это вообще бредятина
Таковы независимые от производителей
исследования.
A version of DeepSeek’s R1 model required 308,186 watt hours to complete a task with reasoning turned on, and only 50 watt hours when the feature was turned off.
One of Microsoft’s Phi 4 reasoning models used 9,462 watt hours to complete tasks with reasoning turned on, and 18 watt hours with reasoning turned off.
OpenAI’s GPT-OSS model showed a reduced disparity, consuming 8,504 watt hours to complete tasks with reasoning set to “high” and 5,313 watt hours with reasoning set to “low”.
Complex tasks
The researchers ran all the models on the same hardware, with the same prompts ranging from relatively simple queries to complex math problems.