Claude Opus 5.5 Cuts Prices, and the Table Says Where

Claude Opus 5.5 became generally available on 22 September as the first model in what Anthropic calls the Claude 5.5 family, and the benchmark table is the least interesting thing about it. The pricing is the story: input at 4 dollars and output at 20 dollars per million tokens against Opus 5's 5 and 25, with cache reads at 20 cents against 50.

That last number is the one that changes a bill. Agent loops re-read the same context on every turn, so cache reads are often the largest line in a long session. Cutting them by 60% does more for the cost of a codebase-wide refactor than the 20% cut to output does.

The published prices, and the 40% claim next to them

Anthropic's published per-million-token prices, as reported on the release page

Line	Opus 5	Opus 5.5
Input	$5	$4
Output	$25	$20
Cache read	$0.50	$0.20
Cache write	$6.25	$5

The company states, on the release page, that at default settings it will cost 40% less than Opus 5 on typical workloads. That figure is a vendor claim about a workload mix nobody outside Anthropic has seen, and it is bigger than the price table on its own supports. Input and output each fall 20%, cache writes fall 20%, and only cache reads fall 60%. A session with almost no cache reuse will not see anything close to 40%. A session that is nearly all cache reads will beat it. Read the claim as a statement about Anthropic's traffic mix rather than as a discount you can expect.

Anthropic also reports output more than 30% faster than Opus 5, and a fast mode priced at 8 dollars input and 40 dollars output per million tokens for up to 2.5 times the speed. The fast-mode premium is a separate calculation again: twice the input price for latency, which is a good trade for an interactive coding session and a bad one for a batch job.

What a cheaper cached token does to an agent loop

The economics of an agentic coding session are dominated by what gets sent repeatedly. A system prompt, a repository map, tool definitions and the accumulated transcript are re-read on every step, and with prompt caching in place most of those tokens are billed at the cache-read rate rather than the input rate. That makes the cache-read price the effective price of keeping a long conversation alive.

At 50 cents per million tokens a careless loop has a real cost. At 20 cents the arithmetic that decides whether to compact a context, restart a session or carry the full transcript forward shifts, and code that previously had to be aggressive about summarising can be allowed to keep more. Teams should re-run their own numbers rather than assume the improvement is proportional, because the answer depends entirely on their cache-hit rate.

The benchmark figures Anthropic published are vendor-reported and should be read as such: Terminal-Bench 4.0 at 66.4%, FrontierCode v1.1 at 54.4%, CursorBench 4.0 at 57.8%, GDPval-AA v2.1 at 1846, OSWorld 2.1 at 81.8% described as partial, and Humanity's Last Exam at 67.7% with tools. The same page credits testers with migrating a 680,000-line codebase in less than a day, which is a marketing claim rather than a measurement.

Latency is the other half of the bill. A model that returns output more than 30% faster finishes a task in fewer wall-clock minutes, and for an agent that makes many sequential calls the saving compounds in a way that per-token prices do not capture. The relevant metric for an agentic workflow is rarely cost per million tokens; it is cost per completed task, which folds in retries, wasted turns and the time a person spends waiting. A cheaper token that needs more turns to finish is not cheaper.

The migration clock has already started elsewhere

Independent confirmation that Opus 5.5 is the current generation comes from GitHub, not from Anthropic. In a changelog published on 15 October, GitHub announced that on 16 November 2026 it deprecates three models across all Copilot experiences: Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra. The suggested replacement for the Claude models is Opus 5.5, and the suggested OpenAI replacements are GPT-6.1 Sol or GPT-6 Luna.

Anyone pinning a model in Copilot's agent or ask modes, or in CI configuration that names a model explicitly, has a dated migration rather than a vague upgrade path. One carve-out is worth noting: GitHub says Claude Opus 4.8 remains available to individual Copilot subscribers on annual plans, so not every customer is affected in the same way.

What else changed for existing customers

Anthropic raised the five-hour usage limits on Pro, Max, Team and Enterprise plans and added a saveable rate-limit reset, according to the release page. Neither detail changes the price of a token, but both change how much work fits into a window, which is the constraint most subscribers actually hit.

The company also says Claude Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks. Since Sonnet and Haiku are the tiers most production traffic sits on, the pricing that matters for a typical workload is still unpublished. Anyone who moved to Opus because it was the only way to get a task done should watch for a cheaper tier that clears the same bar.

What to do

Re-run a real workload rather than a synthetic one. Instrument cache-read and cache-write volumes for a week, then price the same traffic at the Opus 5 and Opus 5.5 rates. If most of your tokens are cache reads, the reduction will exceed the headline; if they are not, the saving is closer to 20% and the decision to move should rest on the latency and quality claims instead.

Then diarise 16 November if you use Copilot. An Opus 5 pin in a workflow file becomes a failure that morning, and the replacement is the model whose pricing this article is about.