DRAGSTRIP

Same agent. Same question. One lane runs through Paritok. Watch the token bills split.
LLM – Paritok GPU –
Both lanes run the identical agent — same model, tools, and turn budget. The blue lane's requests pass through Paritok's compression first; the meters show what that does to the input-token bill. Quality is checked at the end by a judge that can't see which answer came from which lane.
RAW
0 tok in
PARITOK
0 tok in
Cumulative input tokens billed by the LLM API, per lane. Shorter bar = smaller bill.
–
input tokens saved
–
saved this race
–
saved at Claude Sonnet rates
the bill most coding agents pay ($3/M input)
–
tool output kept, after compression