial Analysis or another third party publishes, we will write that up separately with the arithmetic shown, the way we did for Grok 4.7.

We also cannot yet verify capability against the API ourselves - our Anthropic key has been out of credits since September 3, so the launch probe fails open. The detection above rests on the id appearing in Anthropic's own published model list and on three independent catalogs carrying it at a consistent price, not on a completion we ran.

THE NEXT ONE IS ALREADY ARMED

Opus cadence in 2026: 4.7 on April 16, 4.8 on May 27, 5 on July 24, 5.5 on September 22. The hunt for 5.6 and above, and for a 6, went on the board the moment 5.5 landed - armed on the range, not on a leak. That is the same shape that caught this one.