AI coding is taking the lead again. Cognition has raised $2 billion at a $48 billion valuation, nearly doubling what investors believed the company was worth four months earlier. The company behind Devin also says its annualized revenue has grown from $492 million in May to $900 million, with customers including Nvidia, Goldman Sachs, Citi, Mercedes-Benz and GE Aerospace.
Those figures make it one of the most valuable private software companies built during the generative AI era. They also make its latest financing useful evidence of a larger change in software development. AI coding is becoming a category of infrastructure through which engineering work is understood, executed and verified.
The category now stretches well beyond code completion. This helps explain why investors continue to finance several large competitors at once. Engineering organizations vary widely in their codebases, workflows, security requirements and tolerance for autonomous action. The available evidence also shows meaningful differences between agents depending on the work they are asked to perform. A single winner has yet to emerge because the underlying engineering problem has yet to become uniform.
AI coding – a valuation for the system around the model
Cognition announced its latest financing on September 8. The $2 billion round was led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst and Avenir, four months after the company raised capital at a $26 billion valuation.
According to TechCrunch, Cognition leases an Nvidia server cluster costing hundreds of millions of dollars annually and is training its own model using open-source foundations. Its total cash burn could reach $800 million this year. The company expects its annualized revenue to reach between $4 billion and $5 billion by the end of 2026, although that forecast and its current run-rate figures remain company-reported.
The scale of the investment indicates how much value may accumulate above the foundation model. Devin is sold as an AI software engineer capable of completing tasks with access to a repository, shell, editor and browser. Its commercial value depends on the entire environment: how it gathers context, decomposes a request, chooses tools, recovers from failure and demonstrates that the completed work is correct.
Adoption has moved into real repositories
A large-scale study of coding-agent adoption examined more than 129,000 GitHub projects and estimated that agents were already present in 15.85% to 22.6% of them. Their use extended across programming languages, established organizations and projects at different levels of maturity. Agent-assisted commits were also substantial, covering features and bug fixes rather than only documentation or minor cleanup.
OpenAI said in September that Codex had reached 25 million users, according to Reuters. Cognition attributes its growth to adoption among large enterprises in finance, manufacturing, aerospace and technology. These are environments where access controls, legacy dependencies and review requirements place greater demands on an agent than a self-contained programming exercise.
The unit of adoption is changing with the product, and adopting such a system is therefore an architectural and organizational choice. The tool acquires a position inside the software delivery process, along with the context and authority required to do useful work.
AI coding performance depends on the work
Rapid adoption has not produced a clear technical hierarchy.
A study comparing Codex, GitHub Copilot, Devin, Cursor and Claude Code analyzed 7,156 pull requests created with coding agents. It found that no agent performed best across all nine categories of work. Claude Code led in documentation and feature tasks, Cursor performed strongly on fixes, and Codex recorded consistently high acceptance rates across categories.
Task type often mattered more than the difference between products. Documentation pull requests achieved an 82.1% acceptance rate, compared with 66.1% for new features. Devin was the only agent in the study to show a consistent improvement in acceptance over the 32-week observation period, but it did not lead every category.
The results make a winner-take-all outcome difficult to assume. A team maintaining a mature financial platform may value conservative edits, traceability and policy enforcement. A startup building a new product may prioritize speed, broad task completion and the ability to move across unfamiliar technologies. An infrastructure team may care most about terminal operation and debugging, while a large enterprise may require deployment inside a controlled environment.
Passing a task is only the beginning
Many of the benchmarks used to compare AI coding agents present a bounded issue against a repository at a fixed point in time. Production software presents a harder problem. Requirements change, dependencies move, architectural decisions accumulate and an apparently correct modification can create costs several releases later.
The authors of SWE-CI designed a benchmark around this gap. Its 100 tasks cover repository histories averaging 233 days and 71 consecutive commits. Agents must respond to repeated changes while preserving code quality, shifting evaluation from whether a system can solve one issue towards whether it can maintain software through continuing evolution.
Another study, SWE-Skills-Bench, tested whether adding structured procedural instructions improves agent performance on real software-engineering requirements. Thirty-nine of the 49 skills tested produced no increase in pass rate. Some added as much as 451% in token overhead without improving results, while three reduced performance because their guidance conflicted with the project context.
These findings place the central technical challenge of AI coding in context. Giving an agent more instructions or a more capable model does not automatically improve its work inside a particular repository. The system must retrieve the right information, distinguish current conventions from obsolete ones, understand the consequences of a change and use tests that measure what matters.
Verification becomes part of productivity
Code generation is easy to measure because output is visible and immediate. Verification is slower and distributed across reviews, tests, security checks and production behavior.
A 2026 Sonar survey, reported by ITPro, found that 96% of developers did not fully trust AI-generated code, while fewer than half consistently checked it before use. Thirty-eight percent said reviewing AI-generated code could take longer than reviewing code written by colleagues. The survey also found that AI coding was already being used extensively in prototypes, production systems and customer-facing applications.
This creates a productivity question that cannot be answered by measuring how quickly an agent produces a patch. An agent may reduce implementation time while increasing the burden on code review, testing or incident response. It may also improve overall delivery by writing tests, documenting its reasoning and catching problems a developer would have missed.
The relevant unit is the accepted, maintainable change. Products capable of proving their work through reproducible tests, legible diffs and clear execution records will have an advantage over those that simply generate more code. As autonomous tasks become longer, observability and the ability to interrupt or redirect an agent also become part of software quality.
Autonomy has a compute bill
AI coding agents also introduce an economic constraint that code-completion products could largely hide behind a subscription.
An autonomous task may require repeated repository searches, long reasoning traces, several implementation attempts and multiple test runs. Each iteration consumes tokens and infrastructure. Business Insider reported that OpenAI’s heaviest internal users were spending more than $7,000 per day on coding-agent tokens. Median daily expenditure per researcher reportedly increased from $162 in July to more than $600 by mid-August as agents enabled employees to run more experiments and pursue more tasks in parallel.
Cognition’s infrastructure commitments show the same pressure at company scale. Training proprietary models could reduce dependence on external providers, but the operation of long-running agents still requires considerable inference capacity.
A new layer in the development stack
Cognition’s valuation is evidence that AI coding has become a significant software market. It does not establish which company will lead it or whether current valuations can be sustained.
The more durable development is the emergence of a new layer between human intent and the codebase. Several companies can occupy this layer because the work remains heterogeneous.
The contest will be decided inside real engineering systems, where an agent’s output must survive tests, reviews, dependencies and subsequent changes.Cognition’s $48 billion valuation tells us how valuable investors believe that position could become. The harder evidence, from repository adoption to pull-request outcomes and maintenance benchmarks, tells us why the position is still open.
