FrontierRL
Frontier Level Inference
Decoding throughput, tokens per second.
FrontierRL Star
0t/s
GPT-5.6 on Cerebras
0t/s
Xiaomi
0t/s
DeepSeek
0t/s
Claude Opus 4.8 Thinking
0t/s
00
~60× faster than commercial models
Star outputs measured on workstation GPUs, not datacenter-grade.
At Budget-Level Prices
Calculated per 1M output tokens, at each model’s published price.
How far does $100 go?
FrontierRL Star $0.40 / M
0M
Gemini 3.1 Pro $12 / M
0M
Claude Sonnet 5 $15 / M
0M
Claude Opus 4.8 $25 / M
0M
GPT-5.6 Sol $30 / M
0M
00
~75× more tokens per dollar
Watch Star run.
Why Star is different
Star is our own model.
Not a wrapper, not a reseller.
Everyone else is busy making their big models faster.
We started from the other end, then trained in a totally different way.
Built by engineers from
Programs
MIT TNT Research 
Together.ai Accelerator