GPT-6 Sol

GPT-6 Sol arrived on September 22, 2026 as gpt-6-sol, the mid-tier of the GPT-6 generation, trained with similar methods to GPT-6 Astra and sold at half the price of the model it replaces. The rate is $2 per million input tokens, $0.20 cached, and $10 output, against $4 and $20 for GPT-5.6 Sol, and an OpenAI spokesperson confirmed the GPT-6 rates are permanent rather than promotional, which matters because GPT-5.6 Sol's $4 and $20 is itself a promotion that runs at least through November 21. The long-context rule carries over from Astra: prompts above 272,000 input tokens bill the whole request at 2x input and cache and 1.5x output, so $4, $0.40, and $15. Batch is half price and fast mode doubles the standard rate. The window is 1,050,000 tokens with 128K output, text and image input, and reasoning effort from none through max with medium as the default. OpenAI calls Astra its best model across the board and pitches Sol at complex coding and agentic workflows. Independent results put it clearly below the frontier and clearly above its price: Cognition measures 49.3 percent on FrontierCode v1.1 at max effort, ahead of GPT-5.6 Sol at 47.5, the Vals Index places it eighth at 62.57 percent, and Artificial Analysis scores it 48. OpenAI's own chart puts it at 60.5 percent on the offline OSWorld 2.0 set at xhigh effort, level with Claude Opus 5 at medium. It lands at exactly the $2 and $10 of Claude Sonnet 5, which makes that the comparison most buyers will actually run. OpenAI did not publish SWE-bench Verified, MMLU-Pro, GPQA Diamond, BrowseComp, or HLE figures for it.