Anonymous Model Pixel Canary Launches Free, Matches GPT-6 Astra on Coding Benchmark but Is Extremely Slow

Cline announced on X that the anonymous model Pixel Canary is now free on its platform, but what really sparked discussion was its result on Next.js Agent Evals: under the standard setup it completed 28 of 31 tasks, a 90.3% success rate, tied with GPT-6 Astra (high reasoning effort) and above Kimi K3's 84%.

But the same leaderboard gives another set of numbers: Pixel Canary's average task completion time is 1015.80 seconds, about 16 minutes 56 seconds; GPT-6 Astra is 251.89 seconds, and Kimi K3 is 352.34 seconds. Put the success rate and the time taken side by side, and it shows the ability to complete complex tasks, but has not yet proven itself suited to fast-paced daily development.

There is also an easily overlooked reading issue here. The evaluation uses pass@4: each task gets up to four attempts, and as long as one succeeds, the task counts as passed. So 90.3% cannot be directly understood as: hand it ten tasks and it completes nine on the first try.

Pixel Canary first landed on Vercel AI Gateway on September 25, under the model identifier stealth/pixel-canary, free for a limited time during the anonymous open period. Vercel positions it for app building, refactoring existing code, front-end development and mobile app design.

The model information published by the platform shows it supports a context window of about 262,000 tokens and adjustable reasoning effort, but does not disclose the development team. In other words, the capability parameters are public; the identity is not.

The evaluation itself examines how agents handle real Next.js development tasks, covering App Router migration, data fetching, image and font optimization, caching and page transitions. This defines the scope of the 90.3%: it describes performance on this class of tasks, not general capability across a broader range of tasks.

When the evaluation provided Next.js documentation to the agent via AGENTS.md, Pixel Canary completed 30 tasks, and its success rate rose to 96.8%.

Under this setup, however, models such as Kimi K3 and GPT-6 Astra also reached about 97%. This shows that task-relevant documentation can significantly change results, and also reminds developers: code conventions, framework versions and project documentation in real projects may matter as much as model choice.

Looked at another way, the gap of 90.3% versus 84% under the standard setup is essentially erased once documentation is added. Pixel Canary's relative advantage seems to appear more in scenarios with insufficient information, rather than existing stably under all conditions.

Vercel explicitly states in its release notes that Pixel Canary does not offer zero data retention (ZDR), and prompts and responses sent to the model may be used for training and model improvement; its model page also notes that the provider may retain inputs and outputs.

For developers connecting a code repository to an agent, this condition is worth checking before the word "free": code, configuration files or project context may all enter model requests along with the task.

What needs to be distinguished is that this is a condition publicly stated by the platform, a verifiable release note; how long the anonymous open period will last and how pricing will work afterward are currently not public information.

After launch, netizens quickly split into two camps: "testing capability" and "guessing identity." In the capability-testing camp, speed and stability became the focus of controversy. X user codemeoww said that when using Pixel Canary, they repeatedly encountered tasks suddenly interrupting halfway through, eventually stopped using it, and believed the model was too slow for many people to be willing to use long term.

Another X user said Pixel Canary's pelican test ran for 5 hours, with various errors along the way and output speed of only single-digit tps. Netizen VulKan said they gave up testing after waiting 8 hours, arguing that if the service cannot be provided stably, even a high benchmark score is hard to translate into real value.

These feedbacks come from different users' own access conditions and cannot be treated as a uniform performance metric for the model. But together with the nearly 17-minute average task time on the leaderboard, they point to the same practical issue: being able to finish and being able to finish efficiently are two different things.

Whether the latency and instability stem from the model itself or from service resources at the time is still undetermined. This is also why these individual experiences can only serve as reference, and cannot replace uniform performance measurement.

Netizens guessed based on Thai and Hindi text segmentation that Pixel Canary is related to Qwen 3.8, while others guessed from the name that it comes from Google or is Kimi K3.1, but none of these have been confirmed by official information.

Such guessing is not new. Previously in August, Ox Alpha was made available anonymously for trial and was later revealed by Zhipu as GLM-5.3-Flash; in September, Union Alpha appeared on OpenRouter, and anonymous launches have become a recurring pattern.

Anonymous releases have their rationale: temporarily removing the vendor name lets developers test hands-on first, less influenced by brand impressions. But after the same playbook appears frequently, attention can easily shift from "what problem can this model solve" to "who is creating suspense this time." There is currently no evidence that Pixel Canary's publisher deliberately hyped it, but this skepticism deserves to be taken seriously.

Pixel Canary's most certain value right now is giving developers one more option to try. It delivered a competitive pass rate on specific Next.js tasks, improving further with documentation; its provider, long-term pricing and performance on broader tasks still await public information and more real-world testing.

If you want to judge whether it can become a daily workhorse, the next round worth watching is not another guess at "who it is," but under the same project, same tools, same tasks and same time limits, looking at the quality of code it delivers and how long developers waited for it.

As anonymous models become more common, the mystery will become shorter-lived. To keep users after the identity is revealed, vendors still have to answer more plain questions: whether the model is stable, whether the speed is acceptable, and whether performance on different tasks can be repeated.