Technical Context
I dug into what’s really going on with GPT-5.6 Sol Max because the story looked odd: people expect maximum quality, yet they get a scaled-down mode without even realizing it. For AI implementation, that’s a nasty surprise, especially if you’re building chains where model behavior must be predictable.
The confirmation is quite down-to-earth. Regular ChatGPT doesn’t have some universal Max auto-switch for everyone, but in ChatGPT Work and Codex, the reasoning effort level max has to be enabled manually through settings and configurations. If you don’t, the model doesn’t run at its peak reasoning mode.
And here I wouldn’t lump three things together: vanilla Sol, max mode, and separate enhanced modes like ultra or Sol Pro. They’re not the same. Max isn’t a magic “make it perfect” button—it’s a heavier reasoning mode with increased time and token consumption.
Tokens are predictable but unpleasant. OpenAI openly states that max and ultra increase consumption by design, yet it gives no precise public multipliers. The chart from the tweet only confirms what I already see in actual runs: Max guzzles a lot, and on long tasks that’s no longer cosmetic—it’s an architecture factor.
Business and Automation Impact
The first consequence is simple: if a team thinks it’s testing top-tier mode but hasn’t turned on Max, their model and prompt comparisons get skewed. Then false conclusions about quality, SLA, and ROI start cropping up.
Second: I wouldn’t advise building AI automation directly on Max from the get-go. Better to keep routing: default for high-volume tasks, Max only for expensive touchpoints where deeper reasoning is truly needed.
Third: budgets. If an agent writes code, validates hypotheses, or runs multi-step workflows, one config slip easily turns into thousands of extra tokens per run. These are exactly the things we at Nahornyi AI Lab usually clean up before launch, because AI integration breaks not in demos but at scale.
If you have a similar story and costs have already started creeping up, you can just break down your scenario layer by layer: where Max is needed, where the standard mode suffices, and where it’s worth rethinking your AI architecture entirely. At Nahornyi AI Lab, that’s exactly where I normally start, because proper automation with AI should save resources, not burn them by misconfiguration.