So GitHub made it official this week: Copilot is out of technical preview and it's a real product now, with a real price tag. Ten bucks a month, or a hundred a year if you commit up front. Free if you're a verified student or you maintain a popular open source project, which is a nice gesture but also a little funny when you think about where the training data came from.
I've had preview access since some time last summer, back when you had to join a waitlist and just kind of wait, no idea how long, no idea if you'd ever get an email. Mine came through in September. So I've had close to nine months of using this thing for free, and now I have to decide if it's worth paying for out of my own pocket, which is a different question than "is this neat."
Is it neat? Yeah, honestly. When it's good it's very good. I write a lot of small Python scripts for scraping and data cleanup, boring repetitive stuff, and Copilot has saved me real time on writing the tenth argparse block or the twentieth regex for stripping whitespace. It guesses variable names better than it has any right to. Type a function name like parse_invoice_date and it just knows, most of the time, what you probably want next.
But it's also wrong constantly, in a specific way that's almost more annoying than just being wrong. It'll suggest something that looks completely plausible, indented correctly, right style, and then you actually read it and it's calling a method that doesn't exist, or it's subtly off by one, or it hallucinated an entire config key. I've caught myself accepting a suggestion, moving on, and finding the bug three days later. That's the real cost of these tools that nobody puts in the pricing page. Not the ten dollars. The extra fifteen minutes of code review you now owe yourself because the thing writing half your code doesn't actually understand what it wrote.
The bigger thing everyone's been arguing about, and will keep arguing about now that there's a price attached, is the licensing question. Copilot was trained on public GitHub repos, including a lot of GPL and other copyleft code, and there have been enough examples floating around of it reproducing chunks of that code near-verbatim (there's a well known one where it spits out the fast inverse square root trick from Quake, comment and all) that it's hard to pretend this is just "learning patterns" in some abstract sense. I don't have a tidy answer here. I use the tool and I also think the "we trained on your code, now pay us to use a version of your code" framing is a genuinely fair complaint from people whose repos were in the training set. Both of those things are true for me at the same time and I'm not going to pretend otherwise just to land on a clean opinion.
Small tangent: I still think the best autocomplete moment in software history is dumber than any of this. It's Xcode in like 2013 suggesting you finish typing "self" and getting it wrong four times in a row. We've come a long way from that, whatever else is true about it.
Anyway. Ten dollars a month is not a lot of money, less than most people's streaming subscriptions, and I'll probably pay it, at least to see if the GA version behaves differently than the preview build I've gotten used to. The free tier for open source maintainers is a smart PR move regardless of what you think of the underlying training data question, and it'll probably work, because most maintainers I know are broke and will take the free tool and worry about the ethics of it later, same as I'm apparently doing right now.
If you were on the waitlist and never got in, this is your way in now, just with a credit card attached. Curious if the paid version feels any different in practice or if it's the exact same model with a login wall added. I'll report back once I've actually paid for a month and used it properly instead of just poking at it out of curiosity.