gpt-6.1-sol

Overview. GPT-6.1 Sol is OpenAI's closed-weight reasoning model, released on 29 September 2026 as a refresh of GPT-6 Sol and positioned between it and the GPT-6 Astra flagship in the GPT-6 line. It accepts text and image input and returns text, working across a 1,050,000-token context window with up to 128,000 tokens per response. A developer reaches for it when a task needs near-flagship reasoning over very large amounts of material, and is willing to accept a reasoning pass on every request in exchange.

Key strengths. Measured on the xhigh effort configuration, GPT-6.1 Sol records an AA Intelligence Index of 51, MMLU-Pro 0.85, SciCode 0.56, AA-Omniscience accuracy 0.61 and a GDPval-AA Elo of 1510. Long-context work is where it stands out: AA-LCR 0.80 against a 1,050,000-token window, of which up to 922,000 tokens can be prompt input. Tool use is broad and provider-hosted, covering web_search, file_search, code_interpreter, computer_use, hosted_shell, apply_patch, skills, mcp and tool_search, and reasoning effort is configurable across low, medium, high, xhigh and max. Prompt caching is automatic, so repeated long prefixes are cached without a client-side cache primitive.

When to use it. Use GPT-6.1 Sol for research and document-heavy tasks that must reason over hundreds of thousands of tokens, where the AA-LCR result and native file_search and web_search tools do the retrieval work in one pass. It suits agentic pipelines that need multi-step recovery, since reasoning is always on and the hosted tools handle code execution, browser-style operation and external tool servers. Because there is no non-reasoning mode and median output speed is around 63 tokens per second, it fits interactive request-and-response work rather than latency-critical, high-throughput paths.

Limitations. Reasoning is mandatory: effort accepts low through max, but none and minimal are unsupported, so there is no fast no-think path, and temperature and top_p are rejected outright, leaving no sampling control. Function calling is available only through the Responses API, so the same model served over Chat Completions cannot call tools. Fine-tuning is not supported, the knowledge cutoff is 30 April 2026, and output is text only, with no audio input or speech output.

Sovereignty note. GPT-6.1 Sol is built by OpenAI, a US-based creator, so the first-party route carries CLOUD Act exposure. Standard processing supports both US and EU data residency, though Fast mode does not; OpenAI states that data sent to its API is not used to train or improve its models unless the customer explicitly opts in, and offers Zero Data Retention and Modified Abuse Monitoring to approved customers, with abuse-monitoring logs otherwise retained for up to 30 days. EU data residency additionally requires approval for abuse-monitoring controls. On the available evidence it is not a Sovereign option.