OpenAI Put the Frontier Tier on Sale for 92 Days
Sol output tokens fell by a third on 21 August. The guarantee runs to 21 November. That is a promotion, not a price — and it is a different planning object.
Sol output tokens fell by a third on 21 August. The guarantee runs to 21 November. That is a promotion, not a price — and it is a different planning object. What happened On 21 August 2026, OpenAI cut the API price of GPT-5.6 Sol — its flagship tier. OpenAI's own pricing table puts input at $4.00 per million tokens, down from $5.00 (20%) , cached input at $0.40, down from $0.50 (20%) , and output at $20.00, down from $30.00 — a 33% cut . The cut is time-boxed. OpenAI's announcement guarantees it "at least through November 21, 2026" — 92 days — and explicitly leaves open the possibility of continuing past that date. It does not commit either way. Scope, from OpenAI's announcement: API calls, Fast mode, long-context requests, Batch and Flex processing, plus Codex credits and eligible ChatGPT Work plans. Pro, Plus and Business subscription usage is unchanged. This is the second GPT-5.6 cut in 22 days . The 30 July round hit the lower tiers — Terra 20%, Luna 80%. We published a routing playbook for that round on 7 August , with thresholds calibrated to Terra and Luna pricing: The GPT-5.6 Price Cut Routing Playbook. Those thresholds are now too conservative — the Sol cut moves the line we drew three weeks ago, and we say so below. On the why , be careful. Reuters frames the cut as a response to "growing competition from Anthropic and Chinese AI models." That is Reuters' reading, not OpenAI's statement. OpenAI's own framing is narrower: "continue to push the frontier of capabilities while improving efficiency." Neither is independently verified. Why it matters The headline number in most coverage is "more than 20%," which undersells the part that changes decisions. Input fell 20%; output fell 33%. Sol earns its premium on reasoning-heavy, long-generation, agentic work — exactly the workloads that are output-dominated. The cut lands hardest where the money actually goes. The cross-tier ratio moved with it. On post-cut pricing, Terra sits at $2.00 input / $12.00 output. Sol is still 2.0× Terra on input, but on output the multiple fell from 2.0× to 1.67× (our arithmetic on the published figures). If your router was built on "Sol costs twice as much to generate," that premise is stale. The second thing that matters is the shape of the thing. A price cut is a new floor you can plan against. A limited-period cut is an option OpenAI holds and you do not. If you re-architect routing around $20 output and it reverts to $30 on 22 November, your unit economics move 50% against you on your most expensive traffic, without a negotiation and without warning. That is not a reason to ignore the promotion. It is a reason to treat it as a promotion in every model you build with it. For a CPO, the near-term consequence is a re-scored backlog: features shelved as uneconomic at $30 per million output tokens deserve a second look this week, with a note on the ticket that the price is guaranteed only to 21 November. For a founder, it is 92 days of subsidised experimentation o