For the past 3 days I've seen garbage output, API errors, cognitive degradation and it's also been super aggressive about its bias towards action i.e. impossible to have a conversation because it's always defaulting hard to coding. Anyone else seeing this? Was a completely different product early last week.
gemini-cli is trash, use a different agent tool or framework
or consider building your own, it's not that hard and you learn a lot about why these tools have certain peculiarities in the process
tl;dr the system prompts and context engineering matter a lot, they people building these tools have unlimited AI tokens and toy problems, so they aren't great when you try them on IRL problems
No it's not. Gemini 3 is unbelievable when it works. Faster. Bigger context window. Higher intelligence as proven on ARC-AGI-2. Please skip the all-lower-case-troll comments and add actual data to the conversation.
Big projects that are hard. Rust and CUDA for example. Particularly where you have many files that need to be fully read in for context, before planning or implementing something.
Also very hard algorithmic problems. Also bugs that Claude Code or Codex CLI are completely stuck on and can't find, Gemini 3 using Gemini CLI will go in and find the impossible. It's incredible when it does this.
Claude Code burns through context like crazy. With or without Serena.
Codex CLI using GPT 5.1 Codex Max Extra High is really good and uses context more efficiently. But Gemini 3/CLI is 10X more efficient with context.
But for the past 3 days I've had to endure absolute torture and I think it's the actual Gemini 3 back-end model that is having major issues.
I feel like Codex is the middle ground. You can define a project, break it into bite sized chunks, but still lift a reasonable amount. Claude with Opus 4.5 right now chews up context at an eye watering rate. It's really unfortunate because it's really good.
or consider building your own, it's not that hard and you learn a lot about why these tools have certain peculiarities in the process
tl;dr the system prompts and context engineering matter a lot, they people building these tools have unlimited AI tokens and toy problems, so they aren't great when you try them on IRL problems
reply