21.7x
fewer Claude tokens in one matched development task.
In this experiment, Claude token usage fell 95.38% and the Claude account cost fell 78.84%. Both outputs were checked against the same task reference.
The video illustrates the task. The figures below compare the recorded Claude Stock run with the run using happyllama.
These figures describe one matched development task, not a guarantee of savings on other work. Local model computation is reported separately; hardware, electricity and operator time are not included in the Claude account cost.
Same task, same model, same reasoning effort, and the same machine. The scene is a 4096 particle vortex integrated at a fixed timestep; every number is read out of the running page and checked against a reference written before either side saw the task. Costs are what the account was billed.
Have a project in mind?
Start with the problem your customers need solved. We will recommend a scope and quote the work before it begins.
Talk through my project