OpenAI cut API prices on two of its three GPT-5.6 models on July 30, 2026. GPT-5.6 Luna now costs 80% less: $0.20 per million input tokens and $1.20 per million output tokens. GPT-5.6 Terra dropped 20% to $2 input and $12 output. GPT-5.6 Sol, the flagship reasoning model, keeps its $5 and $30 rates but gains a new Fast mode that runs 2.5x faster at double the standard price.
If you built your cost model around the launch-day GPT-5.6 pricing, that math is now out of date. This post covers the new rate card, why OpenAI cut prices, how the numbers stack up against Google’s Gemini Flash tier, and what to change in your own API routing this week. And since the fastest way to compare model costs is to test the same request against each tier, tools like Apidog make it simple to send one prompt through Sol, Terra, and Luna side by side and measure the token spend before you commit.
The new GPT-5.6 rate card
Here is the full before-and-after, per million tokens:
| Model | Old input | New input | Old output | New output | Change |
|---|---|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $5.00 | $30.00 | $30.00 | No change |
| GPT-5.6 Terra | $2.50 | $2.00 | $15.00 | $12.00 | -20% |
| GPT-5.6 Luna | $1.00 | $0.20 | $6.00 | $1.20 | -80% |
The new prices took effect on July 30, 2026 and apply to the standard API tier. The spread between the top and bottom of the family jumped from 5x at general availability on July 9 to 25x today. That gap changes the routing decision. Three weeks ago, sending a task to Luna instead of Sol saved you 80% of the bill. Now it saves you 96%.
Prompt caching discounts still apply on top of these rates. OpenAI has not published separate cached-input prices for the new Luna and Terra rates in the announcement, so check the official pricing page before you update billing forecasts that depend on cache reads.
What happened to Sol
Sol’s list price is unchanged, but its speed options changed. OpenAI introduced a Fast mode for Sol: 2.5x faster token generation at twice the standard rate. Fast mode replaces the previous Priority Processing tier, so if your production traffic runs on Priority Processing today, you need to migrate that configuration and re-check the invoice line it lands on.
The practical read: OpenAI is segmenting Sol by latency instead of discounting it. Teams that need flagship reasoning at interactive speeds pay a premium for it. Everyone else routes down the family, which is exactly where the price cuts push you.
Why OpenAI cut prices
OpenAI frames the cuts as passed-through efficiency gains rather than a loss-leader move. The announcement credits four levers:
- Improved hardware routing across its inference fleet
- Better production inference software
- Smarter context-caching algorithms
- Production code kernels rewritten by GPT-5.6 Sol itself
That last one is the detail worth remembering. OpenAI says Sol autonomously rewrote serving kernels, cutting end-to-end serving overhead by 20% and improving token-generation efficiency by more than 15%. “Making advanced intelligence more abundant and affordable is central to OpenAI’s mission,” the company said in the announcement. Whatever you make of the mission framing, the mechanism is concrete: the model is now cheap enough to serve that OpenAI can drop Luna’s price 80% and still protect margin.
There is also an obvious competitive driver. VentureBeat’s coverage frames the cut as a direct response to Google’s Flash tier. Luna’s combined list price (input plus output) is now $1.40 per million tokens, undercutting both Gemini 3.5 Flash-Lite at $2.80 and Gemini 3.6 Flash at $9. The cheap end of the frontier is where the volume lives, and OpenAI wants it back.
What the cut means for your workloads
The 80% Luna cut is the one that changes architecture decisions, not the 20% Terra trim. A few concrete shifts:
Agent loops get cheap. Agentic workloads burn tokens in long chains: plan, call a tool, read the result, retry. At $1 input and $6 output, running those loops on Luna was defensible. At $0.20 and $1.20, it is the default. OpenAI’s own announcement calls out complex agent loops as the workload the new Luna price unlocks.
Classification, extraction, and summarization move down-tier. If you kept these on Terra because Luna’s quality margin felt thin, re-run the eval. The quality did not change; the price did. A workload that costs $100 a day on Terra costs about $11 on Luna at the new rates.
Terra as the default gets cheaper to defend. Our Sol vs Terra vs Luna breakdown recommended Terra as the sensible default for everyday work. That advice holds, and now costs 20% less to follow.
The effort dial still matters more than the rate card. GPT-5.6 exposes six reasoning effort levels, and higher effort produces more output-side tokens, which is the expensive direction on every tier. A price cut does not fix an over-provisioned effort setting.
How to re-check your own numbers
Do not take the percentages on faith; measure your actual traffic. The workflow that catches surprises early:
- Pull a representative sample of production prompts.
- Send the same requests to
gpt-5.6-luna,gpt-5.6-terra, andgpt-5.6-soland capture token counts per response. - Multiply by the new rates and compare against quality on your own eval set.
Apidog makes step 2 fast: define the chat completion endpoint once, switch the model ID with an environment variable, and run the same test scenario against all three tiers. The response viewer shows the usage block per call, so you get real input and output token counts instead of estimates. If you are wiring GPT-5.6 into an app for the first time, our guide to using the GPT-5.6 API walks through authentication, model IDs, and the request shape. You can also mock the endpoint while you wait for budget approval, then flip to live traffic without rewriting tests. Download Apidog free to run the comparison.
The bigger picture: the price war is on
This is the third major price move at the cheap end of the frontier this summer. Google refreshed its Flash tier in July, Anthropic launched Claude Opus 5 at half of Fable 5’s price, and now OpenAI has cut Luna below both Gemini Flash options. Model quality still separates the flagships, but the workhorse tier is converging on a simple rule: whoever is cheapest this month wins the batch jobs.
For API teams, that means model choice is no longer an annual decision. Build your integration so the model ID is a config value, keep a standing eval suite, and re-run the cost comparison whenever a vendor moves. The teams that treat model routing as infrastructure, with tests and monitoring around it, capture each price cut within days instead of quarters.
FAQ
What are the new GPT-5.6 API prices?
As of July 30, 2026: Luna costs $0.20 per million input tokens and $1.20 per million output; Terra costs $2 and $12; Sol is unchanged at $5 and $30. Sol also gained a Fast mode at double the standard rate for 2.5x faster processing.
Did GPT-5.6 Sol get a price cut?
No. Sol’s standard rates are unchanged. The changes to Sol are structural: a new Fast mode replaces the Priority Processing tier, trading double the price for 2.5x speed.
Is GPT-5.6 Luna now cheaper than Gemini Flash?
On list price, yes. Luna’s combined input-plus-output rate is $1.40 per million tokens versus $2.80 for Gemini 3.5 Flash-Lite and $9 for Gemini 3.6 Flash. Price is not quality, though: run both against your own eval set before switching. If you want the free route first, see how to use GPT-5.6 for free.
Do the price cuts apply automatically?
Yes. The new rates apply to standard API usage from July 30, 2026 with no code changes required. You only need to act if you use Priority Processing on Sol, which is being replaced by Fast mode.
What to do this week
The new rate card rewards teams that move fast. Re-run your cost model with the new numbers, re-evaluate which workloads Luna can absorb at 96% below Sol’s price, and check whether Terra’s 20% trim changes any routing thresholds. Then verify with real traffic instead of list-price math: one test scenario in Apidog, three model IDs, and the usage numbers will tell you exactly what the cut is worth to you.
