I can get behind @bcherny's claim that "coding is solved" if you: have a good software factory, are using the latest frontier model (Fable), and have unlimited spend.
Unfortunately, I work at an enterprise company where we pay full API prices and are now capped at $1k spend a month. Here, coding does not feel "solved" yet.
In this scenario, models like Fable/Opus are out of reach (you'll burn through tokens too fast), and automated software factories are avoided as they also tend to drain token budget (I'm using my own software factory here, so some of this may just be my own learning curve in making it more efficient).
Given the above constraints, at my place of work I tend to use the Pi (omp) harness and reach for models like GLM 5.3 / GLM 5.3 Flash, or use Codex with Terra/Luna models. These tend to give me the best bang for my buck.
On my personal projects, I do still prefer Claude Code and Anthropic models for coding. Here I can pay for the personal Max plan and get 20–30x the API usage I would get at my place of work.
I can only imagine that token costs will continue to come down, and that coding will eventually feel "solved" in the enterprise world too.
I just don't think we're there yet.
this is super cool.
> the feedback can create a github or linear issue, which we later pick up with a cron-backed cloud agent that decides which are duplicates
i’ve been working on a feedback widget (feedbackdrop) meant for this exact use case (issue -> agent handoff). it consolidates duplicate user reports into a single issue before creating or syncing with github/linear, then keeps the status updated both ways, with all the debug context included (logs, DOM state, user journey, etc..).
curious how you’re handling spam or malicious submissions? letting raw user input flow directly into an agent loop feels like where a lot of people are going to run into trouble.