One PR from an AI agent on a real ticket costs about $82 in Opus tokens. The same run on open-weight models costs about $13.
I pulled our agent logs to see how many tokens it actually takes to get from a ticket to a PR. The token counts are real, the prices are just the providers' public list prices.
On small tickets, things like fixing an endpoint or adding a validation, we had 931 PRs across 8 repos on different stacks. The average was around 1.25M tokens per PR: 240k input, 67k output, and the rest cache reads. That comes to $3.35 per PR on Opus, $2.01 on Sonnet and $0.63 on open-weight models like GLM, DeepSeek or Kimi.
Real tickets are a different story. On a live project with features coming straight from Jira, planning included, a PR took 10 to 100 times more tokens. That's where you get ~$82 on Opus, ~$49 on Sonnet and ~$13 on open-weight.
Now imagine a team of 30 engineers, each shipping 20 of these PRs a month. That's around $49k a month on Opus, $29k on Sonnet and $7.6k on open-weight. Over a year it's $590k vs $350k vs $90k for exactly the same work.
To be fair, I priced Opus and Sonnet on the same token counts the open-weight models used, and frontier models sometimes get there with fewer tokens. But even if Opus needed half as many, it would still be 2.5–3x more expensive.
And there's one cost that doesn't show up in this math at all: where your code goes while you're paying for those tokens.
At YAGNI we run these numbers every day, and that's why we build on open-weight models. Have you actually calculated what your team spends per PR?