Reasoning effort levels drive the bill: Low vs Max compared
The same question cost $0.77 at Low effort and $3.69 at Max, scoring 58 versus 66; the post also cites an OpenAI help page saying Astra at Low can outperform Sol at High.
i checked this three times and it is still true someone priced the same question to the same AI at $0.77 and at $3.69, and the only thing that moved was one dropdown: an independent lab ran Claude Fable 5.1 through its full test at every setting, Low scored 58 out of 100 for $0.77 a task, Max scored 66 for $3.69 Anthropic wrote it into its own launch post: "when set to Low or Medium effort, Fable 5.1 achieves similar or better results than Fable 5 at a much lower cost". then it shipped claude ai on Medium and Claude Code on High OpenAI's help page says the same about GPT-6 Astra: "Astra at Low effort can outperform Sol at High effort". a man on OpenAI's own forum posted the receipt: one 40-minute job on the default setting took his week from 87% to 47% i found both lines in public help pages, and the two models carry the identical price tag either way this is Graph Engineering, the layer that decides how much thinking each step of a job is allowed to buy, and it runs inside the plan you already pay for: 1. click the model name next to the send button, open Effort, set Low. in ChatGPT the same menu is called reasoning level. run a normal day like that 2. when one step decides the outcome, raise it for that message only and drop it back after: Fable 5.1 keeps everything it already read, so the switch costs nothing 3. before the next big job open Settings → Usage, note the number, run the job, look again turns out the setting decides the bill, and both companies picked it for you on Low, Fable 5.1 looks things up less and answers from memory more, so anything about this week goes back to High almost five answers from the best model on earth for the price of one, on the plan you already have bookmark this, then open the piece below: how to split one job into steps so each one gets exactly the thinking it needs, and who checks the work at the end ↓
This post quotes another post: x.com/argona0x/status/2080626046903157126