Next.js AI Model Performance Evaluations (nextjs.org)

A broad benchmark of AI models and agent stacks was run on Next.js code generation and migration tasks, measuring success rate, average execution time, token consumption, and qualitative improvements. Among 50 evals per model, gpt-5-codex led on raw success (42%) with relatively high latency (42.8s avg) and heavy token use (186k total), followed closely by claude-opus-4.1 and glm-4.6 (40% success). Many generalist and specialized code models clustered in the 30–40% success range, while several faster models (kimi-k2-0905 at 1.82s, qwen3-coder at 0.78s, gpt-oss-120b at 1.39s) achieved competitive success with dramatically lower latency and token costs. Conversely, gemini-2.5-pro consumed the most tokens (322k) and one of the longest runtimes (~51s) for only 36% success. Agent-level runs showed the claude agent outperforming others at 42% success, with cursor and codex agents around 30% and gemini at 28%, highlighting that orchestration and tool use matter as much as base model selection. The study underscores stark trade-offs between accuracy, latency, and cost: top success doesn’t always mean practical deployment viability. For ML practitioners, these results argue for workload-specific model selection (fast, lean models for iterative dev workflows; higher-capacity models when quality justifies cost), standardized benchmarks for code migration tasks, and continued tuning of agent strategies to raise real-world success rates.