I’m very curious to understand why anyone would use Claude Opus when o3 has become a fixed 1x in cost?
It looks to me this move by OpenAI completely obliterates Claude Opus as a practical API use case. It doesn’t entirely kill Gemini 2.5 Pro usage that still dominates very long context, but I do find myself using it much less.
I’m in the same boat. I was spending close to $100/day on Claude Opus but after o3 got the price cut I’ve almost entirely swapped over.
o3 has been very good at understanding my requirements and coming up with smart solutions that do not overcomplicate things. Quality is on par with Opus at a fraction of the cost. And for software architecture I think o3 beats Opus.
I will sometimes use Gemini 2.5 Pro to identify bugs because of its huge context but my daily driver is now o3.
My experience was the opposite. I had to switch to o3 when sonnet4 and gemini2.5 were unavailable, and I ended up quitting development for the day, o3 was a disappointment. Perhaps I’m just not used to prompting it or something.
gemini2.5 used to be great, but this new June version is a step back. At least that’s show it feels. Several times it changed completely unrelated things at random. Now I review its changes really attentively, which is exhausting.
Claude Max, ChatGPT plus, Gemini Pro plan, Cursor Pro, and also Windsurf.
I have used Claude, and OpenAI’s models, from long time, and even though OpenAI gets to the front of the line, I tend to like the coding abilities, of Claude (hence my Max plan purchase).
Claudes Model’s have been consistently providing me the result that I need. And of course with a proper instruction set. Be anywhere, Cursor or windsurf or even the Claude Code CLI.
I just have a basic setup. a
info.md: A custom built ruleset that mimics, the structural traits, patterns and workflows, that the Model needs to get the ask done.
scratchpad.md: A place where these models track, share, and store knowledge across sessions related to the ask.
ask.md: just a vague description (its vague vague bro trust me lol), because info.md has the way the I write these ask’s and how to approach.
ASK START
Changelog: 1 (main) (success|failure|moderate)
… and more Changelogs
ASK END
I have used the same setup, for code generation, in which OpenAI’s models tend to, take paths that aren’t accurately aligned to the ask, compared to what Claude comes up with.
But that doesn’t mean OpenAI’s models are not good, they are, and I use for brainstorming, general analysis, research, also some personal insights tracking, and project tracking too.
Gemini 2.5 Pro, is great, and I use it, mostly for scheduling and working with other Google applications, calendar, tasks, mails, etc.
Using both o3 and Claude-4 in tandem. I have one review the others work. Claude-4 is better at following rules, but 03 definitely more innovative with hard problems, and simple solutions. I’m still smarter than both of them, for now at least …
I’m curious what techniques others are using because it’s interesting to me that everyone has such different experiences.
In my case, I always ask for a plan, then approve the plan. I’m also working with complex things such as refactors within a mono-repo. Or reducing the image size of my Dockerfile (which it will go into agent mode and test).
Overall I’ve been very impressed with o3.
To be fair I’ve been impressed with Claude Opus too but more-so when the tasks are straightforward. It’s very good at following precise instructions.
I have the same feeling about still being smarter - for now! Clearly that is about to change sooner that I imagined 6 months ago.
venkatd,
In resonse to your post, I am finding that heavy Rule usage is the way to go for all models, I have not embraced mcp servers yet, but for my codebase of 60,000+ lines in 200 files spanning 11 code types - of that 36 files are Rules comprised of 7,500 lines in Rules themselves.
I have found that frequently creating and deleteing expired Rules in my workspace to guide every aspect of development. I even have a rule about making rules.