Transcript
0:00 We just released Opus 5.5, and compared to Opus 5 it's faster, it's a lot better at telling you what it did, and it's 20% cheaper per token. On a Pro, Max or Team subscription your limits go 25% further, and because it needs fewer tokens to get to a result, a task is 40% cheaper. So this all sounds great in theory, let's run them side by side and see what actually changes. Okay so here we have the same project, same prompt, both on medium effort, with Opus 5 on the left and 5.5 on the right.
0:29 So there's a rounding bug, refunds on orders with a discount code come out a few cents off. They both read the refunds code, find where the discount gets split, fix it, run the tests, and so on. But 5.5 is done in under a minute, and its whole response fits right here on the screen. So it just tells me what it changed, the tests it ran, the one thing it ran into, and the one question it has for me.
0:50 And in the meantime Opus 5 is still reading, the README, the money helpers, the test suite, and eventually of course it also comes back with a response, but it just went through way more of the project first, with more tool calls and a more verbose response. So in this case 5.5 was 30% faster than Opus 5. And I'm running this on a Max subscription, so you can see my limits right here. So 5.5 used about 4% of the 5-hour limit, and Opus 5 used 6%.
1:17 Now of course this is just my run and yours will look different, but I ran this a bunch of times and every time 5.5 used less of my 5-hour limit than Opus 5 did. Opus 5.5 also defaults to medium effort, and effort is basically how much thinking it puts into a turn before it acts. So everything you just saw was on medium effort, and for most of what I do on a daily basis that's enough, but there are some cases where I still go to high.
1:43 So for example here I ask it to rename a field in the orders handler. So it reads the handler, renames the field, the tests pass, done. Cool, exactly what I asked for. But the thing is that that field doesn't only live in the handler. The mobile app sends it as well, except there it's called customer_ref, with an underscore, and it goes through the serializer. So when Opus searched for customerRef, that file just never came up, and it didn't touch it.
2:07 So that fix isn't wrong, it's just not entirely complete. So in situations like these, where a fix would span multiple files or the entire project, I still go to high. So now on high I tell it to finish the rename everywhere, and this time it actually traces what sends that payload, finds the mobile client and the serializer, and renames it there as well. And it even adds a little test to make sure it won't happen in the future.
2:28 I do the same for any change that touches the whole project, but once that's done I go back to medium. More generally, another easy win is running your subagents on a smaller model. A subagent that's only exploring the codebase doesn't need Opus-level reasoning. So if you have custom subagents you can set the model to Sonnet right in the frontmatter, or use the CLAUDE_CODE_SUBAGENT_MODEL environment variable in settings.json to set the default for all of them in just one place.
2:52 So as your daily driver, Opus 5.5 on medium effort is great. I only go to high whenever there's a change in many files at once, and any read-only subagents I run them on Sonnet, and sometimes even Haiku. A useful command we have is /claude-api prompt-audit. Even if you're just using Claude Code and not the API, this goes through your CLAUDE.md and your skills and tunes the instructions in there for Opus 5.5. So overall, a big improvement over Opus 5.
3:18 It's faster, your limits go further, and I just find it a lot nicer to work with. So go try out Opus 5.5 and let me know what you think.