GPT 5.6 Sol Pro

GPT-5.6 Sol running in pro reasoning mode -- the same underlying model given substantially more work before it answers. OpenAI recommends pro for difficult tasks that can tolerate higher latency and token usage, which is exactly where a first-pass answer has already failed. Everything else, including the million-token context, matches Sol.

This is not a different model from GPT-5.6 Sol. It is the same weights served at the deliberate end of Sol's own dial, which is exactly how OpenAI sells it: in ChatGPT, Pro and Enterprise subscribers can select GPT-5.6 Sol Pro when they want the highest-quality result on a complex task, while everyone else reaches Sol through the ordinary effort settings. Pick it when a first-pass answer has already failed, not to make a routine answer marginally better.

What the extra spend buys is described precisely in the launch post. Max effort gives the model more time than xhigh to reason, explore alternatives, run checks and revise its approach. Ultra goes further still by coordinating four agents in parallel by default, and OpenAI's own charts show what that is worth: on Terminal-Bench 2.1 the score moves from 88.8% to 91.9%, on BrowseComp from 90.4% to 92.2%, and on SEC-Bench Pro from 71.2% to 74.3%. The lab frames all three as the score-latency frontier shifting up and to the left - a better answer, and sooner, because the work is split rather than serialised.

What it costs is the other half of that sentence, and it is not the rate. The per-token price is Sol's price; what changes is how many tokens a single answer consumes and how long you wait for it. OpenAI is explicit that ultra trades higher token use for stronger results, and that its highest-effort settings are for problems that reward a greater investment of time and compute. On a lookup, a rewrite or anything with a known shape, that investment buys nothing you can measure and the bill is real.

Everything else is unchanged from Sol: the same million-token context, the same February 2026 knowledge cutoff, the same tool use and structured output, and the same caching behaviour, where GPT-5.6 introduced explicit cache breakpoints and a thirty-minute minimum cache life, charged cache writes at 1.25 times the uncached input rate and kept the 90% discount on cache reads. On a long deliberation that caching is the difference between a manageable bill and a surprising one.