OpenAI cut the price of its best model twice in under a month. On July 30, Luna dropped 80% and Terra 20%. Last Friday, Sol’s output tokens fell from $30 to $20 per million. A third, overnight.
For marketers evaluating AI tools for a real workflow, that changes the routing decision.
OpenAI named why: Anthropic and cost pressure from the Chinese labs.
This is an API price cut. Your $20 Plus subscription didn’t change. Neither did Pro or Business. If you use ChatGPT through the app, it does not help your wallet. But your developers, vendors, and tool providers all got a 20% cut. How much did you get back?
The ChatGPT three…
GPT-5.6 ships in three variants. Same 1,050,000-token context window, same February 2026 knowledge cutoff. What changes is price and raw intelligence.
| Model | Input / Output (per 1M tokens) | Intelligence index |
|---|---|---|
| Sol (frontier) | $4 / $20 | 59 |
| Terra (balanced) | $2 / $12 | 55 |
| Luna (volume) | $0.20 / $1.20 | 51 |
Sol costs 20× more per input token than Luna and buys you eight points on Artificial Analysis’ index.
Sol is $4 in and $20 out, it now undercuts Claude Opus 5 ($5 / $25) on both sides. A month ago Sol was the expensive one.
You pay 20× more per token for eight points. Most work doesn’t live in those eight points.
Which one, when
Luna can be your default. Test running it on extra high and you might be surprised by what you get back.
Terra on High is my workhorse. Drafting, analysis, code, anything multi-step, anything running unattended. It matches what was OpenAI’s flagship one version ago at roughly half the cost. This is where I do 70% of my ChatGPT work.
Sol is for the hard things. The one gnarly analysis, the agent run that has to hold a complicated workflow together for an hour, the piece of code you can’t debug. Reach for it deliberately, not by default.
That’s how I run it. My main agent sits on Terra with reasoning on high. A second instance runs Sol for the heavy jobs. Luna handles the routine volume and none of my decisions.
One last trap
Cross 272,000 input tokens in a single request and the price doubles on the entire prompt, not just the overflow. You have a million tokens of room. That doesn’t mean dumping your whole Drive into every call.
What to test this week
- Look at your three highest-volume AI workflows. Test moving anything mechanical down to Luna.
- Pick your default model and hold the line.
- Find the workflow you re-send the same long context to. Make sure it’s actually reusing the cache: same model, same tools, same format, within 30 minutes.
- Note the fine print: Sol’s new price is promotional, guaranteed only through November 21.
Models are becoming a commodity. Two price cuts in a month, and the floor keeps moving.
The model isn’t the edge. The edge is picking the right one for the job, then leaving the workflow alone long enough to keep the cache in a low cost run.
Don’t chase models. Ship workflows, loops, and processes that move you forward at 10x less cost than your competition.