The most expensive model on the board just got cheaper. Only until 21st Nov’26
Since 21st Aug’26, OpenAI has been charging less for GPT-5.6 Sol, its top model for complex coding and agent work. Short-context input drops from $5 to $4 per million tokens; output drops from $30 to $20, that’s a fifth off the way in, a third off the way out. It applies to the pay-as-you-go API, Codex credits and eligible ChatGPT Work plans; consumer Pro, Plus and Business subscriptions are unchanged. Amazon announced matching reduced pricing on Bedrock the same week. The rate is guaranteed until 21st Nov’26, and OpenAI has not said what happens after.
What it actually means
Two weeks ago the cheapest supplier on the board put its prices up. This week the most expensive one put its prices down. Those aren't two stories, they're one and the story isn't the level, it's the volatility. A price with a published expiry date is not a price. It's a customer-acquisition budget with a quarter attached to it, and 21st Nov’26 is not an accident, it lands just before Q4 renewals. That is the mechanism, and it works because architectures set faster than budgets do. So take the discount, and be specific about what you take it for. A 90-day window is not a licence to re-plan your unit economics around $20 output. It's a window to do the things you couldn't previously justify: run the full eval suite you've been sampling, clear the backfill, and most usefully, put a cheaper open-weight model through your actual traffic. The only real hedge against a market where the price is a marketing lever is the ability to switch suppliers without a rewrite, which means the cost worth paying down this quarter isn't the token bill, it's the coupling.
A price with an expiry date isn't a price. It's a customer-acquisition budget.
21 Novthe day the discount expires
Source: WinBuzzer / AWS / Quartz · Monday 24 August 2026