OpenAI finally discounts GPT-5.6 Sol, and puts an expiry date on it

OpenAI lowered the API price of GPT-5.6 Sol on 21 August, bringing its frontier model down to $4 per million input tokens and $20 per million output tokens. The company's API changelog describes the move as 20% lower input pricing and 33% lower output pricing, and attaches a condition that is easy to skim past: the rate is promotional and holds "at least through November 21, 2026".

What actually changed

Working backwards from OpenAI's own percentages, Sol's previous standard rate was $5 per million input tokens and $30 per million output tokens - the price it launched at on 9 July and held for six weeks, including on 30 July when the other two members of the family were repriced and Sol was not.

In its announcement on the OpenAI Developer Community forum, the company framed the change as dropping API and credit pricing "by over 20% for the next 3 months". It named where the reduction lands:

• API requests to gpt-5.6-sol
• Codex credits
• Eligible ChatGPT Work plans
• Not Pro, Plus or Business subscription usage, which OpenAI says is unchanged

The discount has a date on it

The two OpenAI statements phrase the duration differently, and the difference is worth noticing. The developer-forum post says the lower pricing runs "for the next 3 months". The changelog says the promotional pricing is available "at least through November 21, 2026". Three months from 21 August is 21 November, so the two agree on the date. They disagree on what happens afterwards: "at least through" leaves room for the rate to persist, while "for the next 3 months" reads as an end point. Neither commits to anything past that day.

That matters more than a footnote usually would, because this is the number people build spreadsheets on. If you are re-costing a Sol workload against $4 and $20, treat 21 November as a review date rather than a permanent floor. OpenAI has not said what the rate reverts to, or whether it reverts at all.

What it does to the model-selection maths

OpenAI's pricing page currently lists the three GPT-5.6 models, per million tokens, on the standard tier:

• GPT-5.6 Sol: $4 input, $20 output
• GPT-5.6 Terra: $2 input, $12 output
• GPT-5.6 Luna: $0.20 input, $1.20 output

Before today, Sol cost 2.5x Terra on input and 2.5x on output. It now costs 2x on input and roughly 1.7x on output. For output-heavy work - long generations, agent traces, verbose structured responses - the gap between OpenAI's frontier tier and its mid tier is meaningfully smaller than it was yesterday. That is the part worth re-testing rather than assuming: a workload that was routed to Terra purely on cost, and which took a quality hit for it, may no longer justify the downgrade.

The gap to Luna remains large. Sol's output price is still about seventeen times Luna's, so high-volume classification, extraction and routing work has not changed character.

The fine print

• The $4 and $20 figures are standard-tier pricing. OpenAI's pricing page separately lists a long-context tier for Sol at $8 input and $30 output; the changelog entry does not mention it, and the page does not state where the long-context threshold begins.
• Cached input for Sol is listed at $0.40 per million, a tenth of the uncached rate.
• Batch and flex are listed at 50% of standard.
• Fast mode - renamed from priority processing on 30 July - is listed at double the standard rate.

The caching rate is more actionable this week than last. On 20 August OpenAI shipped a Prompt Caching dashboard that reports cache hit rate over time, cache reads per write, and the breakdown between cache-read, cache-write and uncached tokens. If a large share of your Sol spend is repeated system prompts or tool definitions, that dashboard is what tells you how much of the $4 rate you are actually paying.

Six weeks, three repricings

OpenAI's changelog records the sequence plainly. The GPT-5.6 family shipped on 9 July - Sol for frontier capability, Terra for a balance of intelligence and cost, Luna for efficient high-volume workloads. On 30 July, Luna dropped 80% and Terra dropped 20%, with Sol left at its launch price. Today Sol follows. Every model in the family has now been repriced within six weeks of launch.

Worth being clear about what this is not. OpenAI has published no benchmark, capability claim or model update alongside the change, and the changelog offers no stated reason for it. Sol is the same model it was yesterday, sold cheaper for a defined window. That is a real and useful change for anyone paying the bill - and one with a date attached that belongs in the calendar, not the footnotes.

Separately, the same day's changelog adds that API customers can now select regional processing for an individual request by using a prefixed domain with an API key from a project set to Global geography, with existing eligibility, retention-control, endpoint and model-support requirements still applying.