OpenAI cut its prices in half the same week Claude Opus 5.5 launched. That's not a coincidence.
GPT-6 Sol now costs $2 per million input tokens, and Luna costs 10 cents. Here's what those numbers mean in real life, and why OpenAI's own benchmark deserves a raised eyebrow.
Anthropic launched Claude Opus 5.5 and, within minutes by most accounts, OpenAI announced it was halving the price of its newest models. If you’ve ever watched two petrol pumps on the same road, you’ve seen this movie before.
The new prices
OpenAI’s two new models are GPT-6 Sol (the big one) and GPT-6 Luna (the small, cheap one).
| Model | Input, per million tokens | Output, per million tokens |
|---|---|---|
| GPT-6 Sol, before | $4 | $20 |
| GPT-6 Sol, now | $2 | $10 |
| GPT-6 Luna | $0.10 | see OpenAI page |
| Claude Opus 5.5 | $4 | $20 |
“Per million tokens” is how these companies price everything, and it means almost nothing to a normal person, so here’s the translation I used in the reel. A million tokens is roughly 750,000 English words. That’s somewhere around eight full-length novels. For Luna, that’s ten cents.
Ten cents to read eight novels. A few years ago that sentence would have sounded made up.
The benchmark I don’t fully trust
OpenAI says that on a business-workflow test, Sol beat Claude Opus 5 at about 9% of the cost.
Three things bother me about that:
- It’s OpenAI’s test and OpenAI’s numbers. Nobody independent has run it yet.
- They compared against Opus 5, not the new Opus 5.5 that launched the same week.
- On one hard coding benchmark, the new Sol actually scored lower than the previous version, 68.8% against 72.7%.
None of that means Sol is bad. It means the headline number is marketing until someone else reproduces it.
Who actually wins
Developers and anyone building apps on top of these models, mostly. When the cost of calling a model halves, a lot of small ideas that weren’t worth building suddenly are, like a WhatsApp bot for a local shop, or an app that summarises every PDF a student gets in a semester.
For regular users on a ChatGPT or Claude subscription, you won’t see your monthly bill change because of this. You’ll see it later, in cheaper apps and more generous free tiers.
The bit I’m watching is whether Anthropic answers back on price. Opus 5.5 is now twice the price of Sol on paper. If the quality gap isn’t big, that’s a hard sell.
Related: OpenAI’s sandbox incident and training pause.
Sources
- OpenAI: Introducing GPT-6 Sol and Luna
- The benchmark figures (9% cost, 68.8% vs 72.7%) are OpenAI’s own, from the same announcement.