GPT-5.2 sometimes does this too. Opus-4.5 is the best at understanding what you actually want, though it is ofc not perfect.
It's also just not as good at being self-directed and doing all of the rest of the agent-like behaviors we expect, i.e. breaking down into todolists, determining the appropriate scope of work to accomplish, proper tool calling, etc.
I have developed decent intuition on what kinds of problems Codex, Claude, Cursor(& sub-variants), Composer etc. will or will not be able to do well across different axes of speed, correctness, architectural taste, ...
If I had to reflect on why I still don't use Gemini, it's because they were late to the party and I would now have to be intentional about spending time learning yet another set of intuitions about those models.
Codex is the best at following instructions IME. Claude is pretty good too but is a little more "creative" than codex at trying to re-interpret my prompt to get at what I "probably" meant rather than what I actually said.
[0] based on user Thumbs up/Thumbs down voting
I am familiar with copilot cli (using models from different providers), OpenCode doing the same, and Claude with just the \A models, but if I ask all 3 the same thing using the same \A model, I SHOULD be getting roughly the same output, modulo LLM nondeterminism, right?
It's on the top of most leaderboards on lmarena.ai
My default everyday model is still Gemimi 3 in AI Studio, even for programming related problems. But for agentic work Antigravity felt very early-stages beta-ware when I tried it.
I will say that at least Gemimi 3 is usually able to converge on a correct solution after a few iterations. I tried Grok for a medium complexity task and it quickly got stuck trying to change minor details without being able to get itself out.
Do you have any advice on how to use Antigravity more effectively? I'm open to trying it again.
Both Claude and ChatGPT were unbearable, not primarily because of lack of technical abilities but because of their conversational tone. Obviously, it's pointless to take things personally with LLMs but they were so passive-aggressive and sometimes maliciously compliant that they started to get to me even though I was conscious of it and know very well how LLMs work. If they had been new hires, I had fired both of them within 2 weeks. In contrast, Gemini Pro just "talks" normally, task-oriented and brief. It also doesn't reply with files that contain changes in completely unrelated places (including changing comments somewhere), which is the worst such a tool could possibly do.
Edit: Reading some other comments here I have to add that the 1., 2. ,3. numbering of comments can be annoying. It's helpful for answers but should be an option/parameterization.
I've been experimenting with small local models and the types of prompts you use with these are very different than the ones you use with Claude Code. It seems less different between Claude, Codex, and Gemini but there are differences.
It's hard to articulate those differences but I think that I kind of get in a groove after using models for a while.
It fails to be pro-active. "Why didn't you run the tests you created?"
I want it to tell me if the implementation is working.
Feels lazy. And it hallucinates solutions frequently.
It pales in comparison to CC/Opus.
It won't make any changes until a detailed plan is generated and approved.
Using Gemini 2.5 or 3, flash.
With humans you can categorically say ‘this guy lies in his comments and copy pastes bullshit everywhere’ and treat them consistently from there out. An LLM is guessing at everything all the time. Sometimes it’s copying flawless next-level code from Hacker News readers, sometimes it’s sabotaging your build by making unit tests forever green. Eternal vigilance is the opposite of how I think of development.
I'm also mostly on Gemini 3 Flash. Not because I've compared them all and I found it the best bar none, but because it fulfills my needs and then some, and Google has a surprisingly little noted family plan for it. Unlike OpenAI, unlike Anthropic. IIRC it's something like 5 shared Gemini Pro subs for the price of 1. Even being just a couple sharing it, it's a fantastic deal. My wife uses it during studies, I professionally with coding and I've never run into limits.
All providers are opt-out. The moat is the data, don't pretend like you don't know.
the tools its built with seem to suck, but it can cook with serena mcp.
the flash models seem to get better results than the pro ones as far as ive seen, but theres not a big difference
The TLDR: The $20/40m cost is more reflective of what inference actually costs, including the amortised cost of the Capex, together with the Opex.
The Long Read:
I think the reason is because Anthropic is attempting to run inference at a profit and Google isn't.
Another reason could be that they don't own their cost centers (GPUs are from Nvidia, Cloud instances are from AWS, data centers from AWS, etc); they own only the model but rent everything else needed for inference so pay a margin for all those rented cost centers.
Google owns their entire vertical (GPUs are google-made, Cloud instances and datacenters are Google-owned, etc) and can apply vertical cost optimisations, so their final cost of inference is going to be much cheaper anyway even if they were not subsidising inference with their profits from unrelated business units.
It's pretty much trial and error.
I tried using ChatGPT via the webchat interface on Sunday and it was so terse and to the point that it was basically useless. I had to repeatedly prompt for all the hidden details that I basically gave up and used a different webchat LLM (I regularly switch between ChatGPT, Claude, Grok and Gemini).
When I used it a month ago, it would point out potential footguns, flaws, etc. I suppose it just reinforces the point that "experience" gained using LLMs is mostly pointless, your experience gets invalidated the minute a model changes, or a system prompt changes, etc.
For most purposes, they are all mostly the same i.e. produce output so similar you won't notice a difference.