On August 21, 2026, OpenAI used its API changelog to announce a price cut for its flagship model, GPT-5.6 Sol. The new rate is $4.00 input and $20.00 output per million tokens, which the company describes as 20% lower input pricing and 33% lower output pricing1.
The same entry states that this is promotional pricing available at least through November 21, 20261. It is presented as a time-limited rate rather than a permanent revision.
What the pricing page shows today
The changelog names figures only for standard pricing at shorter context lengths, but the pricing page shows the current rates for the other conditions. Under standard processing, GPT-5.6 Sol costs $4.00 input and $20.00 output for contexts at or below 272K tokens, and $8.00 input and $30.00 output for longer contexts12. Cached input runs $0.40 and $0.80 respectively, with cache writes at $5.00 and $10.002.
Choosing batch processing, where immediacy is not required, halves the standard rate at shorter contexts: $2.00 input and $10.00 output2. The context window itself — the unit these price tiers are drawn around — is covered in our explainer on context windows.
Whether the long-context and batch rates changed at the same time cannot be determined from public information, since the changelog does not say. The figures above are simply the current prices.
All three tiers in the family have now been repriced
This cut did not happen in isolation. Going back through the same changelog, GPT-5.6 Luna dropped 80% and Terra dropped 20% on July 30, 20261. Within roughly three weeks, all three tiers of the GPT-5.6 family have seen a price reduction.
Lining up the current standard rates at shorter contexts: Sol at $4.00 input and $20.00 output, the mid-range Terra at $2.00 and $12.00, and the entry-level Luna at $0.20 and $1.202. The gap between Sol and Luna is 20x on input and roughly 17x on output. The fact that Luna became available without usage limits on ChatGPT’s free tier rests on that same low price point.
Prices are falling for the processing itself, however, while speed is sold separately. On July 30, Fast mode replaced the Priority Processing offering, delivering up to 2.5x the speed of standard processing for GPT-5.6 Sol at twice the price1. Then on August 13, Ultrafast mode, said to run up to 14x faster than standard processing, was announced as a limited preview for select customers1. We covered Ultrafast mode separately.
Per-token rates are coming down while the tiers charging for speed multiply. Both together determine what a workload actually costs.
What matters when planning a budget
If you are calling Sol through the API, costs fall by 20% to 33% without changing a line of implementation. Because the output cut is the larger of the two, workloads heavy on output tokens — long answers, code generation — stand to benefit more.
The expiry date deserves attention, though. What OpenAI committed to is “at least through November 21,” with nothing said about what follows1. For budgets spanning a fiscal year or multi-year estimates, it is worth not treating this rate as permanent.
The changelog also offers no explanation for the cut. It is tempting to read competitive pricing or demand into it, but the only material made public is the price itself and the date. For data on which models companies actually pay for, see our separate article.
Sources
- Changelog - OpenAI API - OpenAI official API documentation (entries for August 21, August 13 and July 30, 2026)
- Pricing - OpenAI API - OpenAI official pricing page (confirmed August 24, 2026)