OpenAI launches GPT-6.1 Sol with Astra-level agent coding at one-fifth the price

OpenAI released GPT-6.1 Sol this week, an update to the mid-tier model in the GPT-6 family, claiming it delivers results close to the flagship Astra on agent coding, computer use and professional work at one-fifth of Astra's token pricing. The new rates are $2 per million input tokens, $10 per million output tokens, and 10 cents (about 37 agorot) per million cached input tokens. By comparison, Astra charges $10, $50 and $1 respectively, while the smaller Luna model is priced at 10 cents, 50 cents and one cent.

The cached-input discount is especially significant for agents, which resend the same system prompt, tool schemas and history at every step. In the new version a cache read costs 5% of the regular input price, down from 10% in the previous GPT-6 Sol. The model is available now via the API under the identifier gpt-6.1-sol, in ChatGPT Work and in Codex, with a 1,050,000-token context window, a 128,000-token output limit and a knowledge cutoff of April 30, 2026.

According to OpenAI's published figures, on DeepSWE v1.1 Sol matches Astra at one-fifth the cost and improves the best GPT-6 Sol score by 6.4 percentage points at lower reasoning effort. On GDP.pdf, which tests answers on complex professional PDF documents, Sol beats Claude Opus 5.5 with fallbacks at less than half the cost per task. On AutomationBench 1.0.6 it leads Opus 5.5 by 2.2 points at medium effort, at roughly one-third the cost. On OSWorld 2.0 offline the gap to Astra narrows to 2.1 points at max effort, at about one-seventh the cost per task.

Science and factuality still have room to improve. On Terminal-Bench Science 0.1 Sol more than doubles the GPT-6 Sol score at max effort, with an average cost of $5.47 per task versus $23.21 for Opus 5.5 and $23.80 for Astra, yet Astra still leads at 68.1%, and OpenAI itself recommends Astra for the hardest research. On the factuality benchmark at low effort, the rate of answers containing a factual error fell from 11.4% to 7.7%, a reduction of about 32%, measured on especially difficult conversations that users had flagged as problematic.

API details and limits worth knowing: the API accepts text and image input and produces text output only, with five reasoning-effort levels — low, medium (default), high, xhigh and max — while none and minimal are not supported. Tool calls require the Responses API; Chat Completions works without them. Prompts exceeding 272,000 tokens double the input and cache rates and multiply the output rate by 1.5 for the entire request. Batch and Flex modes are 50% cheaper, Fast mode costs twice as much, and data residency in the US and EU is supported, though Fast is not available with EU residency. Fine-tuning is not supported, and an Ultrafast version of Codex with up to 8x token-generation speed is expected in the coming days.

The competitor comparison is not always apples-to-apples. In the official price table Claude Sonnet 5.5 matches Sol at $2 and $10, but Sol's cached input is half the price. OpenAI's benchmarks compare against the more expensive Opus 5.5, not Sonnet. Gemini 3.1 Pro matches the input price but charges $12 for output and remains in preview. Anthropic notes that its new tokenizer produces roughly 30% more tokens for the same text, so list-price comparison is not a true cost comparison.