Skip to content

Telemetry and savings

Coder telemetry is local product telemetry. It is not a production-cohort median and it is not a guarantee. This page is the methodology for the only numeric Coder claim currently approved for public use.

Visible user turns versus model invocations

Section titled “Visible user turns versus model invocations”

A visible user prompt is counted once even when a client makes multiple model invocations. Auxiliary invocations are excluded from the input-savings total.

The full-repository counterfactual is assigned only to the first eligible invocation of a conversation.

Saved input is max(counterfactual_input_tokens - actual_input_tokens, 0).

The corrected controlled E2E database contains four eligible completed user turns.

MeasureValue
Eligible visible user turns4
Distinct test sessions4
Actual input tokens24,705
Counterfactual input tokens132,733
Avoided input tokens108,028
Weighted input-token avoidance81.3874%

Approved wording:

81.4% input-token avoidance observed in a four-turn controlled Coder E2E test against a one-time full-repository session-start counterfactual. Results vary by repository, client, model, and task.

This number may appear on the Coder product page or this methodology page with the sample size and baseline visible beside it. It is not a median, an average, a production cohort, or a guarantee.

The local dashboard can show input, cache, output, accuracy, Now, and quality views for your own sessions. Local model results are not treated as monetary savings when no paid model cost applies.

MemStrata-managed email is limited to one message per signed-in user per UTC day. Reinstalling does not restart trial access.