I wouldn't recommend it, performance is worse than the alternatives. Its newer tokenizer makes the same task more expensive, and while the automatic self-review Anthropic added sounds useful, it further inflates token consumption. My verdict: Factoring in retries and everything else, Opus 4.8 works out cheaper. Sonnet 5 is a clear improvement over older Sonnets, but it doesn't beat Opus 4.8 as some assume. Thus the applications it's useful for is VERY limited. P.S. what you COULD use it for is like creative writing, etc. for example and set the thinking effort to low, for example.

@Wendy Breakstone No, use either Sonnet at low level thinking, or Opus 4.8 normally as is. Sorry, if it was unclear above...

Great call. The fastest way to "get it": HINT: when stuck: /help shows every built in command. 1. Grab it from claude.com/claude-code and run claude inside a real project folder (not a toy). It's a terminal agent, it reads and edits your files and runs commands, not a chatbot you copy-paste from. 2. First thing: type /init, it reads your codebase and writes a CLAUDE.md (its "memory" for that project). Then just talk to it like a teammate: "explain this repo," "add X," "fix this bug." 3. Learn by handing it one real task you'd normally do by hand. Watch it work, correct it mid-flight, you'll click in a single session. 4. It asks permission before edits/commands. Start there, allow-list the safe stuff once you trust it. 5. And it's not just for coders: any repetitive file/computer task or a small tool/app idea is the sweet spot (people here have built trackers, assistants, etc.). Additional tip: use the slash command /insights, after a while, then an HTML will be written for you that tells you how to improve working by using claude code. None of these commands are required, by the way. You can get everything done just talking to it normally. They're there for when curiosity strikes. My one real suggestion: experiment. Poke at it, try things in a throwaway folder, see what it does. That's the whole unlock. P.S. Hope it helps.

Hi all ,Holger here. 👋 I've been lurking in this group a while (mostly active over in Mark's Early AI-Adopters), and after today's Claude Code thread it felt like time to actually introduce myself. Background: I work as an engineer, day job is industrial software / ops tooling , but the thing I can't stop doing is building my own tools on top of Claude Code: worklog reconcilers, a session-history browser, git helpers, and a bigger research-and-build platform I've been growing called Prometheus-Crystal-Lab. What I'm probably most useful to swap notes on: - Permissive prompting — you get more from these models by granting them room to act than by dictating every step (I posted the core ideas in Mark's group). The way I hold it: permission past the habit, never past the rule. - Tools that referee instead of flatter, I'm a bit obsessed with building things that tell me when I'm wrong, or when a "new" idea is actually already known, instead of cheerleading. The honesty is the feature. Claude Code as a daily driver, happy to help anyone getting started (saw the thread!).

@Endy Cheung Both exist now in a 4.6 version: Opus AND Sonnet.

@Endy Cheung Exactly. It's closing the gaps between opus and sonnet more in some areas as well. So, if you're doing deep researches or scientific stuff, or need the extending thinking capabilities and better reasoning, then opus may be still advised. 😊