GPT-5.6 Sol: Pricing, Reasoning Settings and the Ultra Mode

What GPT-5.6 Sol is

GPT-5.6 Sol is the top model of OpenAI's GPT-5.6 family, released on 9 July 2026 alongside Terra and Luna. OpenAI frame the release around getting more useful work from every token. It carries a 1,050,000 token context window, returns up to 128,000 output tokens, and has a knowledge cutoff of 16 February 2026. Reasoning effort is exposed directly in the API and accepts none, low, medium (the default), high, xhigh and max. If you have seen it described as having a 1.1 million token window, that is a rounding by third-party trackers; OpenAI's own model page states 1,050,000.

Pricing

The standard rate is $5 per million input tokens and $30 per million output. On top of that sits promotional pricing of $4 and $20 — a 20% cut on input and 33% on output — which OpenAI say holds at least through 21 November 2026. Their own page records the history: a price drop on 21 August 2026 of over 20% for three months, following a 30 July reduction on Luna. Prompts over 272,000 input tokens bill at 2× input and 1.5× output for the entire request, not just the excess. Cache writes bill at a premium as well.

What OpenAI claim, and on what

These are vendor figures. OpenAI report state-of-the-art results on BrowseComp at 92.2% and OSWorld 2.0 at 62.6%, adding that on OSWorld it surpasses Claude Opus 4.8 while using 85% fewer output tokens. On Agents' Last Exam, an evaluation of long-running professional workflows across 55 fields, they put Sol at 53.6 — 13.1 points above Claude Fable 5 with adaptive reasoning, and still 11.4 points above it at medium reasoning for roughly a quarter of the estimated cost. Cybersecurity is where the generational jump is largest: ExploitBench 2 at 73.5% against GPT-5.5's 47.9% at a comparable token budget, SEC-Bench Pro at 71.2% against 45.8%, and ExploitGym rising from a 15.1% peak pass rate to 24.9% under a two-hour cap, or 33.7% given six hours. On the Artificial Analysis Intelligence Index, OpenAI say Sol at max reasoning comes within one point of Fable 5 while finishing in 61% less time at around half the cost. Rogo, a customer, reported a 6.2-point rubric-quality gain over GPT-5.5 on their finance benchmark.

The ultra setting

Above max sits ultra. Ultra coordinates four agents in parallel by default, trading token use for stronger results and faster wall-clock completion on demanding work. OpenAI publish comparisons of the four-agent default against a single-agent baseline on BrowseComp, SEC-Bench Pro and Terminal-Bench 2.1, with 16-agent configurations shown for the first two. An ultra score is a multi-agent system's score rather than a single model's. Developers can build equivalent setups through the multi-agent beta in the Responses API.

Independent figures

On LMArena's text leaderboard GPT 5.6 Sol has 1485 Elo (rank 23, xhigh effort); by category, 1533 Elo in Coding (rank 19, xhigh effort), 1494 Elo in Math (rank 23, xhigh effort), 1532 Elo in Expert (rank 12, xhigh effort) and 1471 Elo in Creative Writing (rank 19, xhigh effort). SimpleBench, run independently, records 65. Output runs at 70 tokens per second with a 97.2 second time to first token.

Official Resources