Claude Opus 5: more power for the same price, but not for free

Anthropic has switched on its new flagship model, Claude Opus 5. It went live on July 24 and landed everywhere at once: the API, Amazon Bedrock, Google Cloud and Microsoft Foundry, plus the Claude apps. Anthropic is not selling this as a bit of polish. It calls it a leap. On coding and knowledge work, the company says Opus 5 beats every other model while costing half as much as its own flagship, Fable 5.

Who actually gets Opus 5

On Claude Max, Opus 5 is now the default model. On Pro, it is the strongest option you can pick. Anyone on the free tier stays with Sonnet 5 and does not see Opus 5 at all.

That is the difference from the noise around Kimi K3, which anyone can try without hitting a paywall. Anthropic is sticking to its line: the best model costs money.

The half-price claim is not about the predecessor

Half price sounds like a discount. It is not one. Opus 5 costs 5 US dollars per million input tokens and 25 US dollars per million output tokens through the API, exactly what Opus 4.8 costs. The half refers to Fable 5, the pricier model a tier above. If you need speed, Fast mode doubles the rate and currently runs only through the Claude API.

For private users those token prices barely matter, because the subscription gets billed, not the individual token. We have worked out what the various AI subscriptions cost in 2026 and which one pays off for whom.

The model now thinks on its own

The change you will notice most in daily use is not on the benchmark slides. On Opus 4.8, a request ran without any internal reasoning unless you switched it on. On Opus 5, it is on out of the box. The model decides for itself when to think and how deeply.

Five levels control it: low, medium, high, xhigh and max. High is the default. Anthropic explicitly recommends reaching for low and medium more often, because they deliver usable results on a fraction of the tokens. Two side effects are worth knowing. Answers run longer by default, and the model checks its own work without being asked. Both eat into your quota. If you keep hitting the usage limit on Pro, you are better off dialing the level down.

One million tokens of context, with nothing smaller

Opus 5 works with a context window of one million tokens. That is the default and the maximum at the same time, and there is no smaller variant. Output can run to 128,000 tokens. In practice, long contracts, entire manuals or sprawling project folders fit in one piece, with no need to cut them up first.

Safety: tighter and looser at once

Anthropic calls Opus 5 its best-aligned model so far, with the lowest rate of deceptive behavior it has measured. On security topics it has loosened the reins at the same time. The filters for cybersecurity requests are meant to step in around 85 percent less often than on Fable 5. When a request does get blocked, the older Opus 4.8 takes over automatically inside the Claude apps. At building attack code, Anthropic's own tests still put Opus 5 well behind Mythos 5.

One caveat belongs here. Every top result named above comes from Anthropic's own measurements. Independent comparisons with OpenAI's GPT-5.6 and the Chinese challengers are still missing. Opus 4.8 stays available on all platforms for now.