sdecode / speculative decoding
Every target call
has a price.
A speculative decoder drafts a short run, then asks the target model to verify it. The win is not “more tokens per second” on every machine. The win is fewer expensive target calls for the same tokens. This is the session that sends the bill.
The decision
Given your target-call price, does drafting cut the cost of a million output tokens enough to deploy?
token-identical
Session tape / measured real run
ready
Prompt
Output
Round —
— / —
DRAFTED —
ACCEPTED —
TARGET CALL —
ROUND WALL —
The deployment read
—
—
—
target spend / 1M tokens
—
target calls removed / 1M
—
pair mean tokens / call
—
calls in this session