Cognition's annualized revenue doubled from $492 million to $1 billion in four months1, just as Nvidia published research showing coding-agent token consumption can be cut nearly in half without sacrificing performance3. ANALYSIS Taken together, the two developments mark a turning point for the AI coding-agent market: demand is scaling faster than almost any precedent in enterprise software, and the cost structure underpinning that demand is about to compress.
Why it matters
The AI coding-agent category has moved from demo-ware to genuine enterprise revenue faster than any adjacent AI vertical. Cognition's trajectory, from $492 million in annualized revenue in May to more than $900 million by early September and $1 billion this month, is the clearest proof point yet. But revenue growth alone does not guarantee durable margins. Coding agents consume long chains of reasoning, tool calls, and feedback loops that balloon token usage the longer they run unsupervised. ◆ The sustainability of billion-dollar run rates depends on whether the cost per task falls fast enough to keep enterprise buyers renewing.
The big picture
Three forces are converging on the developer-tools stack simultaneously.
First, Cognition is consolidating distribution. The company owns both the autonomous agent Devin and the Windsurf IDE, and Bloomberg reported on September 25 that the combined business is now generating revenue at an annualized pace of $1 billion. Cognition said enterprise usage of Devin had grown roughly 50 percent month over month for six straight months as of May. Internally, the company said 89 percent of code committed by its engineers was committed by Devin.
Second, competition is intensifying on price. Meta's Muse Code exited beta inside a month with paid tiers starting at $5 a month, undercutting both Claude Code and Cursor by a wide margin2. Cursor's Pro plan sits at $20 a month, with Business seats at $40 per user per month. Anthropic, meanwhile, priced Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, while discounting cache reads to $0.20 per million tokens. ANALYSIS The pricing spread across these three products suggests that model providers are willing to subsidize adoption to lock in developer workflows, a dynamic that compresses the revenue opportunity for any single agent vendor over time.
Third, infrastructure research is attacking the cost problem from below the model layer. Nvidia's SoL-Pi system optimizes the harness, the control layer between the model and its environment used by systems like Codex and Claude Code. A research AI analyzes agent traces, proposes harness changes, and keeps only those that maintain performance while cutting costs. The result: token savings of 50 percent compared to Codex and 54.3 percent compared to Claude Code on EdgeBench.
ANALYSIS Cognition's revenue milestone and Nvidia's harness optimization target different layers of the same stack, but they share a common implication: the coding-agent market is entering a phase where growth depends less on model capability gains and more on operational efficiency. Cognition's doubling happened while the competitive field added a major new entrant in Muse Code and while Anthropic expanded Claude Code's context window ceiling to as much as one million tokens. The fact that Cognition still accelerated through that competitive pressure points to the stickiness of enterprise integrations once an autonomous agent is embedded in production workflows.
Nvidia's contribution is subtler but structurally important. Most efficiency work to date has focused on cutting the cost per token through faster attention kernels, model compression, or swapping in cheaper models. SoL-Pi instead reduces the number of tokens consumed per task by rewriting how the agent sees states, runs actions, and processes feedback. The system explored 152 directions to find leaner control logic. ◆ If harness-level optimization becomes standard practice, it would lower the floor on per-task costs for every agent vendor, not just those running Nvidia's infrastructure, because the technique operates on the orchestration layer rather than the model itself.
What's next
Cognition raised more than $2 billion in September at a $48 billion valuation, up from $10.2 billion eight months prior. ◆ That valuation trajectory prices in continued revenue acceleration, which means the next quarter's retention and expansion numbers will be the real test of whether enterprise adoption is durable or front-loaded. On the cost side, Nvidia's SoL-Pi remains a research result, not a shipped product. The gap between a paper and production deployment will determine how quickly harness optimization reaches the agents enterprises are already paying for.