Valuation acceleration and revenue growth
Cognition reached a $26B valuation in May 2026. This figure sits on $492M of run-rate revenue, creating a 53x revenue multiple. This multiple stays far above the 29x revenue multiple reported for Cursor and significantly exceeds public SaaS companies like Cloudflare at 30.5x or CrowdStrike at 21.7x. Cognition valuation grew through several checkpoints, starting at $350M in early 2024, rising to $2B in April 2024, nearly $4B in March 2025, and $10.2B in September 2025. Between April 2024 and March 2025, the company added $186M in valuation per month. This figure jumped to $1.1B per month between March and September 2025, and then to $1.8B per month between September 2025 and May 2026. Devin alone grew from $1M ARR in September 2024 to $73M ARR in June 2025. After Cognition acquired Windsurf, which brought $82M in ARR, the combined enterprise ARR rose more than 30% in the following seven weeks. Enterprise usage grew more than 10x since the start of 2026. I believe the valuation is aggressive because it assumes Cognition converts engineering payroll into metered AI-agent work. The company counts Goldman Sachs, Citi, Dell, Cisco, and the U.S. Navy among its customers, and a pilot at AHEAD showed Devin delivered 8x to 40x time savings compared to manual baselines.
Task success rates and agent performance
Acceptance rates for AI agents change based on the specific task assigned. In a study of 7,156 pull requests, Devin showed the only consistent improvement, with acceptance rates rising by 0.77% per week over 32 weeks. GitHub Copilot scored 56.00% on the SWE-bench accuracy metric, while Cursor scored 51.70%. OpenAI Codex achieved high acceptance rates across nine categories between 59.6% and 88.6%. Claude Code leads in documentation tasks with 92.3% acceptance and feature tasks with 72.6% acceptance. Cursor excels in fix tasks at 80.4% acceptance. Documentation tasks achieve 82.1% acceptance compared to 66.1% for new features. This 16 percentage point gap between task types exceeds typical inter-agent variance. The data shows that task distribution matters more than the specific agent selected for a coding job. One tester found that Devin sometimes makes unnecessary changes when it assumes a library is used in a project. Furthermore, Devin recommended separating a large task into different sessions, which caused problems because the sessions did not share the same context. The disparity between Claude Code and Cursor on test tasks reached 44.4 percentage points. Additionally, Cursor completes tasks approximately 30.00% faster than GitHub Copilot, with completion times of 62.95 seconds versus 89.91 seconds. Some research notes that velocity gains with AI tools can collapse back to baseline by month three. AHEAD engineers used Devin to build three APIs in a single day, a process that previously required four weeks. The research at NAV IT found no statistically significant changes in commit-based activity for Copilot users after they adopted the tool.
Credit consumption and technical limits
GitHub Copilot transitioned to a credit-based system on June 1, 2026. This change replaced flat-rate premium requests with token-based AI Credits. One Pro+ developer burned 8% of their monthly allotment in just two hours.
| Plan | Price | Credits |
|---|---|---|
| Copilot Pro | $10.00 | 1,500 |
| Copilot Pro+ | $39.00 | 7,000 |
| Copilot Max | $100.00 | 20,000 |
| Copilot Business | $19.00/seat | ~1,900 |
| Copilot Enterprise | $39.00/seat | ~3,900 |
Devin uses Agent Compute Units (ACUs) for its resource consumption. Users can purchase additional ACUs for $2 per unit. One tester exhausted 150 ACUs in less than a week. Performance degrades when a conversation consumes 10 ACUs. Devin also only checks CI, not build errors, which caused failures in local Docker environments. A single complex change request for a Copilot user cost $6. You should remember that the model knows the token counts in advance while the user does not. Models like Claude Opus 4.7 show a quality cliff where retrieval drops to 32% at 1M tokens. NVIDIA’s RULER study showed nearly every model performs below the Llama2-7B baseline at its claimed length. One researcher observed that experienced developers were 19% slower when using Copilot despite believing otherwise. Does the developer need to spend more time fixing Devin’s errors than writing code?
