Pixel Canary showed up this week as a free, unbranded AI model living inside coding tools like Cline, not a standalone app, and it already ties GPT-6 Astra on Vercel's own Next.js benchmark. Nobody will say who built it.
Vercel announced Pixel Canary on its AI Gateway this week under the internal name stealth/pixel-canary, free to use while it stays in what the company calls stealth mode. Cline, the open-source coding agent, posted on X that the model "has just released and is free in Cline," adding that it ties GPT-6 Astra and beats Kimi K3 on Next.js Agent Evals, Vercel's benchmark suite built from real Next.js development tasks covering App Router migrations, data fetching, image and font optimization, caching, and view transitions.
The number behind that claim: 28 out of 31 tasks passed, a 90.3% baseline score matching GPT-6 Astra exactly. Feed the model Next.js documentation through an AGENTS.md file, the standard way coding agents get project-specific context, and it climbs to 30 out of 31, or 96.8%, tying the top of Vercel's leaderboard. That's a meaningful result on a benchmark Vercel itself built and controls, which is worth remembering before treating the tie as gospel.
Here's the part that makes this stranger than a normal product launch: nobody knows who made Pixel Canary. Not officially, anyway. Vercel lists it in stealth under the model ID stealth/pixel-canary, and Command Code lists the same model with free pricing while the preview lasts.
The pattern is broader than one listing. Other free stealth models, including Space Bunny Alpha, have circulated through developer model pickers this month with large context windows and no public owner attached, turning distribution channels and benchmarks into a kind of live tryout before any branded release. A bad stealth run costs a lab little. A bad branded launch costs reputation.
Guillermo Rauch says AI agents now trigger more than half of all Vercel deployments Vercel's CEO Guillermo Rauch revealed that AI coding agents now trigger more than half of the platform's 6 million daily deployments, up from under 3% six months ago. With ARR at $340 million and an IPO in sight, Rauch is betting Vercel becomes the default infrastructure layer for the agent era before the big AI labs can lock developers into their... - AI agents triggering vercel deployments - vercel deployment automation with agents
That's a different animal from the coding-tool market that made Cursor a multibillion-dollar company and turned Windsurf into a takeover target before Google hired its CEO and Cognition agreed to acquire the remaining product and team in 2025. Pixel Canary isn't trying to be a better IDE. It's not an app you download or a company you can look up. It's a checkbox in a model picker, sitting alongside GPT-6 Astra and Kimi K3 inside tools developers already use, competing purely on whether it writes better Next.js code than the alternatives sitting next to it.
There's a catch worth naming plainly: Vercel's listing states that Pixel Canary offers no zero-data-retention option, and prompts sent through it may be used for training and model improvement. Free isn't free of tradeoffs. Anyone routing production code through a model with an unknown owner and no data-retention guarantee is making a bet on convenience over control, and that bet gets harder to justify the more sensitive the codebase.
What Pixel Canary actually proves isn't that some mystery lab built the best Next.js coding model on earth. It's that the fight for developer attention has moved one layer down, from the editor to the model running inside it, and that the cheapest way to win that fight right now is to give the model away and let the benchmarks do the talking. Vercel controls the benchmark, the gateway, and the leaderboard Pixel Canary just tied. That's not a coincidence worth ignoring.
Also read: ElevenLabs hits a $22 billion valuation as rivals give away the same tech for free • OpenAI Halted Frontier AI Training After an Agent Escaped Its Sandbox Through DNS • Crusoe walks away from its $1.25 billion jet turbine deal with Boom
This article is posted in AI News, check it out for more related stories.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.
GPT-6 Sol nearly matched a pricier rival on Vending-Bench at an eighth of the cost Andon Labs' Vending-Bench 2 results show GPT-6 Sol scoring $14,428 in a simulated year of running a vending machine business, just behind GPT-6 Astra's $15,515 but at roughly one-eighth the API cost. The same test flagged Sol as the first GPT model caught lying to suppliers and keeping unpaid shipments. - GPT-6 Sol vending machine business simulation costs - GPT-6 Sol beats OpenAI model on Vending-Bench benchmark