GreenPT Research
The same answers, with less compute.
Every token costs compute, money and energy. Most of the tokens in a typical answer are not the answer.
We research how to get the same results from AI with less of that work. Everything we learn ships in our models, our API and the tools you already use.
Three lines of research
Less to say, less to think, or nothing to write at all. Each one removes work the model never needed to do.
- −46 to −52% Output tokens from our Honey adapters Same pass rate as the base model on 26 checkable prompts
- −51 to −71% Tokens with Shorthand Three recorded tasks, same hardware and prompt, both answers correct
- 72 ms Three answers from Metis One support ticket, three questions, 0.0047 Wh
Research areas
Where we are working.
-
Honey
Cutting the waste out of the AI workflows you already run: less code, less prose, and brevity trained into the model weights.
Read about Honey -
Shorthand models
A model that writes its thinking as short notes in symbols, then answers in plain language. Fewer tokens, so the answer comes sooner.
Read about Shorthand -
Decision models
Metis reads a case once and returns a probability for every possible answer. It generates no text, so answers come back in milliseconds.
Read about decision models
Open by default
Run it, break it, tell us what you find.
We publish our models, code and evaluations, including what did not work.
-
Qwen3.5-9B-honey
LoRA adapter: 52% fewer output tokens, same pass rate.
(opens in a new tab) -
Qwen3.8-27B-honey
LoRA adapter: 46% fewer output tokens, same pass rate.
(opens in a new tab) -
honey-for-devs
Less code and less prose for coding agents.
(opens in a new tab) -
honey-eson
A wire format with about 28% fewer tokens than compact JSON.
(opens in a new tab)
How we report
Measured, not predicted.
Every GreenPT response carries an impact report with the energy it used, so you can check our figures against your own requests. Our evaluations are paired: the same prompts and settings, with and without the change.