Guide
GPT-5.6 Sol's Token Discount Did Not Expand Subscriptions
GPT-5.6 Sol API and eligible purchased-credit rates fell temporarily, while included ChatGPT plan usage stayed unchanged. Separate batch work from seat work.
OpenAI changed the price of GPT-5.6 Sol usage, not the amount included in ChatGPT subscriptions. That distinction matters for a freelancer or small team deciding whether an unattended workload belongs in ChatGPT, purchased credits, or an API project.
OpenAI’s GPT-5.6 product page was updated on 21 August 2026 to record a reduction of more than 20% in Sol API and credit pricing for the next three months. OpenAI’s Business and Enterprise credit rate card states that the promotion applies to eligible usage paid with purchased credits, while included plan usage and five-hour and weekly limits remain unchanged. For personal plans, OpenAI’s credit guide likewise says the plan and its included usage do not change; credits are a pay-as-you-go add-on for supported features.
This was a pricing update, not a new model release or a larger subscription allowance.
The current API meter
The OpenAI API pricing page, checked on 28 August 2026, showed these short-context rates per one million tokens:
| Processing mode | Input | Cached input | Output |
|---|---|---|---|
| Standard | $4.00 | $0.40 | $20.00 |
| Batch | $2.00 | $0.20 | $10.00 |
| Flex | $2.00 | $0.20 | $10.00 |
| Fast mode | $8.00 | $0.80 | $40.00 |
The Sol model page confirms the $4 input and $20 output Standard rates, a 20% input reduction and 33% output reduction. Requests with more than 272,000 input tokens price the full request at twice the input rate and 1.5 times the output rate. Cache writes have their own rate; do not substitute the cached-input number for them.
The pricing page—not a forum recap—also states that promotional pricing is available at least through 21 November 2026. “At least through” is a floor, not a permanent list price and not a promise that the previous price will return on the next day. Reopen the pricing page before quoting a client or approving a run.
Included plan usage did not grow
OpenAI’s official rate cards are explicit for managed plans: included plan usage, five-hour limits, weekly limits, and legacy credit rates did not change with the Sol promotion. The personal-plan guide also preserves included usage, and OpenAI’s GPT-5.6 in ChatGPT guide says the model uses existing ChatGPT limits. Eligible purchased-credit or token-billed activity receives the lower meter; the subscription’s included envelope does not become 20% larger.
That prevents a common comparison error:
- A seat buys a product surface and an included usage envelope.
- Purchased credits extend supported eligible features after included usage, where the plan offers them.
- The API bills a project for measured model and tool usage.
Those products can use the same named model without providing the same interface, storage, connectors, controls, or billing behavior. A lower API token rate does not make ChatGPT memory, files, or the interactive Codex loop free to rebuild.
Keep interactive work on the seat
Keep work in ChatGPT or Codex when its value comes from a person steering the session: exploring a problem, reviewing files, using connectors, refining a draft, or interacting with a coding agent. The subscription is paying for that working surface as well as included model usage.
Consider API or eligible purchased-credit billing for work that is:
- unattended and repeatable;
- measurable in input and output tokens;
- safe to retry and log;
- independent of ChatGPT-only product features; and
- large enough that the separate meter and engineering overhead are justified.
Examples include nightly classification, extraction across a stable document set, an evaluation harness, or a queued transformation job. This is workload separation, not a way to evade a plan limit. Keep each API project inside an approved budget and use the platform’s spend controls.
A hypothetical document run
The following is arithmetic from the 28 August rate card, not a claim about typical document size or model usage.
Suppose a studio processes 2,000 documents in one month. Each request averages 4,000 input tokens and 800 output tokens, with short context, no cache hit, and no separately billed tools.
- Input: 2,000 × 4,000 = 8 million tokens
- Output: 2,000 × 800 = 1.6 million tokens
| Route | Input cost | Output cost | Total |
|---|---|---|---|
| Standard | 8 × $4 = $32 | 1.6 × $20 = $32 | $64 |
| Batch or Flex | 8 × $2 = $16 | 1.6 × $10 = $16 | $32 |
Batch or Flex is cheaper in this illustration because the job accepts those processing modes. It is not the right comparison for live chat or interactive coding. If a request crosses the long-context threshold, uses tools, writes cache entries, or requires regional processing, recalculate with the corresponding rows on the pricing page.
The previous public Sol list was $5 input and $30 output per million tokens. At those rates the same hypothetical run would be $88 Standard. Use that only as a stress test for whether the workload still works at a higher rate; OpenAI has not promised that $5/$30 will be the post-promotion price.
A one-pass routing decision
- Label last month’s Sol work interactive or unattended.
- Leave interactive conversations and coding sessions on the product surface that makes them useful.
- Estimate input, cached input, cache writes, output, tools, and processing mode for each unattended job.
- Price the job at the live API rate card, then test a higher-rate scenario.
- Move only the jobs whose savings exceed the added engineering and operational cost.
- Recheck the official pricing page before 21 November and before any client quote.
For a broader model-routing test, see When Not to Buy the Top AI Model. The useful response to the Sol discount is not to cancel every seat or push every task onto an API. It is to keep the product where a person needs the product and put measurable unattended volume on the meter that now costs less.