Theo ran up to 25 Opus threads on 20% of a $200 plan
The limit on parallel agents is looking less like billing and more like how many threads one person can watch.
Theo says he had between 8 and 25 threads running most of the day, all on Opus 5.5, and still burned less than 20% of a $200 Claude subscription. His verdict on the plan was that the value you can get out of it right now is insane.
The amount of value you can get out of the $200 Claude plan rn is insane.
For reference, my Codex subs are getting absolutely hammered.
The interesting part is where the spend actually landed. Theo says his Codex subscriptions are getting absolutely hammered, because he uses Sol for reviews, and he adds that he has Opus reviewing the same code in parallel most of the time. So the reviewer, not the writer, is the thing eating the budget.
He also qualifies the number, which is worth noting for anyone trying to reproduce it. This was a day of actually working and trying to land a bunch of changes, with no death loops running, and he names the ts-rust rewrite and lust-lang as examples of the kind of long-running grind that would have changed the math. A day of shipping small changes is not the same workload as an agent chewing on a rewrite for hours.
One data point, one vendor claim
Separately, Theo says T3 Code users have consistently more than 3x'd their token throughput after moving to the app, which he frames as a question of how well a client utilizes your AI subscriptions. He ships T3 Code, so treat that as a vendor claim rather than a measurement, and he does not publish the underlying numbers.
For working engineers, the usable signal here is a single person's usage log, not a benchmark. One developer on a top-tier plan did not come close to the ceiling while supervising double-digit parallel threads. If that holds for you, the binding constraint stops being the bill and becomes attention, which is roughly the problem Theo has been building around with his multi-machine orchestration setup.
The other half of the picture is speed rather than quota. Vercel Developers announced DeepSeek V4.1 Flash 1017 Ultrafast on AI Gateway, pitched for coding, reasoning and agents at over 1,000 tokens per second, and Malte Ubl's reaction was that once you have experienced 1,000 tokens per second it is hard to go back to anything else.
Once you experienced 1000 tokens per second it is hard to go back to anything else
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.