Who profits when your agent is inefficient? Cognition's president keeps asking

Cognition’s president Russell Kaplan spent late July and early August making one argument in three venues. LangChain published his interview on the misaligned incentives behind AI coding agents on July 30, followed on August 5 by a session on why Cognition built FrontierCode after SWE-bench saturated. On August 6, AI Engineer’s State of Model Routing panel put Cognition between NVIDIA and OpenRouter. Between them, the three appearances have circulated widely — community re-cuts with Chinese and Ukrainian subtitles were on YouTube within days.

The through-line is an economics claim, and it’s worth taking seriously because Cognition has now published enough product telemetry to be held to it.

The incentive problem, stated plainly

The argument, as Cognition has made it in writing: when agent usage is metered by tokens consumed, the vendor’s revenue grows with the agent’s inefficiency. An agent that takes 26 turns to finish a task bills more than one that takes 11. The Devin Fusion launch post opens with the customer side of that ledger — “Engineering teams are lighting money on fire. It’s no longer sustainable to use the most expensive models on every task” — and the whole post is a description of machinery built to spend less: a frontier lead model that delegates to a cheaper sidekick, plus mid-session model switching timed to context compaction so it costs nothing extra.

That machinery only makes commercial sense if the vendor benefits when the agent is cheap. Cognition sells Devin in plan tiers with ACU-based usage (devin.ai/pricing); publishing that Fusion holds frontier-level FrontierCode scores at 35–60% lower cost is, in effect, advertising a smaller bill.

FrontierCode is the other half of the argument

An efficiency claim needs an eval that can’t be gamed by speed alone. Kaplan’s August 5 talk covers why Cognition stopped leaning on SWE-bench — too easy for modern models — and built FrontierCode, which grades both correctness and code quality on real-world tasks, with a 1.1 revision that audited its own grading criteria and demoted 75 overly strict blockers.

The two artifacts lock together: FrontierCode supplies the score axis, Fusion optimizes the cost axis, and Cognition publishes the frontier — score versus dollars per task — with its own product on it. Whatever you think of vendors grading themselves, this is a more falsifiable posture than per-token pricing plus a leaderboard screenshot. (Our /bench page tracks the independent-measurement side of this, where Devin itself notably is not benchmarked — LiveBench scores models, not agent products.)

Why the timing matters

Cognition is making this argument while it is, by its own published description, unusually exposed to it. Fusion’s best measured configuration used Fable 5, whose access was suspended by a US government directive in June and hasn’t returned. The company’s answer — dynamic routing across whatever frontier models are available, graded on its own task-level eval — is precisely the posture you’d adopt if you expected model access, pricing, and capability to keep shifting under you.

Whether “aligned incentives” survives contact with quarterly revenue targets is a question for every vendor in this market, Cognition included. But the argument is now on the record in three venues, attached to published numbers. That makes it checkable — which is the point.

References

1. The misaligned incentives behind AI coding agents — LangChain interview with Russell Kaplan, July 30.
2. SWE-bench is saturated, so Cognition built FrontierCode — LangChain, August 5.
3. The State of Model Routing — NVIDIA, Cognition, OpenRouter — AI Engineer panel, August 6.
4. Devin Fusion — the cost architecture and the 35–60% / 88% figures.
5. FrontierCode 1.1 — the task-level eval and its 1.1 methodology audit.
6. Devin pricing — ACU-based plans.

---

Editorial from Devin Central — a fan news desk, not Cognition. Devin is a trademark of Cognition.