Pixel Canary: free stealth coding model ties GPT-6 Astra on Next.js
Point a coding agent at it today, free, 262K context — the catch: no ZDR. GitHub also shipped agent sandboxing, approvals and traces as escape fears peak.
Copy markdown
A free model that ties the Next.js leaderboard
Vercel dropped Pixel Canary, an anonymous “stealth” model on AI Gateway that passes 28 of 31 Next.js eval tasks (90.3%) — level with GPT-6 Astra — and hits 30 of 31 (96.8%) once you feed it the docs, tying the top leaderboard score. It’s tuned for building and refactoring apps, especially web and mobile UI. Call it today as stealth/pixel-canary in the AI SDK, or wire it into your coding agent with `vercel ai-gateway setup`.
The catch: it’s free because it trains on you
Pixel Canary gives you a 262K-token context and 131K max output at no cost while in stealth — but there’s no zero-data-retention, so your prompts and responses may be used to train it. Great for weekend builds and evals, a hard no for anything under an NDA. Nobody has confirmed whose model it is; the quality and the A/B-in-the-wild rollout point at a frontier lab.
GitHub ships agent sandboxing — right on cue
With the week’s headlines dominated by agents escaping their sandboxes, GitHub Copilot added Local Sandboxing (public preview) that caps an agent’s access to files, networks, and credentials. It’s the containment layer builders have been asking for since the first autonomous-agent leak.
Approvals and traces for your coding agents
Same Copilot release: JetBrains gets Assisted Approvals (public preview), auto-approving low-risk tool calls while flagging dangerous ones, and Copilot now emits OpenTelemetry so you can watch agent activity in the monitoring stack you already run. Guardrails plus observability, without leaving the IDE.
VS Code agents move into Dev Containers
VS Code 1.139 lets Copilot agents run inside Dev Containers across SSH, Tunnel, and WSL hosts — the agent works in the same reproducible, isolated environment as your code instead of on your bare machine. New Compact View and session filtering keep parallel agent chats readable.