Muse Spark 1.3 reportedly scored higher than GPT-5.6 and Opus 5 on DeepSWE 1.1.
Rolling out through Meta’s model APIs “Watermelon” upgrade coming soon Open-weight Muse Spark version on the way The AI model race is accelerating every week.
Senior devs mostly use AI agents for hard debugging.
But agents don't investigate — devs still feed logs, set breakpoints, paste results into chat. I gave Opus 5 a real bug with a "scientific debugging" prompt (via Superpowers). Result: it just wrapped everything in try/catch.
Claude's new model "Fable 5.1" is twice the price of Opus 5.
However, what the official documentation recommends is not to leave everything to Fable, but to use Opus 5 normally, and use Fable only when the quality is insufficient. The saying that you don't need to use the most expensive tool every time applies to all jobs.
Meet Muse Spark 1.3, Meta's new model that almost beats Fable 5 and Opus 5 but is 12x cheaper.
Same prompt as Fable 5.1 in one test: 1 minute / almost free vs 70 minutes / $13.
You have reached the end of the archive
All of Claude Opus 5