Garry Tan says Fable 5.2 cost 2.8x more per solved task
A 52 task run in Claude Code puts a price tag on a near tie between two models.
Garry Tan says to keep Opus 5.5 as your coding default in Claude Code. He ran 52 coding tasks in Claude Code 2.1.322 and reports Fable 5.2 solved about as many as Opus, 82.7% against 84.6%, but cost 2.80x as much per solved task.
Opus already solves most of these tasks, though, so telling the two apart will take harder ones.
That second number is the one worth copying down. Solve rate tells you whether a model can do the job. Cost per solved task tells you what the job costs you when the model fails a few times and retries, which is the number that actually shows up on your bill at the end of the month. A near tie on accuracy and a 2.8x gap on price is not a close call.
Tan adds the caveat himself, and it matters. Opus already solves most of the tasks in his set, so a two point spread between the two models is not much signal. Separating them properly would take harder tasks than the ones he ran. This is one founder's run of 52 tasks on one version of one tool, not a benchmark anyone else has reproduced.
The business read
Most model comparisons you will see this week are solve rates, because solve rates are easy to publish and look good on a chart. Almost nobody publishes what it cost to get there. For a solo founder running agents all day, the price per completed piece of work is the only figure that changes behaviour, and it moves independently of the accuracy score. A model can be marginally better and still be the wrong default.
The practical takeaway from Tan's post is narrow and that is fine. On everyday coding tasks in Claude Code, he found no reason to switch off Opus 5.5. If your own work skews harder than the average task in a 52 task set, his data does not tell you much, and he says so.
Steal the numbers, not the vibes
One email a day. Under three minutes. Real MRR, real pricing, real launches from founders building with AI.
