OpenAI cut the developer price of GPT-5.6 Sol, its frontier model, by more than 20% on Friday. Standard short-context rates fall from $5 to $4 per million input tokens and from $30 to $20 per million output tokens, according to OpenAI’s own pricing table, with the change also rolling out across eligible credit plans for its coding tool Codex and its agentic product ChatGPT Work.
Two details do most of the work here. First, the headline “more than 20%” is an average that hides the real move: input came down 20%, output came down 33%. Second, the new rate is temporary. OpenAI framed it as applying for the next three months, which puts the reversion date around 21 November.
Subscription pricing did not change. Pro, Plus and Business plans stay where they were. This is aimed squarely at the people who buy tokens by the million, not at the people who pay a flat monthly fee.
Our take: a 33% cut on output, and only on output, is a targeted strike at agentic and coding workloads — the jobs that read a little and write a lot. Those are exactly the workloads Anthropic has been winning, and exactly the ones where a per-token spread compounds into a real monthly invoice. This is not generosity. It is a 90-day rental on price leadership.
Where it lands against the competition
Anthropic lists Claude Fable 5 at $10 per million input tokens and $50 per million output, and Claude Opus 5 at $5 and $25. At $4 and $20, Sol now sits under both. That is the point: OpenAI has not just trimmed a number, it has moved beneath the nearest frontier rival on both sides of the meter.
It is also the second pricing move on the 5.6 family inside a month. In late July OpenAI cut mid-tier Terra by 20% and lower-cost Luna by 80%. Read together, the sequence is a lab walking prices down the entire ladder in about four weeks — the cheap end first, now the top. The company’s stated reason is efficiency gains as capability improves. The unstated one, per Reuters, is pressure from Anthropic and from Chinese models that have turned frontier inference into a commodity fight faster than anyone expected.
The trap in a temporary price
Three months is long enough to change behaviour and short enough to leave you exposed. If you re-architect an agent loop around $20 output — more reasoning passes, longer generations, less aggressive caching — you have built a cost structure that only clears at promotional rates. Come late November, either the promo extends or your unit economics move 50% against you on output, overnight.
The teams that do well out of this will treat it as a windfall rather than a baseline: bank the saving, keep the routing layer that lets you swap models, and price your own product off the old numbers. We covered AT&T cutting coding-AI costs 56% by routing across models for exactly this reason — the durable saving came from optionality, not from any one vendor’s discount.
What to watch
- Whether Anthropic responds. Opus 5 at $5/$25 is now the pricier frontier option on both sides. A matching move would confirm the commodity read.
- The November date. If the rate quietly becomes permanent, OpenAI’s inference costs genuinely fell. If it snaps back, it was a retention promo.
- Cached and long-context rates. The cut is quoted for standard short-context use. Long-context and cached-input pricing is where agentic bills actually live.
- Codex and ChatGPT Work credit burn. Cheaper credits mean the same plan buys more work. Watch whether that lands as higher usage rather than lower revenue.
So what: frontier intelligence just got a third cheaper to generate, for ninety days. Use it. Do not build on it.