Cognition cut the price of using Devin, its AI software engineer, by 15 to 70 percent, depending on which mode a customer runs. In a post on Devin’s blog dated 28 September 2026, the company said the biggest drop hits Devin Review, its code-checking tool, which is up to 70 percent cheaper. Fusion and Normal modes get 30 to 40 percent off, and Ultra gets 15 to 20 percent.
Cognition attributes the savings to two things: newer underlying models, including its own SWE-2, and changes to the software wrapper (the “harness”) that feeds those models their instructions and tools. The company says quality did not slip along the way. It claims it “maintained or improved” intelligence in every mode.
The evidence for that claim is a chart on FrontierCode 1.1, a coding benchmark that Cognition links from its own site. On the Extended set, the company reports Devin Fusion scoring 68.8 at an average of $0.60 per task. Its chart puts Opus 5.5 at 65.2 for $0.90, and GPT-6 Astra at 63.1 for $2.62. Fable 5.1 lands at 63.6 for $2.68. By Cognition’s numbers, Fusion costs roughly a quarter as much as the two most expensive entries while scoring higher. Those are the vendor’s figures on a test the vendor points to, and the post does not mention independent replication.
Fusion works by pairing a strong lead model with a cheaper helper model. Cognition says that design lets it swap in whichever model fits each job as prices and abilities shift. It describes Opus 5.5 and GPT-6 Sol as strong for the money, Astra as good at operating a computer, and GPT-6 Luna as a way to shrink the cost of supporting tasks.
The harness changes are the more instructive part. Newer models can plan several steps at once, so Devin now bundles routine work such as formatting code, linting it, and running tests into one request instead of three. In Cognition’s own illustrative example, that takes a four-turn exchange down to two and sends 49 percent fewer tokens. The company labels the numbers rounded examples, not measurements from customer sessions.
Caching is the second lever. A coding agent resends its whole history with every step, and providers charge less for text they have already processed. In another illustrative session, Cognition shows 71 percent fewer tokens processed from scratch. It also says Opus 5.5 cache reads cost 60 percent less than Opus 5, and that GPT-6 Sol and Luna cut cache-read prices in half against their predecessors. Part of Devin’s discount is therefore a pass-through of price cuts made by the model makers.
That matters for how durable the discount is. A price drop built on cheaper cached tokens depends on providers holding their rates, and it rewards agents that keep their opening instructions stable so the cache keeps hitting. It also puts Cognition in a squeeze with rivals such as Claude Code and Codex, which are sold by the same labs whose models Devin relies on. Competing on the price of the wrapper is easier when the raw ingredients are falling in cost.
One gap stands out: the post gives percentage cuts but no base prices, so buyers cannot tell what a typical Devin task costs now. Teams running Devin at volume should compare their last invoice against the new rates and rerun a sample of their own tasks, since a benchmark average says little about a specific codebase.
Reported by Cognition on Devin’s blog on 28 September 2026.
