Cognition's Devin Now Grades Engineers to Decide Who Gets More AI Budget

Cognition has introduced efficiency-based cost controls for Devin that route requests for additional usage according to each user’s Efficiency Score. Users with healthy grades can receive automatic approval; other requests go to an administrator with the stated reason, grade, and model usage mix. The feature gives teams an alternative to equal per-user caps and puts a vendor-generated estimate of engineering output into spending decisions.

Uniform quotas simplify budgeting by assigning every user the same ceiling. They also ignore the value produced with that allowance. Cognition’s system directs more budget toward users whose sessions generate higher estimated output per billed unit, which can keep productive work running and expose expensive sessions with little estimated return.

The score: output per billed minute

Cognition estimates the equivalent human engineering hours produced by each completed Devin session. The estimate includes implementation, investigation, and planning, including work that produces no pull request. Devin uses those session estimates to assign each user an efficiency grade.

Cognition calculates the metric as follows: estimated human engineering hours divided by ACUs consumed. An ACU is Cognition’s billing unit and corresponds to approximately 15 minutes of active Devin execution. A higher score means the estimator attributes more engineering work to each unit of billed execution.

Cognition compared its hour estimates with feedback from 126 users covering 258 sessions. On held-out sessions reserved from model fitting, the logarithms of predicted and user-reported hours had a correlation of 0.74; a coefficient of 1 indicates perfect co-movement. The result supports directional agreement. Exact hour-level accuracy remains uncertain because the sample is limited and the reference values are self-reported.

The evidence supports using the score as one input to spending triage. Employee performance decisions require stronger validation, broader evidence, and safeguards against tasks that the estimator may undervalue.

Healthy grades unlock more usage

When a user requests additional usage, Devin applies the efficiency grade before deciding whether an administrator must review the request.

The Information identifies Bilt as an early user of the system. Cognition also released Devin Coach, a guidance feature intended to help users improve their efficiency grades through more effective agent use.

The business case for spending fewer tokens

Consumption billing gives Cognition an immediate incentive to sell more usage, while uncontrolled costs can weaken adoption and renewals. Efficiency controls may reduce wasted ACUs in the short term and make enterprise customers more willing to expand access when they can connect spending to estimated results.

Cognition has already made cost a product dimension in earlier releases. The company says Devin Fusion combines models to maintain frontier-level performance on its FrontierCode benchmark at 60% lower cost. It has also reported the following cost reductions:

• Fusion and Normal modes: 30% to 40%
• Ultra mode: 15% to 20%
• Devin Review: up to 70%

Those figures come from Cognition and may vary on a team’s repositories, tasks, and development environment. The new controls give customers a way to measure their own usage at the individual level.

“Tokenmaxxing” becomes measurable

Cognition uses “tokenmaxxing” for the pattern in which a small number of users consume a disproportionate share of a team’s tokens, sometimes through speculative, redundant, or poorly scoped sessions. Flat quotas contain the total exposure but provide little information about whether costly sessions produced useful work.

High ACU consumption with few user messages can reflect useful autonomous execution, repeated retries, missing environment setup, or vague requirements, according to Devin’s documentation. Combining estimated human hours with consumption gives administrators evidence about whether that execution produced value.

Devin’s tiered plans already include daily and weekly quotas. Equal caps give every user the same ceiling regardless of task value, while the efficiency grade adds an outcome-based decision when users request more capacity.

Policies to settle before rollout

• Purpose: Document that the grade governs spending approval and define any permitted secondary uses.
• Overrides: Assign an owner for exceptions and set response times so reviews do not block active work.
• Diagnostics: Inspect low-scoring sessions for environment failures, repeated retries, vague requirements, and unsuitable model choices before coaching the user.
• Local validation: Compare the vendor score with team-owned signals such as accepted changes, resolved incidents, and completed investigations.
• Data governance: Limit access to individual grades and establish retention rules for usage records.

Individual grades shown to managers can influence performance discussions. Cognition presents the metric as a budgeting signal, and its published validation measures correspondence with self-reported hours. Teams should require separate evidence for employee evaluation and audit whether the scores systematically undervalue maintenance, investigation, or complex work.