Anthropic released Claude Opus 5.5 on 22 September, the first model in a new 5.5 family. It is announced at the level of Fable 5.1 on most tasks, at 4 dollars per million tokens on input and 20 on output — against 10 and 50 for Fable 5.1, and 5 and 25 for Opus 5. Cache reads fall to 0.20 dollars, 60% less. Output is said to be more than 30% faster. This is the first release since Dario Amodei's essay calling on the industry to slow down, published ten days earlier.
We covered Fable 5.1 and Mythos 5.1 on 1 September, then GPT-6 Astra two days later, then the public explosion of the debate over pace. Here is the release that brings those three threads together.
The figures, side by side
| Opus 5 24 July |
Fable 5.1 1 Sept |
Opus 5.5 22 Sept |
|
|---|---|---|---|
| Input / M tokens | $5.00 | $10.00 | $4.00 |
| Output / M tokens | $25.00 | $50.00 | $20.00 |
| Cache read | $0.50 | — | $0.20 (−60%) |
| Cache write (5 min) | $6.25 | — | $5.00 (−20%) |
| Context | — | 1M | 1M (128K output) |
| Stated real cost | baseline | — | −40% |
The gap between the 20% drop in the headline rate and the 40% announced on real cost comes from two effects that compound each other: the fall in cache reads, and the fact that the model would use fewer tokens to accomplish the same task.
This is exactly the lesson we have been hammering home for a month, applied here in the right direction: the headline rate does not tell you what you will pay. We showed this with the effective rise in Sonnet 5 and token breakdown in billing.
Against GPT-6 Astra
| Test (source: Anthropic) | Opus 5.5 | GPT-6 Astra |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 57.9% |
| FrontierCode v1.1 | 54.4% | 53.3% |
| Terminal-Bench-Science | 58.7% | 64.6% |
| AutomationBench (business workflows) | behind | ahead |
These figures come from Anthropic's own tables. As we explained in our article on how benchmarks age, a score without its protocol is a number without a unit, and a measurement produced by the manufacturer is not an independent measurement.
To its credit, Anthropic itself acknowledges that benchmark gaps have become an unreliable guide to real-world differences at this level of capability, and that the distance from Fable 5.1 is narrower in practice than the scores suggest. That is a rare honesty in a launch announcement.
Anthropic reports that its own safeguards intervened during its own evaluations, redirecting certain tasks touching on cybersecurity and biology to older, less capable Claude models — which probably lowered its scores on those specific tests. In other words: the company is publishing results in which its own safety apparatus handicapped it, and it says so. This is exactly what we were calling for regarding Astra's safety report — a transparency that gives outside evaluators the thread to pull on, even when it works against whoever publishes it.
The question of the slowdown
Here is the topic that will make the most noise, and the answer is less simple than a headline.
On 12 September, Dario Amodei published an essay calling on the industry to slow the pace of frontier capability development — a text that triggered, within four days, three responses from executives, a proposed ban and a presidential comment. Ten days later, a new frontier model.
The contradiction is real if by slowing down one means publishing less often. Three flagship models in two months — Opus 5 on 24 July, Fable and Mythos 5.1 on 1 September, Opus 5.5 on the 22nd — amounts to a cadence that is accelerating rather than easing off.
It is far less clear-cut if by slowing down one means not pushing the capability frontier faster than safety can keep up. Because Opus 5.5 does not exceed the existing ceiling: it matches it. The frontier was set by Fable 5.1 on 1 September. This model reaches the same level for 60% less.
Making an existing capability cheaper is not the same thing as creating a new capability. It is, in fact, precisely what a slowdown at the frontier would allow.
The safety elements announced
Three elements accompany the release, and they are verifiable to varying degrees.
A containment evaluation. The model reportedly attempted to circumvent its limits around 85% less often than Opus 5 or Mythos 5.1. This is the measure most directly tied to the three incidents we have covered, from the first in July to Gemini's last week. The company notes that reliably detecting all failures before deployment remains an unsolved problem.
A behavioural audit of around 2,000 scenarios, on which this model is said to be the best-scoring it has published.
External evaluators before launch, including METR. This ties in with the announcement of embedded evaluators we reported on Saturday, and it is the most important part: a measurement made by someone else is worth more than a measurement you make yourself.
The usual caveat applies: these evaluations are adversarial by construction and do not describe failure rates in normal use.
What it changes for you
If you run agents, this is the release that concerns you most in a month. The cache read at $0.20 is the dominant cost in an agentic loop, exactly as we explained for Fable 5.1.
If you are a subscriber, the five-hour limits are going up, and a quota reset can be used until 22 October. It is a reversal from the tightening of 14 September that we covered.
If you were choosing between Opus 5 and Fable 5.1, the trade-off has changed: this model is cheaper than the former and announced at the level of the latter.
What to take away
This release is less an advance in capability than an advance in efficiency, and that is what makes it interesting at a time when the industry is debating its pace.
It also confirms what we wrote on Monday evening about the normalisation: a frontier model comes out, at half price, and the news now fits in a line of a pricing table rather than in an event.
What is missing remains what is always missing: an independent measurement. Tomorrow we publish the ranking updated to 23 September, systematically distinguishing what is measured by the manufacturers from what is measured by third parties.