OpenAI Calls 60% of Visible Reasoning 'Non-Load-Bearing,' Announces Plan to Sell More of It

MENLO PARK — OpenAI said Thursday it would increase the default allocation of visible reasoning tokens in its next model family after internal research confirmed that approximately 60% of a model's displayed thinking steps have no measurable impact on the accuracy of its final answers — a finding the company characterized as "a product opportunity."

The research, consistent with academic work published this week by Weiyan Shi and colleagues at Columbia showing 30 to 60% of reasoning steps carry minimal causal weight, indicated that the non-load-bearing tokens function primarily to provide users with what one internal document called "a sense of epistemological companionship."

"The model doesn't need to hedge for three paragraphs before answering a question about train times," said a researcher briefed on the findings. "But users who see the hedging rate the answer as significantly more reliable."

In response, OpenAI plans to offer three Thinking Mode tiers in GPT-6: Standard (answers arrive after the model visibly works through the problem), Deliberate (the model visibly works through the problem and a related problem it did not ask about), and Executive (the model works through the problem, acknowledges several adjacent concerns, schedules a follow-up with itself, and then answers).

The Executive tier will add an estimated 2,100 tokens per query and is expected to account for 70% of API revenue.

Independent researchers noted that the findings suggest modern reasoning models have, without explicit instruction, converged on the communication style of a McKinsey consultant: correct at the end, thorough in between, and operating primarily to manage the client's anxiety.

"The model learned that humans don't trust fast answers," said one researcher. "So it slows down. Not to think better. To seem like it is."

OpenAI declined to comment on whether future models would be trained to appear uncertain in order to build rapport, noting only that the feature is "already shipping."