Theo says Sol Ultrafast just barely fits the $500 plan
He pulled a week of his own prompt counts, then threw out the one job that would have blown the number up.
Theo ran the numbers on his own usage and concluded he can probably main GPT-6.1 Sol Ultrafast on the $500 plan while just barely hitting his limits. His total usage came out at 583 prompts a week, and he says that figure is everything, not only the Sol traffic. He also ran the same comparison against Opus on his $200 plan and attached those numbers alongside.
Turns out I can probably main GPT-6.1 Sol Ultrafast on the $500 plan and just barely hit my limits
The caveat is the one he put first. He excluded the ts-rust work from the count, describing it as a super long running YOLO mission rather than actual day to day usage. That is the same TypeScript compiler rewrite in Rust he shipped earlier, which means the week he measured is a week with the single largest job manually carved out of it. A second pass turned up more prompts overall, but he says the extra ones were mostly noise from broken PR monitors.
What the number is and is not
This is one engineer counting his own prompts, not a benchmark and not a pricing analysis. Prompt counts are a crude unit anyway, since a prompt that kicks off a long agent run and a prompt that fixes a typo both count as one. The useful part is the shape of the conclusion. Theo is not saying the plan is generous, he is saying it just barely holds, which is a different claim and a more fragile one. Anyone whose week looks heavier than 583 prompts, or whose week includes the kind of long running mission he deliberately excluded, should expect to land on the other side of the line.
He had already given a feel-based read on the tier below. On the $200 Codex plan with 6.1 Sol he said it was not hard to hit the limit, but not too easy either, and that the $200 Claude Code plan felt meaningfully less limited with Opus use. In the same post he wrote off one option entirely.
Astra feels pretty close to unusable.
Astra feels pretty close to unusable.
That is a usability judgment, not a limits judgment, and he did not elaborate.
For working engineers
Plan math only transfers if your volume looks like his. The thing worth copying here is not the conclusion but the method, which is counting your actual prompts over a real week before deciding which tier you are on. The alternative is the version ThePrimeagen hit earlier, where an hour of prompting ate most of a weekly allowance and the discovery arrived after the fact.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.