Effective Tokens (ET) is a metric for the token consumption of an agent run in which each kind of
token is weighted by its relative cost and the result is scaled by the cost tier of the model that
consumed it. It was described by the team behind GitHub Agentic Workflows in
Improving token efficiency in GitHub Agentic Workflows to deal with the observation
that raw token counts do not reflect cost: the same workflow run on a cheaper and a more expensive
model produces similar token counts at very different cost, so a switch of model tier looks like no
change in raw tokens. The post defines it as ET = m × (1.0 × I + 0.1 × C + 4.0 × O), where m is a
model cost multiplier (0.25 for Haiku, 1.0 for Sonnet, 5.0 for Opus), I is newly processed input
tokens, C is cache-read tokens and O is output tokens.
Usage
The weights encode the post’s reasoning about token prices: output tokens carry four times the weight of input tokens because they are the most expensive token type across all major providers, and cache-read tokens carry a tenth because they are served from cache at a fraction of the cost of fresh input. The authors state that the normalization means a 10% reduction in ET corresponds to a genuine 10% cost reduction whichever model is in use. They used it to compare runs of their own production workflows before and after optimization, and to express aggregate savings — for one workflow, roughly 7.8 million ET saved over the observation period, assuming the pre-optimization rate.
The same post is explicit about what the metric does not capture. Because the workload of a workflow running against a live repository varies from run to run, a lower ET can mean the workflow did less work rather than the same work more efficiently; the authors track LLM API call counts alongside token counts, reading steady turns per run with falling tokens per call as a genuine improvement. ET also says nothing about output quality. In one of their workflows ET rose 5% after optimization, which they attribute to a shift toward larger pull requests and a 14% rise in output tokens — the most heavily weighted term — rather than to the optimization failing.
When It Applies
ET is a working metric defined by one team, in one blog post, for measuring its own workflows. It assumes that the weights and model multipliers track actual prices, and the post gives them as fixed values. It suits comparing the cost of runs of the same workflow across model changes and optimizations, and is misread when a fall in ET is taken as evidence of efficiency without checking whether the workload or the quality of the output changed.