Same models.Faster route.
Runs on
- gpt-oss
- Qwen
Same prompt. Two routes.
One prompt, one model, sent twice: the usual route, then Deference Overdrive.
openai/gpt-oss-120b, Recorded 2026-10-19 · 6 runs · median. Usual route: First token 0.67 s, Total 3.17 s. Overdrive, Cerebras: First token 0.39 s, Total 0.71 s.
A recorded race, replayed once
Same prompt, same model, same day. Times vary with load.
One suffix. Full throttle.
Add :overdrive to a model id. Your key, SDK and credit work unchanged.
Overdrive guideOne route, and no other
If the pinned provider cannot serve, the call fails.
One line changes
Point the model at its :overdrive id. Nothing else moves.
Priced by the route
Provider rates plus the 5% fee, above or below usual.
Every model, timed.
Each Overdrive model with its provider, measured speed and price per 1M tokens.
Browse Overdrive modelsSpeeds measured, never promised
Each figure names its method and its day.
Check every call.
Activity shows the provider, first token and speed of streamed calls.
ActivityOne call, in full
Provider, first token, speed, tokens and cost, on one record.
Timed when it streams
A call that does not stream shows tokens and cost.
Questions
ProductWhat is Deference Overdrive?
A route on the Deference API. Add :overdrive to a supported model id and the call goes to the one fast provider pinned for that model. Same key, same SDK, same INFERENCE credit.
ProvidersWhich provider serves it?
Each Overdrive model is pinned to one provider, named on the models page and on every call in Activity. Today every Overdrive model runs on Cerebras.
BillingDoes it cost more?
Sometimes. You pay the pinned provider's rates plus the 5% fee, which can be higher or lower than the usual route. Today input costs more on both models, and Qwen's output costs less.
FailuresWhat if the provider cannot serve?
The call fails with an error. Overdrive never falls back to a slower route behind your back. Retry, or call the model without the suffix.
ModesCan I combine it with Incognito?
No. A call takes one suffix, :overdrive or :incognito. Overdrive changes the route, not how your data is handled: it carries none of Incognito's protections.