GPT-6 Sol vs GPT-5.6 Sol

OpenAI launched GPT-6 Sol on September 22, 2026 and framed it as a straight price cut on the model it succeeds: GPT-5.6 Sol at $4 per million input tokens and $20 output becomes GPT-6 Sol at $2 and $10, half on both sides, with cached input at $0.20 against $0.40. An OpenAI spokesperson told VentureBeat the new rates are permanent, which matters because the GPT-5.6 Sol rate is itself a promotion that OpenAI lists as running at least through November 21, 2026. GPT-5.6 Sol is not deprecated and its price did not change. The catch is a step in the rate card. GPT-6 Sol bills any request above 272,000 input tokens at 2x input and cache and 1.5x output for the whole request, $4, $0.40 cached, and $15, which is level with GPT-5.6 Sol on input and still cheaper on output, but it breaks any cost model that assumes one rate at every length. The window grows to 1,050,000 tokens with 128K output, and reasoning effort runs from none through max with medium as the default. On capability the independent evidence is thin and points one way: Cognition's FrontierCode v1.1 Main scores GPT-6 Sol at 49.3 against 47.5 for GPT-5.6 Sol, the Vals Index places it eighth at 62.57 percent, and Artificial Analysis scores it 48. OpenAI's own claim is that it makes about half as many mistakes as its predecessor. What GPT-6 Sol lacks is the stored academic record. GPT-5.6 Sol carries SWE-bench Verified at 96.2 and MMLU-Pro at 89.1 from Vals AI, plus vendor GPQA Diamond at 94.6 and BrowseComp at 90.4, and OpenAI reported none of those for GPT-6 Sol. Vals archived SWE-bench Verified, MMLU-Pro, and GPQA Diamond as saturated on September 1, so those columns will not fill in for the new model. OpenAI does report 60.5 on an offline OSWorld 2.0 set at xhigh effort, which it likens to Claude Opus 5 at medium effort; we cannot confirm that setup matches the run behind GPT-5.6 Sol's stored 62.6, so treat that pair as unconfirmed.