Anthropic Launches Claude Opus 5 With Effort Toggle and Self-Verification

Summary

Anthropic shipped Claude Opus 5 on July 24, bringing a 1 million token context window, self-verification that catches and corrects its own errors without prompting, and a new effort toggle that lets users dial between low, medium, and high processing intensity to balance speed against cost. Priced the same as the previous Opus model at $5 per million input tokens, it actually outperforms Anthropic’s top Fable 5 model on several benchmarks while costing half as much to run.

Why it matters

The self-verification and effort toggle are features everyday users will feel immediately. Less back-and-forth to get a good answer, and more control over how much horsepower the model throws at a task.

HTD Says

Anthropic’s Opus 5 is the first AI model I have seen where the “how hard should it try” dial is actually exposed to the user, and that is an interesting design choice. I personally haven’t tested out Opus 5 yet as I’m super conscious of my token usage as I have a lot of projects underway, but it is on my radar for something to test.

My way of doing things, honestly, is once I find a model (regardless of provider) that seems to do relatively quality work and that doesn’t just slurp away all of my usage, I tend to stick with that model. While I know that all of the AI companies are playing a game of leapfrog and trying to outdo the competition, I do also feel that you need to cautiously wade into those waters.

Test the new models on a task or project, and use your human side to gauge the quality of the output. Then weigh it against the amount of usage that task or project consumed. If it works for you, go for it. If not, roll back to an earlier model that you are more comfortable with.

- Advertisements -

Source:

Anthropic