A much cheaper small model

OpenAI has made the cheapest member of its GPT-5.6 family much cheaper.

From 30 July, GPT-5.6 Luna costs $0.20 per million input tokens and $1.20 per million output tokens in the API. At launch on 9 July, the same prices were $1 and $6. That is an 80% cut on both sides of the bill.

GPT-5.6 Terra also falls, though less sharply. Its new price is $2 per million input tokens and $12 per million output tokens, down from $2.50 and $15. GPT-5.6 Sol stays at $5 input and $30 output.

The change matters most to products that make a lot of routine calls: classification, document processing, background agents and well-defined coding steps. A lower token price does not guarantee a lower cost per finished task, but it changes the starting point by a lot.

For a developer sending one billion uncached input tokens through Luna, the published input price falls from $1,000 to $200. Output, tools, retries and the number of tokens a model actually uses still shape the final bill.

Speed now has a clearer surcharge

OpenAI is also replacing Priority Processing with Fast mode in the API.

For GPT-5.6 Sol, the company says Fast mode can deliver up to 2.5 times faster performance than standard processing. It costs twice as much. Existing requests tagged as priority will be routed to the new mode, so applications do not need an immediate rewrite just to keep working.

This creates a more visible trade-off. Luna is the low-cost lane. Terra sits in the middle. Sol remains the expensive model, with a second price again when waiting is the costly part.

The 'up to' matters. Real latency depends on the request, output length, tools, load and network conditions. OpenAI has not published one guaranteed 2.5-times improvement for every workload.

Teams will still need to test on their own jobs. Paying twice as much for a faster answer only makes sense when the saved time is worth more than the extra inference cost.

Subscriptions are not getting cheaper

The API cut is easy to confuse with a ChatGPT or Codex price cut. It is not one.

OpenAI says ChatGPT and Codex subscription prices remain unchanged. Quota budgets also stay the same. What changes is the rate at which Terra and Luna use credits inside paid subscriptions, so the same allowance should stretch further when those models are selected.

Free and Go users can access Terra in ChatGPT Work and Codex. Plus, Pro, Business and Enterprise users can choose Terra and Luna, according to the announcement.

That distinction is important for ordinary users. A lower API price does not automatically reduce a monthly subscription charge. It changes the economics for developers and may give subscribers more work inside an existing quota.

OpenAI also said the revised prices would begin rolling out through AWS on 30 July. Exact billing surfaces can update at different speeds, so customers should check the price shown by the service they actually use.

The pricing pages are briefly out of step

There is one practical wrinkle. OpenAI's announcement gives the new Luna and Terra prices and says they started on 30 July. When Model Current checked the public developer model catalog on 2 August, it still displayed the July launch prices: $1 and $6 for Luna, and $2.50 and $15 for Terra.

The dedicated Luna model page showed the older figures too. The newer announcement is explicit, dated and specific, but the mismatch makes the buying information less clear than it should be.

A documentation delay is not evidence that the announced cut has been withdrawn. It does mean developers should verify their account-level billing before forecasting a large workload from a marketing post alone.

The same caution applies to comparisons. OpenAI says Luna now delivers performance that was frontier-class a year ago at a small fraction of the cost. Those benchmark and cost-per-task comparisons are company analyses, not independent measurements of every production use.

The price is confirmed by the announcement. The real cost of a finished job remains something each team has to measure.

What is confirmed, claimed and still open

Confirmed: OpenAI announced an 80% API price cut for Luna, a 20% cut for Terra and no change to Sol's standard price. It also introduced Fast mode for Sol at twice the standard rate.

OpenAI's claims: the company says Sol helped its engineers reduce the end-to-end cost of serving the model by 20% and improve token-generation efficiency by more than 15%. Those are internal figures and have not been independently audited in the material published with the change.

Still open: how quickly every API, cloud and documentation surface reflects the new prices; how Luna's lower token rate translates into cost per successful task; and whether Fast mode reaches its maximum speed gain under ordinary load.

The broad direction is less uncertain. Model providers are no longer competing only on the quality of their largest model. They are competing on how cheaply smaller models can carry the unglamorous bulk of production work.

This cut makes that competition harder to ignore. It also gives buyers a simple next step: rerun the same evaluation, with the same quality bar, against the new bill.

Sources

  1. OpenAI - GPT-5.6 price and Fast mode updatePrimary announcement published 30 July 2026. Source for the new Luna and Terra API prices, Fast mode, subscription treatment and OpenAI's internal efficiency claims.
  2. OpenAI - GPT-5.6 general availability and launch pricingPrimary launch post published 9 July 2026. Used to verify the previous Luna and Terra prices and the unchanged standard price for Sol.
  3. OpenAI developer catalog - GPT-5.6 modelsPrimary live product documentation checked 2 August 2026. It still displayed the earlier Luna and Terra prices at publication time, which is disclosed in the article.
  4. OpenAI developer documentation - GPT-5.6 LunaPrimary model page used to corroborate Luna's intended high-volume role and the temporary mismatch between model documentation and the newer pricing announcement.