OpenAI: GPT-6.1 Sol

GPT-6.1 Sol is OpenAI's September 2026 upgrade to GPT-6 Sol in the GPT-6 series, positioned below flagship GPT-6 Astra while targeting near-Astra quality at much lower cost. It is built for agentic software engineering, computer use, and document-heavy professional work, with native text and image inputs, configurable reasoning effort, and a very large context window suited to long codebases and multi-step workflows. The model ships in ChatGPT Work, Codex, and the OpenAI API as gpt-6.1-sol, with standard pricing at $2 per million input tokens and $10 per million output tokens.

On capability evaluations, GPT-6.1 Sol reaches strong scores on graduate science and knowledge suites (including GPQA Diamond and MMLU-Pro), contamination-resistant LiveBench, and multimodal MMMU-Pro at high reasoning effort. For real-world engineering, OpenAI reports 75.2% on DeepSWE v1.1 at high effort, while third-party leaderboards place it competitively on SWE-Bench Verified and terminal-style agent benchmarks. It also improves over GPT-6 Sol on business workflow automation (AutomationBench) and desktop computer-use tasks such as OSWorld 2.0 offline, closing much of the gap to Astra at a fraction of the cost per task.

Safety and deployment follow the GPT-6 Astra safeguards stack under OpenAI's Preparedness Framework, with elevated ratings in cybersecurity and biological/chemical risk domains. GPT-6.1 Sol supports API tools including code interpreter, apply_patch, computer use, MCP, and related agent features, making it a practical default for developers who want Astra-class reliability on coding and agents without paying flagship inference prices.