OpenAI cuts GPT-5.6 Sol 20%: the discount has an end date

OpenAI cut GPT-5.6 Sol API pricing by more than 20% on August 21. Input went from $5.00 to $4.00 per million tokens, output from $30.00 to $20.00, cached input from $0.50 to $0.40. That is a real reduction on a frontier model, and it lands the same week Anthropic and several Chinese labs kept squeezing the same segment. It is also a promotion with a calendar date attached, and the OpenAI price cut is only useful to you if you plan for November 21.

What actually happened

OpenAI's announcement states the new rates are "available at least through November 21, 2026" — three months from the August 21 start. The cut applies to the API and is rolling out across eligible plans for Codex credits and ChatGPT Work. Pro, Plus, and Business subscription prices do not change, per Reuters' reporting.

Read the two numbers separately, because they did not move together. Input fell 20%. Output fell 33%. That asymmetry is the actual news. Workloads that read a lot and write a little — classification, routing, extraction, retrieval scoring — get the smaller discount. Workloads that generate — coding agents, long report drafts, multi-step tool loops that narrate their reasoning — get the bigger one.

Run it on a real month. Take 10M input and 2M output tokens. Before: $50 + $60 = $110. After: $40 + $40 = $80. That is 27% off, better than the headline, because output dominates the bill. Flip the ratio to 40M input and 1M output and you save 20.5%. Your discount depends on your token mix, not on the press release.

Why a temporary AI price cut matters for your business

Do not rebuild your unit economics on a three-month rate. If you are quoting a customer a per-ticket or per-document price today, quote it at $5/$30 and pocket the difference until November. The alternative is explaining a 25% cost increase to a client in Q4 because a vendor promo lapsed.

Instrument the split now. You need input and output tokens tracked separately, per feature, before the reversion — otherwise you will not know which parts of your product got expensive again. If your logging only records total spend, you are flying blind on exactly the number that changed most.

Cached input at $0.40 is the line worth optimizing. It is one-tenth of the cache-miss rate. Stable system prompts, fixed tool schemas, and a consistent document prefix are the cheapest performance work available. That ratio is not promotional in the same way the headline is — cache economics have held across every provider for a year.

Treat the discount as evidence, not as a gift. Frontier inference pricing is being fought over quarterly now. The correct response is not loyalty to whoever is cheapest this month. It is an abstraction layer thin enough that swapping models is a config change and a re-run of your evals.

Sources: OpenAI Developer Community announcement, Reuters via Business Standard.