Pixel Canary: a free model that ties GPT-6 Astra on Next.js code
Frontier-class coding for $0 on Vercel's AI Gateway — but your prompts train it. Plus GPT-6 reasoning controls in the AI SDK and a Claude physics record.

Copy markdown
Free, and it ties the Next.js leaderboard
Pixel Canary is a stealth coding model now live on Vercel's AI Gateway, free while it's in stealth. On Next.js evals it matches GPT-6 Astra at 90.3% (28/31 tasks) out of the box and hits 96.8% (30/31) with docs context — a tie for the leaderboard's top score. It carries a 262K-token context, up to 131K output tokens, vision input, tool use, and reasoning effort you can dial from none to xhigh. Call it as `stealth/pixel-canary` through the AI SDK or the OpenAI/Anthropic-compatible endpoints.
The catch: it trains on your prompts
The price is your data. Prompts and outputs sent to Pixel Canary may be retained for training and model improvement, and zero-data-retention routing isn't available for it. That makes it great for side projects, evals, and throwaway prototypes this weekend — but keep proprietary or client code off it until it graduates from stealth.
GPT-6 reasoning controls land in the AI SDK
Vercel's AI SDK shipped reasoningEffort support for GPT-6 Sol and Luna, validated against each model's capabilities, and migrated the Perplexity provider off Sonar chat completions to the new Agent API with fresh presets and tool handling. Net effect: dial reasoning per call, and pick up Perplexity's agentic search without rewiring your provider setup.
Elsewhere: Claude sets a nine-loop physics record
Anthropic says Claude computed the six-particle scattering amplitude in planar N=4 super-Yang-Mills at nine loops — past the eight-loop human record from 2023 — for roughly $1,000-2,000 in usage fees. Stanford's Lance Dixon spent two weeks verifying all 107,053 coefficients via two independent methods. A concrete marker of what a frontier model plus a tight problem spec can now grind through.