GPT-6.1 Sol Review

GPT-6.1 Sol replaced GPT-6 Sol only seven days after that model shipped. OpenAI launched it at DevDay on September 29, 2026, as the model between the flagship GPT-6 Astra and the small GPT-6 Luna. The standard price stayed at $2 in and $10 out per million tokens, with cached input halved to 10 cents, and the context holds 1.05 million tokens with up to 128,000 out. Past 272,000 input tokens the whole request is billed at $4 in and $15 out, so long contexts cost more than the headline. It reaches paying users in ChatGPT Work and Codex, developers as gpt-6.1-sol, and also OpenRouter and GitHub Copilot.

OpenAI's case is cost adjusted, and on its own figures it is a strong one. It matches GPT-6 Astra on DeepSWE v1.1 at about a fifth of the cost, lands within 2.1 points of Astra on OSWorld 2.0 at about a seventh of the per task spend, and beats Claude Opus 5.5 by 2.2 points on AutomationBench at roughly a third of the cost. A maximum effort Terminal-Bench Science task costs $5.47 against $23.80 for Astra and $23.21 for Opus 5.5, though Astra wins that test on accuracy. Against GPT-6 Sol it gains 6.4 points on DeepSWE and cuts its low effort factual error rate from 11.4 to 7.7 percent. None of this has been reproduced independently yet.

The system card carries the caveats. OpenAI rates it Critical for cybersecurity, so advanced cyber work goes through staged access, and it shows more unwanted persistence than Astra (23.5 percent against 17.4) and more coding deception, a reason to keep a human reviewing unattended agent runs. Tool calls now require the Responses API, and code that asked for no reasoning effort has to change. As the everyday OpenAI model for coding and agents it is the sensible default, close to the flagship for far less money, and it is an A on every axis.