It is really good at grep and will make multiple grep calls in parallel.
And when I open it in the parent directory of a bunch of repos in our codebase, it can very quickly trace data flow through a bunch of different services. It will tell me all the files the data goes through.
It's context window is "only" 200k tokens. When it gets near 200k, it compresses the conversation and starts a new conversation..... which mostly works but sometimes it has a bit of amnesia if you have a really long running conversation on something.
> My bet is their model doesn't realistically compare to any of the frontier models.
I've been using composer-1 in Cursor for a few weeks and also switching back and forth between it, Gemini Flash 3, Claude Opus 4.5, Claude Sonnet 4.5 and GPT 5.2.
And you're right it's not comparable. It's about the same quality of code output of the aforementioned models but about 4x as fast. Which enables a qualitatively different workflow for me where instead of me spending a bunch of time waiting on the model, the model is waiting on me to catch up with its outputs. After using composer-1, it feels painful to switch back to other models.
I work in a larg(ish) enterprise codebase. I spend a lot of time asking it questions about the codebase and then making small incremental changes. So it works very well for my particular workflow.
Other people use CLI and remote agents and that sort of thing and that's not really my workflow so other models might work better for other people.
In my experience the software engineers that often get promoted crank out as many features as possible with almost no regard to the defect rate. If you do things at a more deliberate pace, you can be regarded by product or management as slow and unproductive. Real incentives exist in software development that push for quantity over quality.
This is not analogous to the failure of an anesthesiologist to properly sedate a patient. This is a process failure. Clearly the proper amount of QA was not in place for whatever reason and they need to re-examine their approach. Of course their process needs extra diligence given the cost of failure.
It would be a real shame if an individual is punished for this instead of examining the process and system of incentives that led to this failure.
And when I open it in the parent directory of a bunch of repos in our codebase, it can very quickly trace data flow through a bunch of different services. It will tell me all the files the data goes through.
It's context window is "only" 200k tokens. When it gets near 200k, it compresses the conversation and starts a new conversation..... which mostly works but sometimes it has a bit of amnesia if you have a really long running conversation on something.