Opus 5.5 beats GPT-6 Sol across quality, coding evals and cost, a long review says
A long review lists Opus 5.5's advantages in finished output quality, coding benchmark scores and cost, and notes most people preferred Opus over GPT-6 Sol in same-prompt comparisons.
opus 5.5 is a very interesting release, for a few reasons: 1. it shows claude has pulled ahead on making things that actually look good almost every opus 5.5 post blowing up these past two days is a finished thing it made in one go: a playable mario kart with mario and luigi in it. 100 HTML pages in one run, told to look good and never repeat, and not a single broken file. the previous best on that test was fable 5.1. even the classic pelican-riding-a-bicycle test got turned into a 3D scene. on an eval of professional deliverables, it beats fable 5.1 by 143 Elo, ahead on both analysis and presentation. yesterday i had it make a 29-second animated short. the story, every frame, and the music were all its work. i didn't write a single line of code. i showed a few friends, and the first thing every one of them asked was how i made it. 2. this time medium scores higher than max on the official coding eval, medium scored 54.6%, a bit above max and above gpt-6 astra's best score, at about a fifth of astra's cost. probably the wildest benchmark result of this launch. going up to max just costs more: it uses 1.6x opus 5's output tokens per task, and the extra spend buys almost no coding score. so in claude code, stick with the default medium day to day, and use /effort to bump it up when you get stuck. on 2.1.280+, switching effort mid-session won't break your prompt cache, so no need to restart. 3. quite a few people are moving from codex back to claude people who cancelled over opus 5 are coming back, and some cancelled codex the same day. one person ran high effort for 5 straight hours and used 8% of their weekly limit. same-prompt comparisons with gpt-6 sol keep coming, and most side with opus. openai cut sol and luna API prices in half the same day and said subscription usage would go further too. but some $20 plus users are finding sol still hits the limit after a few rounds of discussion, with no real improvement. i tried it on coding myself, and it did save some. on raw capability, opus is ahead of gpt on most evals right now. but switching costs almost nothing. the people coming back today could leave again next month. when the crown changes hands every few weeks, is anyone still loyal to an AI company?