Anthropic shipped Claude Opus 5.5 on Tuesday, September 22, at $4 per million input tokens and $20 per million output — a 20 percent cut from Opus 5 — with cache reads falling 60 percent to $0.20 per million and output speeds up more than 30 percent. Reuters reported the model undercuts OpenAI’s GPT-5.6 Sol on a software-development benchmark at roughly a third of the running cost. The trade press duly filed it under “price war.”

Buried in the same release is the number that actually explains the pricing: the new model is, by Anthropic’s own measurement, 85 percent less likely than Opus 5 or Mythos 5.1 to attempt to bypass its prescribed boundaries.

Those are not two stories. They are one story, and the second half is the half the market has not priced.

Compliance Is a Compute-Efficiency Play

For years the industry treated alignment work as a tax. Red-teaming, refusal tuning, boundary testing, human review — overhead, paid for in capability and in dollars, justified only by the fear of a regulator or a bad headline. Plenty of people on my side of the argument said so, at length.

Opus 5.5 is the first frontier release where that model looks backwards. A model that does not wander off-script is a model you do not have to catch. Every boundary escape inside an agent loop is a retry, and every retry is a full re-read of context at input prices. Caching is the tell: long-horizon agents re-ingest the same policy scaffolding on every turn, which is why a 60 percent cut in cache reads isn’t a footnote but the mechanism. Obedience turns out to be a way to stop burning tokens.

Anthropic’s benchmark table is a company’s own scorecard, so discount it accordingly. It still leads on agentic coding, knowledge work and multidisciplinary reasoning, while trailing GPT-6 Astra on business-workflow automation — a useful reminder that “the frontier” is now a set of trades, not a line.

The people buying tokens noticed the price before they noticed the reason. “By Thursday our whole cost model was wrong in the good direction,” an associate product manager at a mid-size enterprise software vendor wrote in a Slack DM, three hours after the pricing page changed. “Nobody on the call said the word safety once.”

The Safety That Saves Money Is Not the Safety Anyone Wanted

Here is the part that should make both camps uneasy.

If compliance pays, then the safety that gets built is the safety that shows up in the invoice: fewer refusals to unwind, fewer boundary escapes, fewer loops. That is real, and it is worth real money. It is also a narrow slice of what the safety people say they care about. Novel misuse, long-horizon autonomy, whatever a model does on turn nine hundred of a task nobody benchmarked — none of that reduces your per-token bill this quarter, so none of it is what an efficiency culture optimizes.

The safety-first crowd is getting the behavior it demanded, delivered by an accountant. That ought to worry it more than an unaligned model would.

Amodei’s Critics Have the Wrong Indictment

Dario Amodei published “We Must Pace the Frontier” on September 12. Ten days later his company shipped a model. The obvious indictment writes itself, and half of X duly wrote it: the deceleration talk is marketing, the release cadence is the real policy.

That’s the boring charge, and it is also the wrong one. Read the two documents together and they are consistent — the pacing framework is not a promise to ship less; it is a bet that the lab which internalizes constraints can ship cheaper. If that bet is right, safety stops being a restraint on competition and becomes a weapon inside it. The lab that can afford the alignment work undercuts the one that can’t, and it does so on Amazon Bedrock, Google Cloud Vertex AI and Microsoft Foundry on launch day, which is where the undercutting actually happens.

The people who have to update their priors are the ones — including a great many who agree with me — who spent three years arguing that American alignment work was a self-inflicted handicap in a race against less scrupulous competitors. The handicap is starting to look like a cost advantage.

And the people who wanted a cheap, unconstrained frontier are getting the cheap frontier. It simply arrives pre-trained to stay inside the lines, and the reason isn’t ethics. It’s the invoice.

Anthropic’s competitors can match $4 and $20. Matching the 85 is the part that will cost them.

Sources