Skip to content
All news

Devin adds efficiency scoring and cost controls for AI coding budgets

Cognition introduced Efficiency Score, automatic approval for extra spend and Devin Coach in Devin. The score attempts to estimate productive engineering hours relative to ACU spend, but it is a vendor metric, not an objective measure of employee performance. Before connecting it to budget policy, teams should define outcome checks, permissions and retention controls they own.

Otomatzia editorial · Analysis published

Source published: · Cognition / Devin

Illustration: a team setting controls and human approval for an AI agent budget
Original illustration made with AI for Otomatzia. For illustration only.

What changed in Devin

On October 19 Cognition introduced three capabilities: Efficiency Score, efficiency-based spending controls and Devin Coach. According to the company, the score estimates productive engineering hours for sessions, including investigation and planning without a PR, relative to ACU spend. This is a vendor claim, not an independent test by us.

Who may benefit and what to decide

Teams managing usage across several developers may get a mechanism for allocating extra budget according to rules. According to the company, customers on a Devin Enterprise plan can enable automatic approval for extra spend for users with a Healthy or Satisfactory grade; other requests go to an admin with the reason, efficiency, model mix and consumption. The key decision is not whether to trust the score, but which amount, user and action type may proceed without human approval and what requires an outcome check.

The limitation of a productivity metric

Estimated productive hours against ACU spend are not the same as business value, code quality or employee performance. A long investigation session may prevent a failure, while a short result may look efficient without being correct. A vendor metric is also not a substitute for measurement in your team's context. Do not use the score alone for employee evaluation, compensation or access denial.

A rollout checklist

System recommendation: define testable outcomes in advance, such as passing tests, time to merge and repeat-failure rate, and compare them with cost. Set a spend threshold and approval path, document who can change the policy and check that the admin sees the needed context. Restrict permissions, define a usage-data retention and deletion period, and start with a pilot where the score is only one budget signal. After several weeks, check whether it supports decisions without creating incentives for work that only looks efficient.

How do you apply this in a business?