Devin Gets 30%-70% Cheaper: Cognition Cuts Prices While Topping Benchmarks, and $0.60/Task Is Becoming the Industry Anchor

What happened: a price cut and a benchmark top in the same week

On September 29, Cognition cut Devin prices across the board: 30%-40% off Fusion and Normal modes, 15%-20% off Ultra, and up to 70% off Devin Review. The stated reason wasn't a promotion but engineering-driven cost reduction — smarter model routing, batched tool calls, and context-cache reuse.

The same week, Cognition published Devin Fusion's score on its own FrontierCode 1.1 Extended dataset: 68.8 points at an average cost of $0.60 per task. Fusion's approach is a "lead + sidekick" dual-model architecture: the capable model handles planning and final review while the cheaper model does the execution, backed by Cognition's own SWE-2 model (post-trained from Moonshot AI's Kimi K3), which scored 50.0% on the FrontierCode 1.1 main track at 64% lower cost than Fable 5.1.

The timing is worth noting: the cut landed less than a week after Bloomberg reported Cognition had crossed $1 billion in annualized revenue (doubling in four months), closed a $2 billion round, and reached a $48 billion valuation.

Why this time is different: the discounter isn't a struggling company

The usual business logic is that price cuts mean a product can't compete on merit, so it competes on price. Cognition is doing the opposite — it just proved itself one of the fastest-monetizing companies in AI coding, then chose to cut prices anyway.

There is only one explanation: it's buying market share and pricing power, not cash flow. A company that just raised $2 billion and passed $1B in ARR doesn't need to explain survival to anyone; it just needs to cement the "cheap and strong" perception before rivals react. It's the classic "spend runway to buy share" playbook — and Cognition happens to have both.

The cut is also technically defensible. Fusion's "lead + sidekick" architecture is essentially about spending every cent of the expensive model where it counts: planning, disambiguation, and final review on the frontier model; codebase exploration, code writing, lint fixes, and test writing on the cheap one. Batched tool calls and cache reuse compress repeated inference further. Every cent of the reduction comes from engineering, not subsidy — or so Cognition claims, with the 68.8 benchmark score as its backing.

The real weapon: the number $0.60 per task

But more interesting than the cut itself is the number Cognition published: $0.60 per task on average.

GitHub Copilot and Cursor still charge per seat: a fixed price per human per month. Seat pricing assumes "one human logs in and does a fixed amount of work a month." But autonomous coding agents don't work that way at all: they run unattended, burn tokens in amounts that swing wildly with task size, and the vendor's true cost moves with every underlying model call. Under that working model, seat pricing is essentially guesswork.

Cognition's move is simple: cut prices while publishing a per-task cost benchmark. That gives enterprise buyers an anchor to hold against every vendor — "Devin costs $0.60 a task, what about you?" Whoever defines the per-task pricing reference point first owns the next round of pricing negotiations. Copilot has 4.7 million paid users and Cursor is backed by SpaceX, but neither has published an equivalent number.

It's also why "pay per task / pay per output" is the endgame for AI coding: once an agent can genuinely deliver a PR, a migration, or a code review on its own, charging per human head looks absurd. Cognition is just pulling the endgame forward.

Three caveats for vibe coders

The good news first: a price war is good for individual developers. Devin's Core tier at $20/month plus usage is now meaningfully cheaper for small-to-medium tasks; more importantly, cost transparency is becoming an industry norm — when picking an agent tool you can now ask three direct questions: what's my typical task's average cost? Who pays for failed retries? Is the touted routing-and-caching optimization real engineering or marketing?

Now the cold water. First, FrontierCode is Cognition's own dataset — 68.8 is a "wrote the exam and took it" score. Treat benchmarks as reference, not scripture; until independent replication exists, every "number one" deserves a discount. Second, $0.60 is the average task — your tasks aren't average: a medium-sized refactor burning tens of dollars is still common, so do the math on your own workflow. Third, 70% off Devin Review sounds delicious, but code review is exactly where "cheap" is riskiest — money saved is worthless if it buys missed bugs.

One-line summary: Cognition is using price cuts to rewrite the rules of the game, and rewritten rules are always good for buyers. But before buying, do your own math — don't pay for a vendor's benchmark score.