Vercel AI Updates: September 26, 2026

1. Stealth Coding Model Pixel Canary Is Free on AI Gateway

Vercel added stealth/pixel-canary to AI Gateway, an unnamed model aimed at coding, refactoring, and frontend and mobile app work such as responsive layouts, navigation, and interactive components. Vercel reports a 90.3% baseline success rate on its Next.js evals, rising to 96.8% when the model is given documentation. It is free for a limited time, but zero data retention is not available and prompts and responses may be used for training. Source

2. The skills.sh Registry Hit One Million Agent Skills in Seven Months

Vercel published a State of Agent Skills report showing the skills.sh registry reached one million skills and nearly 280 million installs in seven months. Software engineering makes up over 25% of listings but only 18% of installs, with agent workflows at 15% and business operations and writing at about 11% each. Installs are heavily concentrated, as 375 skills (0.04% of the registry) account for 62% of installs, and cross-industry skills get 3.6 times more installs per listing than industry-specific ones. Source

3. AI SDK Perplexity Provider 5.0 Moves From Sonar Chat Completions to the Agent API

Vercel released @ai-sdk/perplexity@5.0.0, a breaking change that migrates language generation from Perplexity’s Sonar Chat Completions API to its Agent API. Model identifiers, configuration options, request and response shapes, streaming events, and usage and cost reporting all change, and Sonar PDF input plus image and video results are dropped. In exchange the provider gains native tool traces, preserved URL citations, deduplicated sources, and better recovery of text from incomplete responses. Source

4. Vercel Sandbox Now Reports Memory Usage

Vercel added memory observability to Sandbox alongside existing CPU and data transfer metrics, with a dashboard card showing average, P75, and P95 memory and detail charts scaled to each sandbox’s limit with an 85% reference line. The data is also available from the CLI via vercel metrics under vercel.sandbox.memory_used_bytes and in the Observability query builder as memoryUsedBytes for custom queries and alerts. For agent workloads running in sandboxes, this makes it possible to catch runs approaching out-of-memory before they fail. Source