Anthropic shipped Claude Opus 5.5 on September 22, 2026, and the first line I opened was not a benchmark table — it was the cache read price. Claude Opus 5.5 charges $0.20 per million cached input tokens, sixty percent under what Opus 5 charged, and cache reads are where most of a long-running agent bill actually sits. Input fell to $4 per million and output to $20, both twenty percent under the model they replace. A frontier lab shipping a stronger model at a lower price is not the pattern I have gotten used to this year, and it changes which question is worth asking.
In this article
- Why the cache read price matters more than the headline scores
- What Claude Opus 5.5 costs across every tier, Fast mode included
- Why the new benchmark names make year-over-year comparison harder
- Why the usage limit increase beats the price cut for subscribers
- What I am changing in my own stack this week
The Price Went Down While the Scores Went Up
Cheaper and stronger in the same release is the whole story, and it is rare enough to slow down for. Anthropic’s launch table puts Claude Opus 5.5 at 66.4% on Terminal-Bench 4.0 against 52.3% for Opus 5, and 57.8% on CursorBench 4.0 against 46.6%. On FrontierCode v1.1 it records 54.4% against GPT-6 Astra’s 53.3%. On GDPval-AA v2.1, a knowledge-work Elo board, Claude Opus 5.5 posts 1846 against 1735 for Fable 5.1 — Anthropic’s own more expensive flagship.
The published API rates are the cleanest fact of the launch:
- Input: $4 per million tokens, twenty percent under Opus 5
- Output: $20 per million tokens, down from $25
- Cache reads: $0.20 per million tokens, sixty percent under Opus 5
- Cache writes: $5 per million tokens
- Fast mode: $8 input and $40 output, running up to 2.5x quicker
Anthropic puts the combined operating cost at forty percent below Opus 5, with output generated more than thirty percent faster. Computer use lands at 81.8% partial on OSWorld 2.0, and AutomationBench comes in at a 40.0% pass rate — a reminder that the agentic frontier is still failing three tasks in five.
Claude Opus 5.5 Cache Reads Are the Line Item That Decides This
The cache read cut is the number that will show up on my invoice, and nobody put it in a headline. Agentic coding sessions re-read the same context on every turn: the same repository map, the same instruction file, the same half-finished diff. Those tokens are cached, and under Opus 5 they billed at $0.50 per million. Claude Opus 5.5 bills them at $0.20. For a workload that is mostly re-reading, that is a materially different bill from the twenty percent the press coverage led with.
I run Claude Code daily on file automation, CSV cleanup and the scripts that keep this site publishing, which is the kind of work I have written about before. Long sessions on one repository are exactly the shape of task that leans on cached context, so the cache line is not an abstraction for me — it is the dominant term.
There is a second-order effect worth naming. When a task gets forty percent cheaper to run, the honest response is usually not to bank the difference but to attempt jobs that were previously not worth the tokens. That is how every price drop I have tracked has actually played out, and it is why cheaper models rarely reduce anyone’s monthly spend.
The Benchmark Names Changed, and That Makes Comparison Harder
Anthropic’s launch tables use Terminal-Bench 4.0, FrontierCode v1.1 and CursorBench 4.0 rather than the SWE-bench family that anchored comparisons through 2025. That is defensible — saturated benchmarks stop discriminating — but it quietly costs you the ability to line this release up against older reports without doing the translation yourself.
It also means the only clean comparisons available on launch day are the ones the vendor chose to publish. Claude Opus 5.5 against Opus 5 on Terminal-Bench is Anthropic measuring Anthropic. Claude Opus 5.5 against GPT-6 Astra on FrontierCode is a 1.1-point gap on a benchmark I cannot independently reproduce, and a 1.1-point gap is well inside the range where run-to-run variance and prompt scaffolding decide the winner.
A 1.1-point lead on a vendor-selected benchmark is a press release, not a procurement decision. The number that survives contact with real work is the one on your own invoice at the end of the month.
None of that makes the scores dishonest. It makes them provisional. When Opus 5 landed in July I said the launch-day numbers would look smaller in a month, and they did — the changes that stuck were duller and narrower than the table implied. I expect the same curve here and would rather be proven wrong in October with evidence than right today by guessing.
Higher Usage Limits Beat the Price Cut for Subscribers
For anyone on a subscription rather than the API, the limit change is the real release. Anthropic raised the five-hour usage limits on Pro, Max and Team alongside the launch, and added a rate limit reset that subscribers can bank and spend when they choose. Claude Opus 5.5 is available to Pro, Max, Team and Enterprise users, and Fast mode is live in Claude Code as well as on the Claude Platform.
That distinction matters because the token prices apply to API calls, and a Pro subscriber does not pay per token. What a Pro subscriber pays is in interruptions — the session that stops mid-task because the window is exhausted. Raising the ceiling and letting a reset be saved for a deadline is worth more to my week than twenty percent off a rate I am not billed at.
It is also consistent with what the company has been signalling about its own path to sustainable margins. Cheaper inference lets a lab widen subscriber limits without widening losses, which is a healthier version of the same story than raising prices would have been.
Sonnet 5.5 and Haiku 5.5 are due in the coming weeks with the same efficiency work. For most of what I run, that pair matters more than the flagship does.
What Claude Opus 5.5 Changes in My Stack This Week
Two things change, and both are small. The Claude Code work I already run gets cheaper without me touching a configuration file, which is the least interesting kind of win and the most reliable. The second change is that I am moving one job back: the long repository audits I had been splitting into chunks to control cost are worth attempting in a single session now that cached context bills at $0.20 per million.
What does not change is the routing. Claude still holds long-form strategy documents and client research briefs. Perplexity still runs competitor research with citations attached. Anything covered by an NDA still goes to Ollama locally, because a launch-week endpoint is not where client material belongs, regardless of price.
My standing rule has been that an upgrade must return roughly three times its added cost before it displaces something that works. Claude Opus 5.5 is the first release in a long stretch where that rule does not apply, because the added cost is negative. The replacement question is simpler and duller: does it produce the same result or better? If yes, the cheaper model wins by default.
For me, the honest summary is that Claude Opus 5.5 is a good release that demands almost nothing of me this week. I will run my normal Claude Code sessions, watch the cache line on the usage dashboard for two weeks, and compare it against August before I claim any number. If Sonnet 5.5 arrives with the same cache economics, that will be the release that actually reshapes my stack. Until then this is a price cut I get to accept without changing anything, and those are rare enough to take quietly.
Sources
AI-assisted research and drafting. Reviewed and published by ToolMint.