OpenAI cuts GPT-5.6 Sol API prices by over 20%: what to recalculate
Original publication: · Sergey Zhadov
This is a translated archive article. The date and pricing discussion refer to the original publication, not a new announcement today.
OpenAI lowered GPT-5.6 Sol API and credit prices by more than 20% for three months. That is a reason to recalculate a working workflow, not to blindly move an entire product to the flagship model.
The price is lower, but only temporarily
OpenAI's GPT-5.6 announcement carries an update dated 21 August: Sol API and credit pricing was reduced by over 20% for the following three months. It does not mean every ChatGPT subscription automatically becomes cheaper.
The announcement does not promise that the offer will stay after that window. A pilot therefore needs a second budget, based on what happens when the temporary terms end.
Who the change affects
The change directly concerns developers, agent platforms and teams paying for GPT-5.6 Sol calls or consuming credits. Multistep agents are especially exposed: one user task can trigger several model calls, tools and verification steps. On a fixed-price subscription, check your own plan separately.
- Input
Count system instructions, user data and retrieved context, not just the last question.
- Output
Long answers and intermediate reasoning can cost much more than the input.
- Retries
Failed calls, checks and regenerated answers all belong in the cost of the finished result.
Recalculate the budget without guessing
The official page lists $5 per million input tokens and $30 per million output tokens for Sol. For a first estimate, multiply actual input and output volumes by their respective rates. Then include tools, retries and application checks. Do not count the same retry tokens twice: they may already be included in your usage export.
Suppose an unchanged workload previously cost RUB 100,000. A cut of more than 20% would put it below RUB 80,000 only if the consumption mix stayed the same. This is a hypothetical calculation, not a forecast. Taxes, exchange rates, platform markups and longer responses can change the total. Compare real invoices for the same week's workload, not the pretty percentage in the headline.
- Week one
Freeze a baseline task set and record the current cost per successful result.
- Month one
Review quality, latency, repeat calls and the actual invoice together.
- Before the offer ends
Prepare a post-offer budget and a route back to a cheaper model.
Three months is a pilot, not a one-way migration
Three months is enough for a pilot and too short for a migration with no way back. Keep the model and rates in configuration, set a spending cap and retain routing to Terra or Luna where appropriate. Check live billing again before the final billing period. An August screenshot will not be a current price reference forever.
Run a fair production test
Collect 30-50 representative tasks: a short answer, a long analysis, file work, a tool call and a difficult case the model often gets wrong. Run them on Sol and a cheaper model under the same rules. If possible, keep the evaluator blind to which model produced each answer.
Measure more than the price of a call: success rate, manual corrections, latency, total token use and cost per accepted result. A model that costs more per request can still win if it sends far less work back for repairs. If quality is the same, the flagship badge does not pay the difference.
- Route tasks
Send straightforward work to a cheaper model and reserve Sol for difficult cases.
- Set limits
Cap context size, retry counts and the maximum output per task.
- Raise alerts
Warn the team before a day's experiment eats the whole monthly budget.
The limitations you should not skip
A lower rate does not guarantee every mode is available or that every product uses the same billing rules. The API, credits, ChatGPT, Codex and third-party platforms may calculate costs differently. Caching has its own rules too: the official material describes a premium for cache writes and a discount for reading cached input.
Do not confuse a published token rate with the full cost of a feature. Data storage, document search, code execution, monitoring, support and human review still count. If spending is not linked to individual workflows, the team sees a big bill without knowing which feature is gobbling up the money.
- Record
Log the model version, effort settings, date, input and expected result.
- Measure
Track the cost of an accepted result, not the cost of one lucky answer.
- Decide
Keep Sol where the quality improvement justifies the full expense.
What the team should do now
Export 7-14 days of Sol spending, split it by feature and set two budgets: one for the discount window and one for afterwards. Compare models on a fixed sample. Short classification and field extraction often do not need Sol; complex code, research and multistep checks may justify it. Test that assumption against your own quality threshold.
KGAM's take: measure first, enjoy the discount second
The price cut makes Sol more interesting to test. Treat the three months as a controlled experiment: a fixed sample, a budget, a stop rule and a cost per accepted result. Without measurement, cheaper tokens may just let you burn through more of them.
Translation reviewed:
Primary source: OpenAI · GPT-5.6
