Users spotted that GPT-5.6 Sol's internal reasoning budgets had been slashed across all tiers, with the top tier dropping from 960 to 128. The suspicion? OpenAI was quietly nerfing the model to save compute.
https://twitter.com/thsottiaux/status/2076495156757577895
Tibo Sottiaux, who leads the Codex team, responded that the team had adjusted reasoning intensity to investigate why subscription quotas were draining so fast. The experiment has been rolled back, and OpenAI says it wasn't a permanent downgrade.
OpenAI also confirmed that the 372K context window was causing higher-than-expected quota deductions, so it temporarily reverted to 272K. The high and xhigh tiers were also making more multi-agent calls than intended — that's being fixed, too.
On the upside, OpenAI says it completed some reasoning optimizations, and GPT-5.6 Sol subscribers are now getting roughly 10% more usage per tier.