OpenAI says Codex reaches 25 million users as Luna usage jumps after price cuts
OpenAI says Codex has reached 25 million users, while usage of its lower-cost GPT-5.6 Luna model increased roughly tenfold following an 80% price reduction, according to CFO Sarah Friar’s September 7 remarks at the Goldman Sachs Communacopia + Technology Conference.
The figures received fresh coverage on September 9, but the underlying price change dates to July 30; they describe adoption following an earlier commercial decision, rather than a new model release or a price cut announced today.
The combination raises an economic question for both customers and the supplier: lower unit prices can make substantially more work affordable while still increasing total spending when consumption grows faster than prices fall.
··········
CODEX ADOPTION AND LUNA USAGE MEASURE DIFFERENT THINGS.
A product user count cannot be read as a model’s customer base or paid subscriber total.
Codex’s reported 25 million users and Luna’s reported 10× usage growth should remain separate metrics: the former counts people using a product, while the latter describes an increase in model consumption.
The cited remarks do not establish a daily or monthly activity window for the Codex figure, identify how many users pay, or specify whether Luna’s usage multiplier measures tokens, requests or another unit.
Without those definitions, dividing the two figures cannot produce a reliable measure of consumption per user, revenue per account or conversion to paid subscriptions.
........
Figure | What is reported | What remains unspecified |
|---|---|---|
25 million | Codex users, according to Friar. | Activity window, paid share and user-level consumption. |
Approximately 10× | Luna usage growth after the price reduction. | Measurement unit, comparison period and workload mix. |
80% lower prices | Official Luna price change announced July 30. | The change in any individual customer’s total bill. |
........
··········
THE PRICE CUT REDUCED LUNA’S UNIT COST.
The July announcement changed API rates and subscription credit consumption.
OpenAI’s July 30 announcement set Luna’s API prices at $0.20 per million input tokens and $1.20 per million output tokens, while lowering Terra’s rates by 20% to $2 and $12 respectively.
OpenAI said subscription prices and quota budgets remained unchanged, with Luna and Terra consuming fewer credits in Codex and ChatGPT Work; a reduction in model usage costs therefore should not be described as an 80% reduction in the subscription fee.
At those announced Luna rates, an illustrative uncached workload containing one million input tokens and 200,000 output tokens costs $0.44 in model charges: $0.20 for input plus $0.24 for output.
This calculation excludes tool charges, infrastructure and human review, and serves as a reproducible workload example using the announced rates rather than an estimate of a typical customer’s bill.
··········
TEN TIMES THE VOLUME CAN STILL MEAN TWICE THE SPENDING.
A normalized scenario separates lower prices from lower aggregate expenditure.
An 80% reduction leaves the unit price at 20% of its previous level, so a customer could purchase five times as many identical billable units before returning to the original spending level.
If comparable billable volume instead increases tenfold, the arithmetic becomes 0.20 × 10 = 2.00: twice the original expenditure despite the lower price.
........
Scenario | Unit price index | Volume index | Spending index |
|---|---|---|---|
Before the reduction | 100 | 100 | 100 |
Same volume after reduction | 20 | 100 | 20 |
Five times the volume | 20 | 500 | 100 |
Ten times the volume | 20 | 1,000 | 200 |
........
This is a conditional calculation, not an OpenAI revenue estimate: it assumes identical units, an unchanged input/output and caching mix, and full exposure to the reduced price.
Because Friar’s usage multiplier is not defined in those terms, it cannot establish that Luna revenue doubled, and fixed subscriptions further weaken any direct link between consumption growth and recognized revenue.
··········
THE BUSINESS TEST IS COST PER ACCEPTED RESULT.
Retries, review time and task quality determine whether cheaper inference creates savings.
Friar’s July framework for evaluating AI emphasizes completed work, the full cost of successful tasks, reliability and value as deployment expands; it explicitly includes employee time, retries, review and rework.
For example, a team spending $100 on model calls and $400 on review for 100 accepted tasks incurs $5 per accepted task; cutting model charges by 80% lowers that total to $420, or $4.20 per task, if review effort and accepted output remain unchanged.
The resulting saving is 16%, rather than 80%, because inference represented only one fifth of the original cost; additional retries or longer review could reduce that benefit further.
Codex adoption and Luna consumption indicate scale, but the operational evidence of value is whether a deployment delivers more accepted work within its total budget, with quality and human supervision measured alongside model charges.
··········
FOLLOW US FOR MORE.
DATA STUDIOS
