OpenAI cut GPT-5.6 Sol to $4 per million input tokens and $20 per million output tokens on August 21, 2026, and says the rate holds at least through November 21. The headline framing was "over 20 percent," but the output side fell 33 percent, from $30. Separately, OpenRouter is listing the same model from the OpenAI provider at roughly half OpenAI's own promotional rate.
Key facts
- New pricing: $4 per million input tokens, $20 per million output, down from $5 and $30.
- The cut also covers Fast mode, long-context requests, and Batch and Flex processing, per OpenAI's announcement.
- The model card says the rate is available at least through November 21, 2026.
- Primary source: OpenAI's pricing announcement.
Price cuts are the least glamorous AI news and among the most consequential, because they change which products are buildable. A 33 percent reduction on output tokens is not a rounding adjustment; it is the difference between an agent workflow that pencils out and one that does not.
Here is why the two numbers differ. Language model billing is asymmetric: you pay one rate for the tokens you send and a higher rate for the tokens the model generates. Output costs more because generating is the expensive part -- each new token requires another full pass through the model, while input tokens are processed in parallel. Any workload that reads a little and writes a lot -- code generation, long reports, agent loops that reason at length before acting -- is dominated by the output rate. So a cut that reads as "over 20 percent" in the announcement is closer to a third off for exactly the applications people are building right now.
OpenAI's post is specific about the scope. "The price reductions also apply to Fast mode, long-context requests, and Batch and Flex processing," the announcement states, while making clear that paid subscription usage and limits are not changing. The older Sol preview post still carries the launch-era $5 and $30 figures, so the change is visible in the gap between the two pages.
Then there is the part that is not OpenAI's doing. OpenRouter, the routing marketplace, currently lists GPT-5.6 Sol from the OpenAI provider at roughly $2 in and $10 out, marked as a 50 percent discount, on its discounted models collection and on the Sol model page. The same page shows Azure at $5 and $30 and Amazon Bedrock at $5.50 and $33 for the identical model. This has circulated as evidence that OpenAI quietly cut twice. It did not. What happened is a reseller stacking its own promotion on top of OpenAI's promotional rate -- which is arguably more interesting, because it means a routing layer is now setting the effective market price of a frontier model.
The analogy is airline fares. The carrier publishes a price; the consolidator sells the same seat for less to win the booking; the passenger flies the same plane. What is new is that in AI, the consolidator layer is only a couple of years old and is already strong enough to reprice the flagship product of the largest lab.
Why it matters: the floor keeps dropping, and not because any single company decided to be generous. Anthropic's public pages still list Fable 5 at $10 and $50 and Opus 5 at $5 and $25 per million tokens. DeepSeek's pricing page sits far below all of it, with its Flash tier in the low cents per million depending on peak and off-peak hours. Open weights apply pressure from a third direction -- when self-hosting a frontier open model takes a whole node but a 22 GB local coder can hold its own on a real repair benchmark, the hosted price has to answer for itself. Ground Truth has an explainer on why LLM inference is memory-bound, which is most of the reason output tokens cost what they do, and a new one on inference cost and token economics.
The reaction in OpenAI's own forum was telling. Developers accepted the cut as real and immediately moved on to complaining about something else: the churn in pricing pages and calculators, the difficulty of forecasting spend when rates move on promotional windows, and the fact that subscription value did not budge while API costs fell. That is what a maturing market sounds like -- the price is no longer the story, the predictability is.
The honest caveat is right there in the phrase "at least through November 21, 2026." This is a promotional rate with a stated floor date and no stated ceiling. Building a business model on a promotional price is a choice, and the OpenRouter discount is even less durable than OpenAI's. Anyone doing capacity planning should model both the promo rate and the rate it replaced.
Originally published on Ground Truth, where every claim is checked against the primary source.
