OpenAI cut GPT-5.6 Sol pricing by over 20 percent for three months

The one OpenAI model that survived July’s price cut untouched is now the one getting cheaper.

What changed

On August 21, 2026 OpenAI dropped API and credit pricing for GPT-5.6 Sol by over 20 percent for the next three months. That phrasing is OpenAI’s own.

Pricing trackers and reporting on the change put the new rates at 4 dollars per million input tokens, down from 5, and 20 dollars per million output tokens, down from 30, guaranteed through November 21. ChatGPT Work and Codex credits get the same treatment on eligible plans. Pro, Plus and Business subscription usage is unchanged.

The contrast with the previous round is the point. On July 30 OpenAI cut Luna by 80 percent and Terra by 20 percent, and left flagship Sol exactly where it was.

Why the output side matters most

Agents generate far more tokens than they read, so a cut on output moves the unit economics of anything running long autonomous loops in a way that a cut on input does not. Sol is the model teams reach for when a task is hard enough that a cheaper tier keeps failing, which is also where agent loops burn the most tokens retrying.

The surrounding market is competing on price rather than capability. Google shipped Gemini 3.7 Flash this month at half the previous Flash tier’s cost, Alibaba published open weights for Qwen3.8-Max under Apache 2.0, and DeepSeek still undercuts Western pricing even after raising its own rates.

The date to keep

The guarantee ends on November 21. Anything rebuilt around these numbers, including agent designs that only pencil out at 20 dollars per million output tokens, needs a plan for what the rates look like after that.