GPT-6.1 Sol: API Specs, Reasoning Controls, and Costs

GPT-6.1 Sol is OpenAI’s updated Sol-tier model for complex coding, computer use, and professional workflows. OpenAI positions it as a balance between capability and cost in the GPT-6 family. For an existing Sol deployment, the useful question is whether it completes the application’s work with fewer failures or less repair effort.

OpenAI lists a 1,050,000-token context window and a 128,000-token output maximum. Supported reasoning efforts are low, medium, high, xhigh, and max; medium is the default. None and minimal are unsupported. Use the Responses API for tool calling; Chat Completions supports this model without tools. Official model documentation

A migration needs to check both the model ID and the surrounding request protocol. Code that already uses Responses can evaluate the new model behind the same application interface. Code that expects Chat Completions function-call messages needs to account for the documented tool contract before changing the selected model. The API endpoint is part of the integration, along with the name in the model selector.

For an application evaluation, keep tool definitions, task permissions, and source context consistent across runs. Track whether the model calls an allowed tool with valid arguments and whether the eventual result passes the application’s checks. A well-written answer is only one part of success when the system can operate on files or services.

At the base $2/M input and $10/M output rates, 100,000 input tokens plus 10,000 output tokens cost $0.30. That example excludes caching, long-context pricing, tools, and retries. Reasoning tokens contribute to billed output, so a deeper reasoning setting can increase cost even when the visible answer stays short. OpenAI reasoning documentation

GPT-6.1 Sol was a paid route when checked, so this pass did not send the SVG, globe, or wizard-game prompts to it. An appropriate next comparison would rerun a fixed set of real coding tasks, record all attempts, and use correctness tests alongside latency and token totals. No new hands-on score or claimed speed improvement is assigned here.