OpenAI cut GPT-5.6 Sol pricing on August 21. The current standard API rates are $4 per million uncached input tokens and $20 per million output tokens. The saving on a complete workload depends on its input and output mix.
Read the current rates and their conditions
The official model page lists $4 for input, $0.40 for cached input and $20 for output per million tokens. Its promotional wording says at least through November 21, 2026. That does not announce an automatic return to previous rates on November 22.
The August 21 announcement describes a 20% input reduction and a 33% output reduction, with lower pricing also applied to other processing modes. It covers API usage and eligible Work/Codex credit plans; Pro, Plus and Business subscription usage remains unchanged.
An example with actual request sizes
Suppose an application makes 100 separate requests, each with 10,000 uncached input tokens and 2,000 billed output tokens. The aggregate is one million input tokens and 200,000 output tokens. None of these individual requests crosses the long-context threshold.
At the previous $5/$30 standard rates, input costs $5 and output costs $6, for $11 total. At $4/$20, the same counts cost $4 + $4 = $8. The saving is $3 ÷ $11, or about 27.3%.
This original example is a token-cost calculation, not a measured production bill or a prediction of output quality. It assumes unchanged token counts and excludes separate tool, storage and other service charges. If a revised workflow produces more billable output, a lower unit price need not lower its total bill by the same percentage.
Keep long context and cache billing separate
The model page specifies that prompts above 272,000 input tokens use 2× input and 1.5× output pricing for the whole request. It also lists cache writes at 1.25× the uncached input rate. A single million-token prompt therefore cannot be substituted into the short-request example while retaining its rates.
Cached reads and cache writes are different billing events. The presence of repeated prompt text does not justify marking every input token as a discounted cache read. Use the usage categories returned for the actual requests and the rules for the selected processing mode.
Reprice a workload before changing it
For a useful comparison, keep a dated record of request counts, input categories, billed output and processing mode. Calculate the same workload under both schedules first. Changes in model behavior, retries or application design can then be evaluated separately from the price change.
November 21 is a sensible date to recheck the offer's continuation. The source does not establish a compulsory expiry or a subsequent price, so a future budget should label any assumed post-promotion rate as a scenario.
Recheck official standard and long-context prices; fix the invented hard expiry and calculate savings across separate short requests.