OpenAI cut the API price of two GPT-5.6 models on 30 July, about three weeks after the family launched. The cheapest tier fell by 80 per cent and the mid tier by 20. The flagship did not move.
The company attributes the reduction to efficiency gains, not competitive pressure, though it is arriving in a market where cost has become the main argument.
What actually changed
| Model | Input, before | Input, after | Output, before | Output, after | Change |
|---|---|---|---|---|---|
| GPT-5.6 Luna | $1.00 | $0.20 | $6.00 | $1.20 | −80% |
| GPT-5.6 Terra | $2.50 | $2.00 | $15.00 | $12.00 | −20% |
| GPT-5.6 Sol | $5.00 | $5.00 | $30.00 | $30.00 | Unchanged |
Prices per million tokens, as announced on 30 July 2026. Both tiers were cut by the same percentage on input and output, so the shape of each model's pricing is unchanged — only its level. Sol is the flagship, built for the hardest reasoning and agentic coding work.
The ladder inside OpenAI's own range
The flagship price held firm while the cheapest model fell by four fifths. The real story is not the new price of any one model, but the new gap between them.
| Comparison | Before | After |
|---|---|---|
| Sol against Luna, input | 5× | 25× |
| Sol against Luna, output | 5× | 25× |
| Terra against Luna, input | 2.5× | 10× |
| Terra against Luna, output | 2.5× | 10× |
RECATOOLS arithmetic on the announced prices. OpenAI's internal price range has gone from top-to-bottom five to one, to twenty-five to one, without the flagship changing at all.
A five-fold gap invites developers to reach for the better model when unsure, because the penalty for over-specifying is modest. A twenty-five-fold gap does not. Routing work to the cheapest capable model is no longer a fine-tuning step; it is the default way to build. That changes application architecture, not just the final bill.
What a job costs now
Token prices are hard to feel. Applied to a single job — one million tokens in, two hundred thousand out — the difference is legible.
| Model | Cost before | Cost after | Change |
|---|---|---|---|
| GPT-5.6 Luna | $2.20 | $0.44 | −80% |
| GPT-5.6 Terra | $5.50 | $4.40 | −20% |
| GPT-5.6 Sol | $11.00 | $11.00 | Unchanged |
RECATOOLS calculation on a hypothetical job of 1,000,000 input tokens and 200,000 output tokens, at the announced rates. Real workloads differ in their input-to-output ratio, and output tokens dominate the bill in generative work while input dominates in retrieval-heavy work.
The same job now costs twenty-five times more on Sol than on Luna. Three weeks ago it cost five times more.
Fast mode, and the arithmetic to check
Alongside the cuts, OpenAI replaced its Priority Processing option with a Fast mode, described as delivering up to 2.5 times standard speed at twice the price. Existing priority-tagged requests remain compatible.
| Model | Standard input | Standard output | Fast mode input | Fast mode output |
|---|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $30.00 | $10.00 | $60.00 |
| GPT-5.6 Terra | $2.00 | $12.00 | $4.00 | $24.00 |
| GPT-5.6 Luna | $0.20 | $1.20 | $0.40 | $2.40 |
Per million tokens. The Fast mode columns are RECATOOLS arithmetic — each tier's post-cut rate doubled, per the stated "twice the price" multiplier — not a separately published rate card. Check them against OpenAI's own pricing page before budgeting against them.
The phrase carrying the weight is "up to". Paying double for 2.5 times the speed is a 20 per cent improvement in cost per unit of work — but only at the ceiling. At exactly twice the speed it is a wash. Below that it costs more per unit of work than the standard tier, and the guarantee is on price rather than on throughput.
The reason OpenAI gives
The day before the cut, five members of OpenAI's technical staff published an engineering post describing optimisations across inference and the agent harness behind Codex and ChatGPT Work. The post cites a 20 per cent cut in end-to-end serving costs from GPU kernel rewrites, and a gain of more than 15 per cent in token-generation efficiency from a speculative-decoding redesign.
The company said it would pass "under-the-hood improvements back to our users and customers in the form of more widely available, cost-efficient intelligence".
That framing is the company's own and should be read as such. An 80 per cent price cut is larger than a 20 per cent cost reduction and a 15 per cent efficiency gain can explain on their own. Efficiency may explain what made the cut affordable; it does not by itself explain the size of it.
Why cheaper tokens may not mean cheaper bills
A price cut of this size reads as a saving only if consumption holds still, and in this market it has not.
The workloads absorbing most of the growth are agentic rather than conversational — systems that plan, call tools, read the results and iterate. A single such task can consume orders of magnitude more tokens than a chat exchange, because the model re-reads its accumulated context on every step. Teams that moved from chat interfaces to agent loops over the past year have generally watched their token counts rise faster than any list price has fallen.
The mechanism to watch is not the rate card. It is that cheaper units make previously uneconomic workloads viable; those workloads get built, and aggregate spending rises even as the price of each token falls. That explains how the same week can bring an 80 per cent price cut and record infrastructure spending without any contradiction.
Where the pressure is
Most reporting on the cut points to pressure on OpenAI from cost-sensitive enterprise buyers and competing labs, especially Chinese model providers. Luna now undercuts Anthropic's Haiku 4.5 by roughly a factor of five on input and four on output, and Terra sits below the rates Claude Sonnet 5 moves to after 31 August.
The open-weight end of the market applies a different kind of pressure, since a model with published weights has no per-token price at all. We covered Kimi K3's weights going out earlier this month, and the constraint there was hardware rather than licence fees.
The other half of the picture
This lands in the same week that the largest infrastructure buyers reported their quarters. Alphabet, Amazon, Microsoft and Meta between them spent more on capital projects in three months than in any quarter on record, nearly double the same period a year earlier.
Capital costs are climbing steeply while the price of the output falls. Those two lines can only diverge for as long as someone is willing to fund the gap — which is the same question sitting underneath Nvidia's reported talks to guarantee a quarter of a trillion dollars of OpenAI's data-centre financing.
The caveats
- List prices are not what large buyers pay. Enterprise agreements, committed-spend discounts and cached-input rates all sit outside these figures.
- Price is not cost. Nothing here reveals what it costs OpenAI to serve these models, or whether any tier is sold above or below that.
- Cheaper tokens do not mean cheaper bills. Falling unit prices historically increase usage, and agentic workloads consume far more tokens per task than chat.
- Our worked example is illustrative. It assumes one input-to-output ratio; yours will differ.
- The competitive comparisons are as reported, and rival pricing moves frequently — Claude Sonnet 5's own rates change on 31 August.
Key takeaways
- Luna fell 80 per cent, to $0.20 input and $1.20 output per million tokens.
- Terra fell 20 per cent; the flagship Sol did not move.
- The internal price spread widened from 5× to 25×, which changes how applications should route work.
- Fast mode charges double for up to 2.5× speed — a saving only if you reach the ceiling.
- The stated efficiency gains do not fully account for the size of the cut.
View author profile → · Editorial policy
