Asana's $12K Codex Migration vs. a $6M Staffing Estimate: What the Cost Gap Really Means

OpenAI published one of the clearest coding-agent cost case studies yet: Asana says it removed the Enzyme testing framework from its codebase in roughly two calendar weeks, spending about $12,000 on models and infrastructure. Its earlier staffing plan was estimated at roughly $6 million and at least five years.

Those numbers come from OpenAI's August 18 customer story, not an independent audit. They should not be generalized into a promise that every migration becomes 500 times cheaper. But the project is still unusually valuable because it reports an actual bill, a defined engineering outcome, an operating pattern, and the abandoned alternative. That is far better evidence than a benchmark score without a workload.

What Asana Actually Did

Enzyme had become a modernization blocker after falling out of active maintenance. Asana started from a five-sentence prompt and ran as many as four coding agents in parallel, each with a separate copy of the repository. An engineer checked progress twice daily and reviewed every proposed change. OpenAI reports 1.5 weeks of engineering effort spread across two calendar weeks before Enzyme was fully removed.

The $12,000 figure covers model and infrastructure cost. It does not mean the total economic cost was $12,000. Human planning, twice-daily supervision, code review, validation, and organizational context still had value. The fair comparison is therefore not $12,000 versus zero people. It is the agent bill plus concentrated expert oversight versus a multi-year staffing plan that the company believed was impractical.

The Right Unit Is Cost per Accepted Migration

Token price alone cannot explain this result. Four parallel agents can consume tokens aggressively, and a frontier model can look expensive per million tokens. The economic result came from converting that spend into an accepted repository-wide change. Divide the reported model and infrastructure bill by the completed outcome and the direct AI cost is $12,000 per accepted migration. Compare that with the value of removing an upgrade blocker, not with a cheap chat request.

This is why cost-per-token leaderboards can mislead engineering managers. A cheaper model that repeatedly misses obscure test behavior, creates review churn, or stalls halfway through a codemod may cost less per request and more per completed migration. For long-running agent work, measure accepted pull requests, passing test conversions, defects found after merge, and human review hours alongside tokens.

Why This Workload Was a Strong Candidate

• The destination was testable. Removing a framework creates a crisp end state: no remaining dependency, migrated tests, and a passing suite.
• Tasks could be partitioned. Separate repository copies allowed several agents to work in parallel without sharing one fragile working tree.
• The work was repetitive but contextual. Mechanical conversion helped, while agents could still inspect local patterns and adapt.
• Experts stayed in the loop. Review was not outsourced. Human attention moved from typing every edit to checking proposed changes.

A Budget Template for Your Own Migration

Start with a paid pilot that covers one representative package or five percent of the affected tests. Track agent spend, wall-clock time, reviewer hours, first-pass test success, and regressions. If a $500 pilot converts 400 tests with 90% acceptance, you have evidence for roughly $1.25 of direct AI cost per attempted test before review. If acceptance falls to 30%, the same headline token rate hides a much worse outcome.

Set separate caps for discovery, implementation, and repair. Parallelism shortens calendar time but can multiply cost quickly, so raise the agent count only when work units are independent and merge conflicts are controlled. Preserve a clean baseline, require deterministic tests, and stop a shard after repeated failure rather than funding endless retries.

Keep a control group of work completed with the current process. A migration pilot can appear spectacular if the old estimate was never tested or if easy files are selected first. Compare similar packages, record which tasks were excluded, and calculate the cost of merging parallel branches. Also monitor the code for several weeks after completion: regressions, flaky tests, and reviewer corrections belong to the project even when they arrive after the agent invoice.

The Practical Takeaway

The most important figure is not the dramatic $6 million comparison. It is that Asana found a previously deferred maintenance project with measurable completion criteria, gave agents parallel but isolated workspaces, and paid for review. That pattern is repeatable even when the exact savings are not. Old test migrations, API upgrades, repetitive type fixes, and dependency removals are good places to search for similar economics.

Before copying the result into a board deck, include human oversight and risk in the denominator. Then compare total cost per accepted outcome. Use our AI Cost Calculator to model the token component, and add reviewer hours to build the honest business case.