Anthropic released Claude Opus 5.5 on 22 September 2026. It is the first model in the Claude 5.5 family, and it is positioned as the company’s leading model for complex coding, research, computer use and professional knowledge work.

The launch story is not capability. It is cost. Anthropic has cut the per-token price of its flagship by a fifth, and has cut the price of cached input far more than that. The interesting question for anyone running agents or coding tools against the API is whether the cheaper bill is the headline change or a side effect of a more consequential one.

The honest framing has to come first, because it shapes everything below: almost every number in this article comes from Anthropic’s own launch materials. No independent benchmark evaluation of Opus 5.5 existed when it shipped. Anthropic says external evaluators, including METR and Frontier Design, ran pre-release tests, but those results are not public, so nothing independent to compare the launch table against. Treat the benchmarks as vendor claims until a third party reproduces them.

The price cut, in plain numbers

The API price is $4 per million input tokens and $20 per million output tokens. That is 20% below Opus 5.

Cache writes dropped by about the same proportion, from $6.25 to $5 per million tokens.

Anthropic goes further than that. It claims that typical workloads cost roughly 40% less overall, and the reason is not entirely the headline rate: the company says the model also needs fewer tokens to finish the same task. A lower rate multiplied by less token consumption gets you to a bigger number than a 20% cut alone.

There is a catch attached to that 40%, and it is the kind that quietly invalidates comparisons. The saving depends on the new medium default effort setting, not on the effort level you tuned for Opus 5. If you were running Opus 5 at a high effort setting and you switch to Opus 5.5 without changing anything, you are comparing a cheaper per-token price against a different amount of thinking. Anthropic’s own guidance is to measure cost per completed task on your own workload rather than to accept the 40% as a property of the model.

Caching is the bigger change

Cache reads fell from $0.50 to $0.20 per million tokens. On most Claude models a cache read costs 10% of the input price. On Opus 5.5 it costs 5%.

That sounds like a rounding difference. It is not, because cache reads are where the money is in agentic work. A coding agent does not send one prompt. It sends the same repository context, the same tool definitions, the same prior conversation, over and over, across dozens of turns. After the first turn, most of the tokens in that traffic are text the model has already seen. The first turn is a purchase. Every turn after it is a re-read.

So the price that matters is not the input price. It is the price of the twentieth, the fiftieth and the hundredth read of the same context block. When that number falls by 60%, the effective cost of a long-running session falls much further than the headline 20% suggests.

Anthropic’s worked example makes the shape of this concrete. Take a task that uses 2.8 million input tokens:

  • No caching: $11.20.
  • A 90% cache hit rate: $1.62.
  • A 96% cache hit rate: about $0.99.

Two things are worth pulling out of those three figures.

The first is the cliff. Going from no caching to a good cache hit rate is the single largest lever in the whole list — roughly a seventh of the original cost. Any agent that is not caching at all is leaving most of the available saving on the table, and it is a cheaper fix than switching models.

The second is the flattening. The step from 90% to 96% buys far less in absolute terms than the step from zero to 90%, because there is little left to remove. The curve runs out. Once you are deep into the nineties, chasing the last few points of hit rate is diminishing returns, and the remaining uncached tokens are still carrying the full input price. Chasing a perfect hit rate is not a strategy.

This is also why a per-token comparison between two models tells you very little about a caching workload. The two variables move independently: you can run the same hit rate on a cheaper model, or the same model at a different hit rate, and the bill is a product of the two. Measure both.

The benchmarks, and the caveat attached to them

Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 on most work, and that it is smaller and cheaper to serve.

The launch table reports:

  • Terminal-Bench 4.0: 66.4%, against 52.3% for Opus 5 and 57.9% for GPT-6 Astra.
  • GDPval-AA v2.1: 1846 Elo, ahead of Fable 5.1 at 1735 and Opus 5 at 1708.
  • Cost per task: matching Astra on Terminal-Bench 4.0 at roughly 40% of the cost per task.

One note on the first line: that 66.4% is Opus 5.5 run at its highest effort setting, and the published table gives the new medium default effort setting no score of its own — which matters, because that default is where much of the 40% saving is said to come from.

The cost-per-task claim is the one that matters for the price argument, and it is the one that depends most on the effort setting described above.

Then Anthropic attached a caveat that is unusual for a launch post. Its position is that benchmark margins have become a less reliable guide to real-world differences, and that the gap between Opus 5.5 and Fable 5.1 is narrower in its own internal use than the scores suggest. In other words, the company is telling you not to over-read the table it just published.

That is worth taking seriously rather than treating as politeness. A 14-point margin over its own predecessor on a benchmark is not a 14-point margin on your codebase, and a vendor publishing its own numbers is not in a position to adjudicate its own caveat. Until an external body runs the same evaluation, the honest reading is: Opus 5.5 is plausibly at or near the top of this class, and the specific margins are unverified.

Parameter count and architecture are not disclosed, so none of this can be sanity-checked from the outside.

Safety, availability and the migration

Opus 5.5 launched with the same style of safeguards Anthropic applied to Fable 5.1, covering cybersecurity, biology and distillation. External evaluators including METR and Frontier Design tested it before release. It is also the company’s first release since CEO Dario Amodei called for pacing the frontier — slowing progress so that safety practice stays ahead of capability.

It is available through Anthropic’s API and across AWS, Google Cloud and Microsoft platforms, and on the Pro, Max, Team and Enterprise plans. Migration from Opus 5 is described as a one-line model-name change, which should be read as a starting point rather than the whole task: the release includes breaking API changes, and those are what will actually cost you an afternoon.

For Claude Code users there is a practical gain. Anthropic raised five-hour session limits by 20%, and says Opus 5.5 gets about 25% further within those limits because it is cheaper. Two separate multipliers on the same budget.

What this means for you

The practical decision is not which model is better. It is which problems justify the price difference.

Use Opus 5.5 for hard, judgment-heavy and long-running work: the tasks where the model has to hold a plan across many steps, where a wrong turn is expensive to unwind, and where the output is the thing you are paying for rather than the tokens.

For well-scoped everyday tasks, Sonnet 5.5 at half the per-token price is likely enough. That is a sensible default, and it is the right one: start there, and escalate to Opus 5.5 when you can name the reason.

Before you upgrade anything, do three cheap things. Turn on caching if you have not, because that is worth more than the model swap and works on whatever model you are running. Measure cost per completed task rather than cost per million tokens, on your own workload, with your own effort settings. And read the breaking changes, because the migration is only one line until it is not.

Then wait a few weeks for someone who is not the vendor to run the benchmarks. The pricing is published and the caching economics are arithmetic, so those parts you can rely on today. The capability claims cannot.

Sources

Reporting on this release, none of which is an independent model evaluation:

  • VentureBeat, “Anthropic releases Claude Opus 5.5, beating Fable 5.1 on key agentic benchmarks at 60% cheaper API price”
  • Technology.org, “Claude Opus 5.5 Launches at 40% Lower Cost”
  • Digital Applied, “Claude Opus 5.5: Pricing, Benchmarks and Breaking Changes”
  • Kingy AI, “Claude Opus 5.5: Specs, Benchmarks, Pricing and How It Stacks Up”
  • AlphaCorp, “Claude Opus 5.5 Launch: Benchmarks, Pricing, Migration”
  • Finout, “Claude Opus 5.5 Pricing 2026”
  • Cosmic JS, “Claude Opus 5.5 vs Opus 5: Benchmarks and Cost”