Pixel Canary: The Free Stealth Model Beating Expensive Models

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Pixel Canary: The Free Stealth Model Beating Expensive Models · Channel: Fahd Mirza · 08:52 · Watch on YouTube

Topics: AI agents, AI evaluation and benchmarks, Software engineering with AI, AI models

Summary

Fahd Mirza tests Pixel Canary, a free stealth model with no publicly known creator or paper, on a 3D suit configurator and an Oracle SQL debugging task. The model takes over an hour to complete the first task, and the resulting suit interface works in several ways but has weak-looking rendering and controls. In the SQL test, it identifies a missing region ID in a self-join, explains the impact, and follows the instruction to provide only the fix. The presenter says Pixel Canary’s combination of benchmark performance and zero price could make it useful for developers, while noting that its provider and future availability are unknown.

Key points

Pixel Canary is presented as a free stealth model, with no publicly known owner or research paper.
The presenter cites a 90% score on Next.js Agent Eval, placing it near GPT-6 Astra and seven points behind Claude Fable 5.1.
The first task asks the model to build a 3D suit configurator using a Blender-generated GLB and a 3.js front end.
The model took 1 hour and 4 minutes to complete the suit task.
The configurator’s color controls, scrolling, and mouse rotation worked, but the presenter criticized the suit’s rendering and some controls.
In the Oracle SQL test, the model found a missing region ID in a self-join and identified downstream rank contamination.
The model followed the instruction to return only the fixed chunk rather than rewriting the full query.
The presenter highlights the model’s reported zero cost as a potential advantage for side projects, prototypes, and agentic workloads.
The presenter suggests stealth releases may help providers gather real-world usage data or build a reputation before a public launch.

Chapters

Start	Chapter	Summary
00:01	Introducing Pixel Canary	The presenter introduces the free, anonymously developed model and describes the benchmark comparison before setting up a full 3D suit configurator task.
02:05	Testing the 3D suit configurator	After a lengthy run, the presenter tests the generated suit interface and finds that some controls work, though the rendering and details are underwhelming.
05:12	Oracle SQL debugging test	Pixel Canary locates a missing region ID in a self-join, explains the resulting errors, and obeys the request to return only the fix.
07:17	Cost and possible reasons for stealth releases	The presenter emphasizes Pixel Canary’s zero price and discusses potential benefits of free models and possible reasons providers release models quietly.

Quotes

Nobody knows who built it. — 00:01

the suit rendering was better in the other one — 04:40

It found the real bug, which was missing region ID in the self-join. — 06:14

Pixel is sitting right between them on performance at absolutely zero cost. — 07:17

to get the real-world usage data and stress test performance before committing to a public launch — 08:20

Concepts and entities

Pixel Canary, Next.js Agent Eval, Hermes Agent, Blender, Three.js, GPT-6 Astra, Claude Fable 5.1, Mask Compute, Fahd Mirza

Full transcript (original language)

00:01 Stealth models are becoming a daily norm now. Today, we are dealing with Pixel Canary, which is available for free at few of the platforms. Nobody knows who built it. It has no public paper, no or behind it. But the name Pixel matches with Google’s Pixel phone, but you know what? No one knows who owns this model, but look at this benchmark. It’s sitting at 90% on Next.js Agent Eval. Right next to GPT-6 Astra and just seven points

00:32 behind Claude Fable 5.1, which costs 72 % million tokens, the most expensive model out there. Pixel Canary, it is completely free at the moment. No one knows what happens in the future. So, what I’m going to do, I’m going to put it through the paces through the first test, where we are going all in. I already have configured it with Hermes Agent, and this is the prompt which I’m going to use. No app to fix at the

01:03 moment. No bug to hunt. We are building from scratch a full 3D bespoke suit configuration. Blender models the suit headless using a Python script, exports it as a GLB, and then a 3.js front end loads it in the browser with live fabric, lapel, and button controls. The model has to do all of it end-to-end. Start the server itself and verify it works before it’s done. One heads-up,

01:34 in my experience so far, this model is quite slow slow. So, anyway, we will see how it goes. It has started thinking. If you want to guess which model is this and who is the provider, just let me know in the comments and we will see who is right at the end of the day. Also, recently, I have started this free weekly AI newsletter, which you can subscribe at fahadmirza.substack.com. And if you want to help out the channel, please become a member.

02:05 After more than an hour, the model has come back with the response, and it says that it has done the task. Before I go and run this in the browser in front of you, I haven’t tested it yet. What it is telling us in terms of features, it has covered it all. And if you look at the thinking, it has thought a lot. Eerily reminds me of CLO 3D point 8. Plus, um, some of the thinking streaks are very eerily similar to CLO’s model, but I’m I

02:35 don’t know which model and provider that is. And you can see that it has thought a lot. And the reason why I show you this raw data is that because this is what shows you, um, the difference with other models. Let me now run this. What it is telling me to run. So, you can see it is loading at the moment. The turntable is also turning, which is good. Auto rotate is on. So, I will just let it auto rotate. The suit texture is not good, as you can already see.

03:07 So, I’m just going to first, just click on maybe peak. The shadow looks good, by the way. And takes bit of a time to come through, it seems. And you see the button, can you spot them? Well, it could have been better. Okay, now let’s maybe turn it to red, which will be visible. Yeah, the color texture is there. Green. Orange. The fabric color is working fine. How about this suit?

03:38 Can I increase the buttons live? Let’s do three. Where’s the three? This is the three, I don’t know. Okay, let’s reset the camera. I have reset the camera. Now it is moving through. The one, two, three buttons. Is there any fourth button? But, could have been much better, right? It says scroll to zoom, so I’m just going to scroll now with my mouse. Yep, scroll works. And can I rotate it like this? Yes, I can with my mouse.

04:09 At least this part is better than the yesterday’s stealth model, which was uh which one was that? I actually forgot the name of. Uh we did it yesterday. Anyway, I’ll check it later. Something like bunny something uh space bunny. Anyway, so let’s go back to nosh. Not much is happening with this label style, but maybe it is quite subtle because this rendering is not that good in three.js.

04:40 Anyway, so this is what it is. I’m just going to do the 360 zoom to do the orbit. There you go. Okay, but there is the suit rendering was better in the other one, but this one I don’t think so it is even a suit. Okay, let’s do another test because it’s a coding model, so I’m just going to go with the coding stuff. And by the way, as I cancelled that session, this is how long it took, 1 hour 4 minutes.

05:12 Okay, so the next test is quite a brutal one. It is a complex Oracle sales report, materialized views, CTEs, Windows functions, year-on-year growth, quota tracking, all-in-one query. As you can see, I’m creating a materialized view. It runs clean. Oh, by the way, no errors, but returns wrong numbers. And we want to see if Pixel Canary can find exactly where and why there is a subtle bug in this query. And I have already asked this model to go and find

05:43 that bug and only give me that fixed chunk. Not only I’m just looking at if it is able to fix it, but also if if it follows the instruction or not. It must not rewrite the query the whole query. And by the way, if you’re looking to rent a GPU on very good price, you can find the link to Mask Compute in videos description. And this time, the model has come back with a genuine impressive response. It found the real bug, which was missing

06:14 region ID in the self-join. There was also a decoy which I planted. It reproduced it with actual data, showed exact wrong numbers before and after the fix, and even flagged the downstream rank contamination. I think it’s a solid analytical performance, just like a DBA who would be doing a SQL debugging with performance optimization. And the thinking is quite good. It even imported SQLite just to see if it could test out the query by converting it to

06:45 SQLite syntax, while totally understanding it was a SQL Oracle SQL report. In addition to that, it didn’t really um rewrite anything. It just fixed the bug, and that’s about it. So, instruction following is also quite good, which really we want for our coding models. One thing I really do want to say out loud is that what makes this Pixel Canary interesting is not just the benchmark number. It’s the price. As you

07:17 can see on this chart, Claude v1.5.1 sits at 72 cents per million tokens, GPT-6 Astra from OpenAI at 89 cents, and Pixel is sitting right between them on performance at uh absolutely zero cost. And And if it is an open model, um the cost will stay zero in addition to our infra cost, and even in contest format. For developers, that changes the calculations entirely. You are getting frontier level coding capability for

07:48 your side projects, prototypes, and agentic workloads without touching your budget. And I can tell you from experience that when, you know, I don’t use Fable. Not just because I don’t like it. It is really good model, but I simply cannot afford it. Um same goes for GPT-6 extra for the major production uh projects. And even, you know, making these videos while I’m evaluating, I really um cannot use this model due to the affordability. That is where these open

08:20 weight models come very handy, and especially these free models, even if they are through API based. One final point is that some of the comments are asking that why would they release a stealth model? So, I think maybe to get the real-world usage data and stress test performance before committing to a public launch, just like we did here. Um and maybe sometimes to build reputation quietly before revealing who is behind it. And remember, there are also model providers platforms who are trying to get into the game, and

08:52 they are providing it for free. Uh just like I’m using it from one of the model platform. And I’m not doing free marketing for anyone. So, but you can see that that is uh another angle of doing these stealth models. Anyway, that’s it. Let me know what do you think. Please become a member and share the channel among your network. Thank you for all the support.