5 Things to Try with Gemini 3 Pro in Gemini CLI(developers.googleblog.com) |
5 Things to Try with Gemini 3 Pro in Gemini CLI(developers.googleblog.com) |
Gemini 3 Pro Preview Live in AI Studio - https://news.ycombinator.com/item?id=45967211 - Nov 2025 (385 comments)
Gemini 3 Pro Model Card - https://news.ycombinator.com/item?id=45963670 - Nov 2025 (263 comments)
Currently my ranking is
* Cursor composer: impressively fast and able but not tuned to be that agentic, so it's better for one-shot code changes than long-running tasks. Fantastic UI.
* Claude Code: Works great if you can set up a verifiable environment, a clear plan and set it loose to build something for an hour
* Grok: Similar to cursor composer but slower and more agentic. Not currently using.
* ChatGPT Codex, Gemini: Haven't tried yet.
Gemini CLI has the lowest rate limits, lowest inability to steer the models (not sure that's a model or tooling thing, but I cannot get any of the Google models to stop outputting code comments constantly and everywhere) and seemingly the API frequently becomes unavailable for some reason.
Claude Code is fast, easy to steer, but the quality really degrades really quickly and randomly, seemingly by time of day. I'm not sure if they're running differently quanitized models during different times, but there is a clear quality difference depending on when in the day I use it, strangely. Haven't found a way of verifying this though, ideas welcome.
Codex CLI is probably what I use the most, with "gpt-5+high", which is kind of slow, a lot slower than Claude Code, but it almost always gets it right on the first try, and seemingly no other model+tool does instruction following as good, even if your AGENTS.md is almost overflowing with rules and requirements, it seems to nail things anyways.
For Gemini 3.0, the rate limits are very very generous. Google says rate limits refresh every five hours, and that only “a very small fraction of power users” will ever hit the limits.
> For Google AI Ultra subscribers and paid Gemini and Vertex API key holders, Gemini 3 Pro is already available and ready to enable. For everyone else, we're gradually expanding access through a waitlist.
So, not available yet. Tried with a free API key and I did not have access. I do have access on a paid API key, but I'd rather not use that with an agent. The rate limits page the docs link to currently has no info on gemini-3-pro rate limits. Seems to me this is really only for users on the $200/mo subscription plan. Somewhat odd, given the model is already GA in every other coding agent as I understand
Going to download Gemini CLI right now™ and see how it performs™ against Cursor, Claude Code, Aider, OpenCode, Droid, Warp, Devin, and ForgeCode.
In the end, the model only ran `git bisect` (if we're to believe the video at least) for various pointless reasons, it isn't being used for what it's usually used for. Why did it run bisect at all? Well, the user asked the LLM to use `git bisect` to find a specific commit, but that doesn't make sense, `git bisect` is not for that, so what the user is asking for, isn't possible.
Instead of the model stopping and saying "Hey, that's not the right idea, did you mean ... ?" so to ensure it's actually possible and what the user wants, the model runs its own race and start invoking a bunch of other git commands, because that's how you'd find that commit the user is looking for, and then finally does some git bisecting stuff just for fun, it had already found the right commit.
I think I see the same thing when letting LLMs code as well. If you give them some work to do that is actually impossible, but the words kind of make sense, and it'll produce something but not what you wanted, I think they're doing exactly the same thing, bypassing what you clearly instructed so they at least do something.
I'm not sure if I'm just hallucinating that they're acting like that, but LLMs doing "the wrong thing" has been hitting me more than once, and imagining something more dangerous than `do a git bisect`, it seems to me like that video is telling us Gemini 3 Pro will act exactly the same way, no improvements on that front.
Also, do these blog posts not go through review from engineering before they're published? Besides the video not really showcasing anything of interest, the prompt itself doesn't make any sense and would have been caught if a engineer who uses git at least weekly reviewed it before.
> gemini
It seems like you don't have access to Gemini 3. Learn more at https://goo.gle/enable-preview-features To disable Gemini 3, disable "Preview features" in /settings. • 1. Switch to gemini-2.5-pro • 2. Stop Note: You can always use /model to select a different option.
Google never disappoints with their half-ass-launches.
In an ideal world, this workaround would not be needed
¯\_(ツ)_/¯
Video is really a terrible format for terminal demos, you've got to pause it as the screen flashes text faster than you can read...
But what is that actually doing? It looks like when it's running the git bisect, it already knows what the commit is, and could have just returned it. The only reason it ran any bisecting at all, was because the user (erroneously) asked it specifically to use git bisect. It didn't have to.
And the MCP field is already pretty heavily saturated with TypeScript and JSONSchema, so using TS for it is a very ergonomic experience. Also since it's written in TypeScript, it's much more easy to integrate it with editors like VSCode (or Google's new Antimatter) which are built on top of Electron.
think about how it would do the same for python/golang/lua/etc.
I would never write a TUI like this and simply `eval` in the same scope as the TUI frames.
you shell out and run it with the language engine required.
Opencode has the ability to lean on LSP and Formatters, but these are not required to be written in the same language as the TUI