Anthropic’s most powerful Claude model is leveling up, with the company saying in a blog post Thursday that Claude Opus 4.6 will be even better at coding and creating projects on the first go.
Claude Opus 4.5 is already a robust coding model, with its November release sparking Claude Code’s viral vibe-coding moment over the holidays. Claude’s proven coding prowess and new Cowork feature have Wall Street anxious, with many tech stocks falling in recent weeks, over concerns that people won’t need software products in the future.
Anthropic said the new model is more focused on solving the biggest challenges, such as the inner workings of complex apps, while also handling the simpler steps more quickly.
As a reasoning model, Opus 4.6 works by breaking down the steps it needs to take so that it can do what you ask and putting together a plan before starting. It’ll also go back and check its work on those steps, sometimes making multiple attempts without you asking.
Sometimes the model can spend too much effort on a task, which Anthropic said can be resolved by reducing its effort level from the default “high” setting.
Anthropic Super Bowl Commercials Pinky Promise No Ads in Claude
The Claude Opus models are available for paying Claude users on the Pro, Max, Team and Enterprise plans. The cheapest of those, Pro, costs $20 a month (or $17 a month if you pay annually). The Pro plan comes with usage limits for Opus, which users can hit after a few hours of vibe coding and then have to wait several hours for it to reset.
Aside from Opus, Anthropic has smaller, less powerful models in Sonnet 4.5 and Haiku 4.5.
A first look at Claude Opus 4.6
To test out the new model, I tasked it with creating a trivia app that operated by voice. This process took several iterations over about an hour, but Claude churned each one out pretty quickly. It was by no means autonomous — I identified glitches and offered ideas for solutions, although some of my suggestions backfired as we ran up against the constraints of building entirely within an HTML file.
Enlarge Image
The app Claude Opus 4.6 built for me really leaned in on the Jeopardy! question style.
The experience was not much different from when I tried similar tests with Opus 4.5, although this seemed to go a little bit faster. The model grasped the idea of what I was trying to do from the get-go, which has not always been the case with AI projects, and the trivia questions it came up with, once I told it to make them tricky, were pretty well crafted. Most of them were accurate, too, although one of the (many) art history questions asked me to name the artist (Edvard Munch) but told me the correct answer was the painting’s title (The Scream).
The downside of the speed is that I burned through the usage limit on my Pro plan in about 90 minutes — just as I got the app to work pretty much seamlessly — and couldn’t make one final request: for a database of more than 100 questions. That will have to wait a few more hours.
Read the full article here
