The frontier is open.
And with think, it is already here.
The most intelligent open-weight models on earth, side by side, running on the most efficient AI infrastructure on earth.
One dial. Every machine resizes itself to fit.
Native precision
Change the precision. Every machine resizes to fit.
The frontier moves every few weeks.
With think, you are always already on it.
No compromise on performance,
efficiency, or price.
The same silicon, the same ILM, the same Constellation fabric. Pick how you want to hold it, and the numbers underneath follow.
Ninety-six cents, against seven to thirty dollars.
What it costs to produce a million tokens on a node you own, against calling a frontier model you do not. Blended input and output, published July 2026 rates, no discounts on either side.
- think node, owned outrightyours $0.96
- Gemini 3.1 Protheirs $7.00
- OpenAI GPT-5.4theirs $8.75
- Claude Sonnet 4.6theirs $9.00
- Claude Opus 4.8theirs $15.00
- OpenAI GPT-5.5theirs $17.50
- Claude Fable 5theirs $30.00
Frontier tier only. Smaller and cheaper API models exist, but nobody buys a node to serve a small model, and none of them run inside your own perimeter. Owned cost is sustained throughput on a think node over a 36 month life including power. Model names are trademarks of their respective owners.
Fifteen thousand a month, all in.
384 GB of the same Blackwell silicon. Their rates are normalised upward to match ours: block storage brought to parity, a support plan at 8%, and 5 TB of egress. think Grid includes all three in the rate. None of them include any of it.
- think Grid, Riyadhyours $15,000
- Google Cloud, Dammamtheirs $15,765
- Azure, East UStheirs $16,060
- Google Cloud, Mumbaitheirs $17,081
- Azure, East US (4x)theirs $18,630
- AWS, Mumbaitheirs $19,759
- AWS, Frankfurttheirs $20,573
Published on-demand rates, August 2026, normalised for storage, support and egress.