On July 30, OpenAI lowered API prices for GPT-5.6 Luna and Terra[1]. Luna, the cheapest model in the family, now costs 80 percent less, while Terra, the balanced model for everyday work, is 20 percent cheaper, and usage of both now consumes proportionally fewer credits in Codex and ChatGPT Work. Sol keeps its existing pricing but gains a new Fast mode that replaces Priority Processing. The move passes the efficiency gains described in the previous day's engineering post straight through to customer pricing[1].
Luna down 80 percent, Terra down 20 percent
As of July 30, API pricing for Terra is 2 USD (about 320 yen) per million input tokens and 12 USD (about 1,900 yen) per million output tokens. Luna is priced at 0.20 USD (about 32 yen) per million input tokens and 1.20 USD (about 190 yen) per million output tokens[1].
Before the change, Terra cost 2.50 USD (about 400 yen) for input and 15 USD (about 2,400 yen) for output, while Luna cost 1 USD (about 160 yen) and 6 USD (about 960 yen)[2].
| Model | Before (input / output) | After (input / output) |
|---|---|---|
| GPT-5.6 Terra | 2.50 USD / 15 USD | 2 USD / 12 USD |
| GPT-5.6 Luna | 1 USD / 6 USD | 0.20 USD / 1.20 USD |
※Prices per million tokens. 1 USD = 160 JPY (as of July 31, 2026)
The same reductions apply to how usage is counted against paid subscriptions in Codex and ChatGPT Work. Subscription prices and quota budgets are unchanged; Terra and Luna simply consume fewer credits[1]. Availability is unchanged as well: in ChatGPT Work and Codex, Free and Go users can access Terra, while Plus, Pro, Business, and Enterprise users can choose Terra and Luna. OpenAI said the pricing change would begin rolling out in AWS later the same day[1].
On Luna's positioning, OpenAI says it delivers performance comparable to models that were frontier-class a year ago at roughly 6 percent of the cost per task, and at nearly nine times the speed. On professional work as measured by Agents' Last Exam, Luna outperforms Fable 5 at an estimated cost per task nearly 99 percent lower[1].
Fast mode for Sol: 2.5 times the speed at twice the price
Sol's pricing is unchanged, but the API gains a new Fast mode. It replaces Priority Processing and delivers up to 2.5 times faster speeds than standard processing at twice the price, with no change in intelligence[1].
Backward compatibility is preserved. Existing API requests tagged priority will automatically use Fast mode, aligning with /fast in Codex[1]. That makes it easier to route latency-sensitive work to Sol while pushing high-volume processing to Luna.
The savings come from efficiency at every layer
The price change is not the product of model improvements alone. OpenAI says it worked across three layers: the models, the inference systems that run them, and the agentic harness that connects them to tools and context. Better routing keeps hardware productive, optimized production software generates tokens more efficiently, and smarter context management stops agents from repeating completed work[1].
The company also restated how much of that optimization work GPT-5.6 Sol handled itself. Within a human-led process, Sol autonomously rewrote production kernels, designed and ran hundreds of experiments to improve token generation, and monitored training while intervening when problems arose. The kernel work cut the end-to-end cost of serving the model by 20 percent, and the experiments improved token-generation efficiency by more than 15 percent[1].
That picture, in which inference costs fall through hardware, software, and model design at once, matches outside analysis. Pareekh Jain, principal analyst at Pareekh Consulting, argues the cuts are broadly sustainable in the long run because new chips, better software, and more efficient model designs will keep pushing down cost per token[2].
Enterprises are likely to use more AI, not spend less
Cheaper tokens do not automatically shrink corporate AI budgets. Jain says the biggest impact for CIOs is scaling adoption rather than cutting costs: moving pilots into production, expanding AI across more employees and business processes, and making complex agentic workflows that require multiple model calls economically viable. Savings, in his view, get reinvested into higher usage, so companies get more reasoning and automation for roughly the same spend[2].
He frames this as the Jevons paradox, where efficiency gains tend to increase overall consumption of a resource rather than reduce it. Chandrika Dutt, research director at Avasant, echoes the point, saying enterprise teams will likely use the new headroom to build sophisticated agentic workflows that were previously hard to justify economically[2].
OpenAI frames the change around workload selection too. In one example it offers, a coding workflow uses Sol to resolve uncertainty and define the plan, then hands off to Luna to implement well-specified changes, write and run tests, and evaluate the results[1]. Large-scale document analysis, customer-interaction classification, and routine implementation become economical to run broadly, while only the most consequential steps pay the premium for speed.
Summary
GPT-5.6 Luna and Terra API prices dropped by 80 percent and 20 percent respectively on July 30, and both models now consume proportionally fewer credits in Codex and ChatGPT Work. Sol keeps its price while gaining Fast mode, which offers up to 2.5 times the speed at twice the cost. The savings come from efficiency work across every layer, some of it performed by Sol itself. Analysts expect enterprises to respond by running more AI work on the same budget rather than spending less.
Source[1]: https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/
