Intel / article

Pixel Canary is now available in stealth for free on AI Gateway

Source: vercel.com
Author: Zachary Chen
Date: Published Sep 25, 2026

From the Daily Brief

Vercel added Pixel Canary to AI Gateway as a stealth model, meaning the developer is not disclosed, and is offering it free for now. According to Vercel, it ties GPT-6 Astra on the company's Next.js benchmarks and reaches a 96.8% pass rate when the model is given AGENTS.md documentation for the project. That second figure is the more useful one: it measures the model with repository context rather than cold, which is closer to how coding agents run in practice. Both scores come from Vercel's own evaluations on a single framework, so they say little about performance on other stacks. Stealth releases on gateways are typically a pre-launch test, so the free period is a low-cost window to run the model against a team's own tasks before its origin and pricing are announced.

Read the 2026-09-26 Brief →

Context

Pixel Canary is a coding model released in "stealth" on Vercel's AI Gateway, and Vercel's release note reports a Next.js score for it. Vercel says Pixel Canary passed 28 of 31 tasks on its Next.js (90.3%) on its own and 30 of 31 (96.8%) when Next.js documentation was supplied through an AGENTS.md file, a file of instructions that coding agents read. Scores use pass@4: a task counts as passed if any of up to four attempts succeeds.

The 96.8% figure therefore describes coverage across a small task set under a four-attempt allowance. It does not tell a developer the probability that a single attempt will succeed. The comparison also shows that the information supplied to the coding was part of the evaluated setup, and the results are Vercel's own report.

Key takeaways · AI-distilled

Vercel's Next.js eval scores use pass@4, so a task counts as passed if any of up to four attempts succeeds; Pixel Canary's 28/31 baseline and 30/31 with AGENTS.md docs are on that basis.
Zero data retention is not available for the stealth model, and Vercel says prompts and responses sent to it may be used for training and model improvement.
Running `vercel ai-gateway setup` detects installed coding agents, provisions or reuses a Gateway API key, and wires them to the gateway so stealth/pixel-canary can be selected.

Terms in this piece · Glossary

benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source vercel.com