Vercel puts Pixel Canary coding model on AI Gateway free during stealth

What Vercel shipped

Vercel has added a new AI model, Pixel Canary, to its AI Gateway under the identifier stealth/pixel-canary. According to the company's changelog, the model is free to use for a limited period while it remains in stealth.

Vercel positions it as a coding specialist. The announcement says the model is strong at building applications and refactoring existing code, with a particular lean toward frontend development and mobile app design — responsive layouts, app screens, navigation and interactive components are called out as its sweet spot.

Benchmark results

On Vercel's Next.js evaluation suite, Pixel Canary passed 28 of 31 tasks for a 90.3% baseline success rate, which Vercel says ties GPT 6 Astra running at its high setting. When Next.js documentation was supplied to the model through an AGENTS.md file, it passed 30 of 31 tasks (96.8%), a score Vercel says matches the leaderboard's top result in that configuration.

The tasks in the suite cover App Router migrations, data fetching, image and font optimization, caching, and view transitions — work that maps closely to maintaining real Next.js applications, which Vercel cites as evidence the model can implement and update the code behind web interfaces.

One methodological detail worth noting: Vercel says the scores use pass@4, meaning a task counts as passed if any of up to four attempts succeeds. That is a more permissive measure than single-attempt scoring, so the headline percentages shouldn't be compared directly against pass@1 results from other benchmarks.

The privacy tradeoff

Zero Data Retention is not available for this model. Vercel states that prompts and responses sent through it may be used for training and model improvement. That is the main caveat attached to the free access: code and instructions you route through Pixel Canary could end up feeding its development. Teams working with proprietary or client code should weigh that against alternatives that offer stronger data guarantees before pointing their agents at it.

How to try it

Access runs through the AI Gateway rather than a dedicated endpoint. Vercel's setup flow starts with installing the latest CLI globally via npm i -g vercel@latest, then running vercel ai-gateway setup. According to the changelog, that command detects which coding agents are installed, provisions or reuses an AI Gateway API key, and configures the agents to connect through the gateway. From there, you select stealth/pixel-canary in your agent's model configuration. Vercel also offers a model playground for trying it without wiring up an agent.

The gateway itself provides a unified API for calling models, with built-in usage and cost tracking, budgets for API keys, and routing rules — so Pixel Canary sits alongside other models behind a single interface rather than as a standalone product.

Why it matters

A model that ties GPT 6 Astra (high) on a real-world web framework eval, at zero cost during stealth, gives developers a low-friction way to test a challenger in their existing tooling. The distribution model is also notable: instead of shipping yet another chat interface, Vercel is surfacing the model as a selectable option inside coding agents people already use, with billing and routing handled centrally by the gateway.

Two caveats temper the enthusiasm. First, the benchmark numbers are Vercel's own, run on Vercel's eval suite for a model Vercel is distributing — self-reported figures deserve independent verification. Second, the absence of zero data retention means the free tier is effectively paid for in training data. For non-sensitive work and quick experiments, though, it is an easy model to spin up and judge for yourself.