Vercel has released Pixel Canary, a new coding-focused model, free on its AI Gateway under the identifier stealth/pixel-canary. On the Next.js evals benchmark, it scores 96.8% when provided Next.js documentation via AGENTS.md, passing 30 of 31 tasks and matching the leaderboard's top score. At baseline, it ties GPT 6 Astra (high) at 90.3%, passing 28 of 31 tasks using a pass@4 scoring method.
The model targets frontend and mobile development: App Router migrations, data fetching, image and font optimization, caching, and view transitions. It runs through Vercel's AI Gateway, which handles unified API access, cost tracking, API key budgets, and routing rules. Zero Data Retention is not available, and prompts may be used for training.
The benchmark methodology and the gap between baseline and AGENTS.md-assisted performance are worth reading in full. A 6.5 percentage point lift from supplying documentation is a concrete data point about how context provisioning affects model output on real engineering tasks. Setup runs through the Vercel CLI, which auto-detects installed agents and provisions gateway credentials.
[READ ORIGINAL →]