Zhipu AI Unmasks Ox Alpha Model at 40-Fold Discount to Claude

Fireship

Zhipu AI's release of GLM-5.3-Flash demonstrates that open models running on domestic Chinese hardware can undercut Western competitors on price by 40-fold while capturing developer traffic despite data privacy concerns.

AI-assisted summary of The mystery is solved... and the answer is 40x cheaper than Claude — original source. Check the original for full context.

Source published: 2026-09-01

Key ideas

An anonymous model codenamed Ox Alpha dominated OpenRouter by serving 42 trillion tokens in six days with a 1 million token context window and native video support. (0:00)
Zhipu AI unmasked Ox Alpha as GLM-5.3-Flash, an open-weight 320B parameter MoE model released under an MIT license and trained on domestic Chinese accelerators. (2:01)
Commercial pricing for GLM-5.3-Flash was set at $0.15 per million input tokens and $0.50 per million output tokens, making it roughly 40 times cheaper than Claude during introductory discounts. (2:01)
In practical testing, GLM-5.3-Flash modernized an outdated 2016 AngularJS codebase into vanilla JavaScript and used native vision capabilities to process video frames via FFmpeg. (3:03)

Detailed notes

## Background Published by the tech channel Fireship on YouTube on September 1, 2026, the video covers the sudden rise of a mystery model dubbed 'Ox Alpha' that dominated OpenRouter. According to the description, the anonymous model served 42 trillion tokens across six days before being unmasked as Zhipu AI's GLM-5.3-Flash, offering frontier capabilities at a fraction of the cost of competitors.

## TL;DR A stealth AI model named Ox Alpha stunned developers by processing 42 trillion tokens for free on OpenRouter, sparking widespread industry panic before being revealed as Zhipu AI's 320-billion parameter open-weight model, GLM-5.3-Flash. Operating on domestic Chinese chips, the newly launched model delivers native video support and a 1-million token context window at rates roughly 40 times cheaper than Claude. To evaluate its real-world performance, Fireship puts GLM-5.3-Flash to work modernizing an abandoned 2016 AngularJS side project into vanilla web code and testing its native computer vision on raw video slices. The video concludes with an overview of Exa, an agent-focused search engine built to query structured databases.

## Atmosphere & style The video moves at Fireship's signature breakneck speed, brimming with cynical internet humor, rapid-fire editing, and deadpan satire aimed at Silicon Valley leadership. The narrative arc pivots smoothly from investigative tech drama into chaotic, hands-on software development as the host revives a bizarre legacy app called 'Horse 2.' It feels like an irreverent, sharp-witted dispatch from the bleeding edge of the AI race that never takes itself too seriously while delivering genuine technical breakdowns.

## Bulletin - [00:00](https://www.youtube.com/watch?v=r-tzcMlQISk&t=0s) An anonymous model codenamed Ox Alpha dominated OpenRouter by serving 42 trillion tokens in six days with a 1 million token context window and native video support. - [02:01](https://www.youtube.com/watch?v=r-tzcMlQISk&t=121s) Zhipu AI unmasked Ox Alpha as GLM-5.3-Flash, an open-weight 320B parameter MoE model released under an MIT license and trained on domestic Chinese accelerators. - [02:01](https://www.youtube.com/watch?v=r-tzcMlQISk&t=121s) Commercial pricing for GLM-5.3-Flash was set at $0.15 per million input tokens and $0.50 per million output tokens, making it roughly 40 times cheaper than Claude during introductory discounts. - [03:03](https://www.youtube.com/watch?v=r-tzcMlQISk&t=183s) In practical testing, GLM-5.3-Flash modernized an outdated 2016 AngularJS codebase into vanilla JavaScript and used native vision capabilities to process video frames via FFmpeg.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=r-tzcMlQISk&t=0s) Ox Alpha silently appeared on OpenRouter, serving 42 trillion tokens in six days and doubling the usage volume of DeepSeek. - [00:00](https://www.youtube.com/watch?v=r-tzcMlQISk&t=0s) The anonymous model offered a 1 million token context window, native video input capabilities, and zero-cost access during its launch window. - [00:46](https://www.youtube.com/watch?v=r-tzcMlQISk&t=46s) Diagnostic stack traces, error codes, and tokenizer fingerprints quickly linked Ox Alpha's infrastructure directly to Zhipu AI. - [00:46](https://www.youtube.com/watch?v=r-tzcMlQISk&t=46s) Developers routed high volumes of proprietary codebases and tools like Claude Code through Ox Alpha to bypass Anthropic's rate limits. - [00:46](https://www.youtube.com/watch?v=r-tzcMlQISk&t=46s) Viral claims suggested an 80% score on SWE-bench, though actual performance figures sat closer to 58%. - [02:01](https://www.youtube.com/watch?v=r-tzcMlQISk&t=121s) Zhipu AI formally unmasked Ox Alpha as GLM-5.3-Flash, releasing the 320-billion-parameter MoE model under an MIT license on Hugging Face. - [02:01](https://www.youtube.com/watch?v=r-tzcMlQISk&t=121s) The model's underlying stealth testing infrastructure ran on 100,000 domestic Chinese accelerator chips rather than relying on Nvidia hardware. - [02:01](https://www.youtube.com/watch?v=r-tzcMlQISk&t=121s) At standard pricing of $0.15 per million input and $0.50 per million output tokens, introductory discounts placed GLM-5.3-Flash at roughly 40x cheaper than Claude. - [02:01](https://www.youtube.com/watch?v=r-tzcMlQISk&t=121s) Real-world trade-offs of the model include slow execution speeds, verbosity, and occasional recursive doom loops during complex planning tasks. - [03:03](https://www.youtube.com/watch?v=r-tzcMlQISk&t=183s) When refactoring the legacy 'Horse 2' AngularJS app, GLM-5.3-Flash chose pure HTML, CSS, and vanilla JavaScript over modern frameworks like React or Vue. - [03:03](https://www.youtube.com/watch?v=r-tzcMlQISk&t=183s) Using visual reasoning, the model diagnosed mobile CSS overflow issues and resolved container layout bugs using min-width: 0. - [03:03](https://www.youtube.com/watch?v=r-tzcMlQISk&t=183s) Multimodal testing showed GLM-5.3-Flash slicing raw video via FFmpeg into frames to analyze tone and generate in-character titles, descriptions, and comments. - [04:31](https://www.youtube.com/watch?v=r-tzcMlQISk&t=271s) Exa operates as an AI-specialized search engine indexing structured databases of research papers, corporate filings, and profiles rather than standard web scrape copy.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=r-tzcMlQISk&t=0s) The Mysterious Rise of Ox Alpha

While putting the final touches on a "Dolly Parton is too fast for the Grim Reaper" tattoo, a brand-new frontier-grade model silently appeared on the internet under the alias Ox Alpha. Within its first six days online, the stealth model served a staggering 42 trillion tokens, completely dominating OpenRouter and doubling the usage volume of the next closest runner-up, DeepSeek.

The anonymous engine arrived packing a 1 million token context window, native video input capabilities, and coding skills described as writing code "like it has trouble making eye contact." Best of all for developers, it was entirely free to use at launch. Speculation ran rampant across developer communities: some hypothesized it was from Zhipu AI, others pointed fingers at Xiaomi, and at least one hopeful observer was convinced that Google had finally shipped something genuinely useful.

- **Token Volume:** 42 trillion tokens served within the first six days of deployment. - **Platform Dominance:** Ranked as the #1 most utilized model on OpenRouter, outpacing DeepSeek by 2x. - **Specifications:** 1,000,000 token context window, native multimodal video support, and zero cost during the initial evaluation run.

---

[00:46](https://www.youtube.com/watch?v=r-tzcMlQISk&t=46s) Clues, Leaks, and Silicon Valley Panic

The release dropped less than two weeks after widespread anxiety hit Silicon Valley over Chinese lab advancements following the DeepSeek Harness release, compounding a brutal month for industry leaders like Dario Amodei and Sam Altman. Ox Alpha bore every hallmark of an origin in China. It represented the fifth anonymous animal-themed model deployed to OpenRouter in recent memory, with all previous four tracing back to Chinese research institutions.

Savvy engineers quickly investigated under the hood. API stack traces and diagnostic error codes matched Zhipu AI's endpoints, and tokenizer fingerprint tests lined up precisely with Zhipu's GLM architecture. Despite fine print warning users that all prompts would likely be retained by a foreign entity, developers threw caution to the wind because "nobody actually cares about privacy when the magic box is free and works this well." Programmers dumped proprietary internal codebases into the prompt box, routed Claude Code through Ox Alpha to bypass Anthropic's rate limits, and drove the mystery model to consume nearly a third of OpenRouter's total weekly bandwidth. Wild rumors followed, including an unverified viral screenshot claiming an 80% score on the SWE-bench benchmark (the actual figure sits closer to 58%), with some commentators going as far as claiming the model was a stolen Claude checkpoint.

---

[02:01](https://www.youtube.com/watch?v=r-tzcMlQISk&t=121s) Unmasking GLM-5.3-Flash and Its Pricing

On August 26th, Zhipu AI officially stepped forward, confirming that Ox Alpha was actually GLM-5.3-Flash. The model is an open-weight, natively multimodal Mixture of Experts (MoE) architecture packing 320 billion total parameters, with its model weights dropping under an MIT license on Hugging Face the same day. Silicon Valley took a notable hit upon learning the entire stealth testing infrastructure was operated on just 100,000 domestic Chinese-made accelerator chips—a development cheekily noted as having "really hurt Jensen's feelings."

```mermaid flowchart TD accTitle: GLM-5.3-Flash Video Processing Pipeline accDescr: Diagram illustrating how GLM-5.3-Flash breaks down video inputs and generates application assets for the legacy app modernization. subgraph Ingest["1. Video Ingestion"] V["Raw Horse Video"] --> FF["FFmpeg Frame Slicing"] end subgraph Analysis["2. GLM-5.3-Flash Multimodal Core"] FF --> Multi["Multi-Frame Vision Analysis"] Multi --> Tone["Context & Humor Extraction"] end subgraph UI["3. Web App Synthesis"] Tone --> Meta["Title & Description"] Tone --> Comments["In-Character Comments"] Tone --> DOM["Vanilla DOM Elements"] end ```

The unmasking brought an end to the free tier, replacing the anonymous endpoint with commercial pricing. However, rates remain remarkably low: $0.15 per million input tokens and $0.50 per million output tokens. With an introductory 50% discount running through September 9th, the model lands at roughly 40 times cheaper than Claude. While many competitor launches lean on questionable "trust-me-bro benchmarks," GLM-5.3-Flash's figures hold up to external testing. The primary drawbacks in day-to-day use are that it runs noticeably slow, "loves to yap," and occasionally falls into recursive doom loops during complex planning tasks.

---

[03:03](https://www.youtube.com/watch?v=r-tzcMlQISk&t=183s) Modernizing Legacy "Horse 2" with Vanilla Stack

To put the engine through its paces, it was tasked with reviving an abandoned personal side project: "Horse 2." Conceived after failing to secure startup funding for "Horse Tinder," Horse 2 was an ancient 2016 AngularJS web app that had sat untouched without a single commit for ten years. The challenge was simple: modernize the legacy codebase without gutting the humor, aesthetic quirks, and custom vibes baked into the original version.

GLM-5.3-Flash started by analyzing the dependencies, correctly identifying the obsolete libraries, and then narrating its refactoring plan in painstaking, uninterrupted detail for roughly 20 minutes straight. Once coding began, it assembled a custom design system in "raw dog CSS," producing a purple sloth backdrop paired with vibrant rainbow gradient buttons—an aesthetic reflecting clear inspiration (or perhaps "distillation") from existing mainstream frontier models.

Leveraging its computer vision abilities, the model spotted a CSS overflow glitch on simulated mobile screens, diagnosed the container layout, and rectified it using the classic, somewhat cursed `min-width: 0` technique. It implemented smooth enter animations on video cards and preserved the original app's comedic tone, generating whimsical toast alerts to match. Rather than swapping AngularJS for React or Vue, the model opted for "no stack" at all: pure HTML, CSS, and vanilla JavaScript, resulting in code that curiously resembled a completely different legacy codebase.

For a true multimodal vision test, a raw horse video was submitted to the model with instructions to embed it into the application alongside contextually accurate titles, descriptions, and mock comments: - The model instructed FFmpeg to slice the video file into individual image frames. - It reviewed the frame sequence to grasp the action and emotional tone. - It produced a context-aware video description and mock user comments that matched the platform's idiosyncratic humor.

---

[04:31](https://www.youtube.com/watch?v=r-tzcMlQISk&t=271s) Sponsor: Supercharging Agents with Exa

The video wraps up with a spotlight on Exa, a dedicated search engine designed specifically for AI agents. Rather than scraping standard web page copy, Exa indexes structured databases containing over 350 million scientific research papers, financial filings, and deep profile data spanning more than 1 billion individuals and 70 million companies, updating with new sources weekly.

Demonstrating its analytical power, Exa was linked to Polymarket tools to power speculative prediction bets—such as tracking tech executives selling personal shares immediately prior to quarterly earnings announcements. Exa successfully retrieved raw SEC filings and normalized the data into structured JSON objects in a single shot while appending full source citations. Developers can interface with Exa directly through native integrations for Claude, ChatGPT, or the dedicated Exa Agent for complex multi-dataset queries, with $20 in introductory API credits available for new users.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=r-tzcMlQISk&t=0s) Dolly Parton is too fast for the Grim Reaper

> [00:00](https://www.youtube.com/watch?v=r-tzcMlQISk&t=0s) like it has trouble making eye contact

> [00:46](https://www.youtube.com/watch?v=r-tzcMlQISk&t=46s) nobody actually cares about privacy when the magic box is free and works this well

> [02:01](https://www.youtube.com/watch?v=r-tzcMlQISk&t=121s) really hurt Jensen's feelings

> [02:01](https://www.youtube.com/watch?v=r-tzcMlQISk&t=121s) trust-me-bro benchmarks

> [02:01](https://www.youtube.com/watch?v=r-tzcMlQISk&t=121s) loves to yap

> [03:03](https://www.youtube.com/watch?v=r-tzcMlQISk&t=183s) raw dog CSS

> [03:03](https://www.youtube.com/watch?v=r-tzcMlQISk&t=183s) min-width: 0

## Mentioned Ox Alpha, OpenRouter, Zhipu AI, GLM-5.3-Flash, DeepSeek, Xiaomi, Google, Dario Amodei, Sam Altman, Claude, Claude Code, Anthropic, SWE-bench, Hugging Face, AngularJS, Horse 2, Horse Tinder, FFmpeg, React, Vue, Exa, Polymarket, ChatGPT, SEC, Dolly Parton

## Video description Exa gives your agents the largest index of public and private data. Get $20 in credits - https://dashboard.exa.ai/billing?coupon=EXA20FIRESHIP

An anonymous model called Ox Alpha just conquered OpenRouter, serving 42 trillion tokens in six days before being unmasked as Zhipu's GLM-5.3-Flash. Let's break it down.

#coding #programming

Want more Fireship?

🗞️ Newsletter: https://bytes.dev 🧠 Courses: https://fireship.dev

[Source](https://www.youtube.com/watch?v=r-tzcMlQISk)

中文简报

智谱AI揭晓Ox Alpha模型，价格比Claude低40倍智谱AI推出的GLM-5.3-Flash表明，在国产硬件上运行的开放模型能以低40倍的价格对西方竞品形成冲击，并在尽管存在数据隐私担忧的情况下吸引开发者流量。一款代号为Ox Alpha的神秘AI模型在OpenRouter上免费处理了42万亿token，令开发者感到震惊并引发了行业恐慌，随后该模型被揭晓为智谱AI的3200亿参数开放权重模型GLM-5.3-Flash。该新模型运行在中国国产芯片上，提供原生视频支持和100万token上下文窗口，其费率比Claude便宜约40倍。为了评估其实际表现，Fireship让GLM-5.3-Flash将一个废弃的2016年AngularJS项目重构为原生Web代码，并在原始视频切片上测试了其原生计算机视觉能力。视频最后介绍了Exa，这是一款专为查询结构化数据库而打造的面向智能体（agent）的搜索引擎。代号为Ox Alpha的匿名模型在OpenRouter上占据主导，在六天内处理了42万亿token，具备100万token上下文窗口并原生支持视频。 (0:00) 智谱AI揭晓Ox Alpha即为GLM-5.3-Flash，这是一款基于MIT协议发布的3200亿参数MoE开放权重模型，并在中国国产加速器上训练而成。 (2:01) GLM-5.3-Flash的商业定价定为每百万输入token 0.15美元、每百万输出token 0.50美元，在首发折扣期内比Claude便宜约40倍。 (2:01) 在实际测试中，GLM-5.3-Flash将一个过时的2016年AngularJS代码库现代化改写为原生JavaScript，并利用原生视觉能力通过FFmpeg处理了视频帧。 (3:03)

Reading summary

Chinese AI lab Zhipu AI has officially unveiled GLM-5.3-Flash—previously deployed as the anonymous stealth model Ox Alpha—offering a 320-billion-parameter open-weight model at roughly one-fortieth the price of Anthropic's Claude. The Stealth Debut of Ox Alpha (0:00) Before its formal attribution, the model debuted silently on OpenRouter under the pseudonym Ox Alpha, equipped with a 1-million-token context window, native video input, and free access. Within six days, the engine served 42 trillion tokens, outpacing the next closest model, DeepSeek, by twofold and commanding nearly a third of OpenRouter's total weekly traffic. Despite service terms indicating prompts would be stored by a foreign entity, developers adopted it rapidly, running proprietary codebases and routing Claude Code through the endpoint to evade Anthropic rate limits. Community speculation was fueled by API stack traces, error codes, and tokenizer fingerprints that aligned with Zhipu AI's GLM architecture. While viral rumors alleged that the model achieved an 80% score on the SWE-bench benchmark or was a checkpoint derived from Claude, its validated SWE-bench score sits closer to 58%. Architecture, Domestic Hardware, and Aggressive Pricing (2:01) On August 26, Zhipu AI confirmed ownership and branded the release GLM-5.3-Flash, releasing its weights under an MIT license on Hugging Face. The system uses a natively multimodal Mixture of Experts (MoE) architecture with 320 billion total parameters. The testing infrastructure operated entirely on 100,000 domestic Chinese accelerator chips, demonstrating frontier-tier capability without reliance on Western semiconductor suppliers. With commercialization, the free tier transitioned to standard pricing of $0.15 per million input tokens and $0.50 per million output tokens, alongside a temporary 50% discount running through September 9. At these rates, the model is roughly 40 times less expensive than Claude. However, daily usage exhibits drawbacks: the engine runs noticeably slow, produces verbose explanations, and occasionally enters recursive loops during complex reasoning steps. Real-World Code Modernization and Video Ingestion (3:03) To evaluate its practical coding abilities, GLM-5.3-Flash was tasked with modernizing 'Horse 2,' a 2016 AngularJS codebase untouched for ten years. The model analyzed the project's dependencies and articulated a refactoring plan over a 20-minute planning phase before replacing the framework entirely with vanilla HTML, CSS, and JavaScript. It also used vision capabilities to diagnose a mobile layout overflow, applying a standard CSS fix to resolve container bounds while matching the project's original styling and comedic tone. For multimodal video processing, the model parsed raw video files by directing FFmpeg to slice the media into discrete image frames. Analyzing the sequential frames allowed the model to interpret visual action and humor, successfully generating contextual titles, summaries, and synthetic user comments that mirrored the platform's established voice. GLM-5.3-Flash confirms that competitive, open-weight multimodal models can be trained and served on domestic Chinese hardware at a fraction of prevailing market costs, though users must balance dramatic price reductions against slower generation speeds and verbosity.

Source quotes

[0:00] Dolly Parton is too fast for the Grim Reaper [0:00] like it has trouble making eye contact [0:46] nobody actually cares about privacy when the magic box is free and works this well [2:01] really hurt Jensen's feelings [2:01] trust-me-bro benchmarks [2:01] loves to yap [3:03] raw dog CSS [3:03] min-width: 0

Original transcript

[0:00] A few weeks ago, while I was busy [0:01] putting the final touches on my Dolly [0:03] Parton is too fast for the Grim Reaper [0:05] tattoo, is someone anonymously dropped a [0:07] new frontier grade model online. They [0:09] called it AUX Alpha and in its first 6 [0:11] days it served 42 trillion tokens, which [0:14] easily made it the most popular model on [0:16] Open Router, doubling the next closest, [0:18] which was Deep Seek. It comes with a [0:19] million token context window, accepts [0:22] video input, writes code like it has [0:24] trouble making eye contact, and until [0:26] just a few days ago, it was completely [0:27] free. This some assumed it was Zeku, [0:30] others Xiaomi, and at least one guy was [0:31] convinced that Google actually shipped [0:33] something useful. In today's video, [0:35] we'll look at how this anonymous model [0:36] won over the internet, who's behind it, [0:38] and see if it can finally help me [0:39] realize my horse SAS dream. It is [0:41] September 1st, 2026 and you're watching [0:44] The Code Report. It hasn't even been 2 [0:46] weeks since we talked about how Silicon [0:48] Valley is terrified of what's happening [0:49] in China after the release of Deep Seek [0:51] Harness, and unfortunately, Dario and [0:53] Sam's terrible, horrible, no good, very [0:55] bad month may have gotten even worse [0:57] because this new anonymous model had [0:59] made in China written all over it. And [1:01] as we know, things that escape out of [1:03] Chinese labs tend to have a big impact [1:05] on the rest of the world. The first clue [1:07] was that AUX Alpha was the fifth [1:08] anonymous animal model to show up on [1:10] Open Router recently, and the four [1:12] before it all turned out to be from [1:13] China. And even more convincing, some [1:15] devs noticed that its stack traces and [1:17] error codes matched Zeku's API, and a [1:20] tokenizer fingerprint test came back [1:22] matching their GLM series. But nobody [1:24] waited for a confirmation to start using [1:26] it. Yes, the fine print did say that all [1:28] your prompts would be retained by [1:30] probably foreign entity, but nobody [1:32] actually cares about privacy when the [1:34] magic box is free and works this well. [1:36] So, devs happily pasted their companies [1:38] proprietary code into it, routed Claude [1:40] code through it to dodge their rate [1:41] limits, and at its peak, AUX Alpha [1:43] accounted for nearly a third of Open [1:45] Router's entire weekly traffic. The hype [1:47] machine got so out of control that a [1:49] misleading screenshot showing an 80% [1:51] score on the Deep SWE benchmark went [1:53] viral, even though the real number [1:55] turned out to be closer to 58, and [1:57] someone even speculated it was so good [1:59] it might be a stolen Claude checkpoint. [2:01] But sure enough, on August 26th, Zhipu [2:03] came clean and announced that Aux Alpha [2:05] was GLM 5.3 Flash, a natively multimodal [2:09] mixture of experts model with 320 [2:11] billion total parameters, and its MIT [2:13] licensed weights hit Hugging Face the [2:15] same day. But the detail that really [2:17] killed the vibe in Silicon Valley was [2:18] that the entire stealth run was served [2:21] on just 100,000 Chinese-made chips, [2:23] which I'm told really hurt Jensen's [2:25] feelings. Of course, the reveal also [2:27] meant the free samples were over, and [2:29] the anonymous endpoint was replaced with [2:31] an actual pricing page. But it's still [2:33] stupid cheap at just 15 cents per [2:35] million input tokens and 50 cents per [2:37] million out. And with a 50% discount [2:39] through September 9th, it's currently up [2:41] to 40 times cheaper than Claude. And as [2:43] we know, normal launches like this come [2:46] with a bunch of trust-me-bro benchmarks [2:48] that usually fall apart the second [2:49] anyone touches them. But their [2:50] benchmarks hold up pretty well under [2:52] outside scrutiny. The catch is that it's [2:54] slow, loves to yap, and occasionally [2:56] gets stuck in doom loops. But in my [2:58] opinion, there's only one way to test if [3:00] it's actually any good, and that's by [3:02] using it on a dear project of mine. [3:04] After failing to raise funding for Horse [3:05] Tinder, I've pivoted to its once-beloved [3:07] cousin, Horse 2, a 2016 Angular JS app [3:11] whose last commit was 10 years ago. So I [3:13] asked it to modernize the legacy stack [3:15] without losing the humor and vibes I had [3:17] so carefully crafted. It correctly [3:19] detected the outdated technologies and [3:21] then proceeded to narrate its plan of [3:23] attack in every little detail for at [3:25] least 20 minutes. It then started [3:26] building a design system in raw dog CSS. [3:29] It went with a beautiful purple sloth [3:31] background and rainbow gradient buttons, [3:33] showing true inspiration, sorry, [3:35] distillation from other popular LLM's [3:38] visual tastes. Because of its vision [3:39] capability, it was able to detect a CSS [3:42] overflow bug on mobile screens, properly [3:44] diagnose it, and fix it with the cursed [3:47] min-width equals zero trick that always [3:49] seemed to work. It added some nice enter [3:51] animations for video cards, which was a [3:53] really nice touch. All the bits of humor [3:55] are still there, and it even added some [3:56] new ones like these toast notifications. [3:59] The code-wise, it replaced the legacy [4:00] stack with no stack. It wrote the app in [4:02] HTML, CSS, and vanilla JavaScript, and [4:05] wrote code that looks like a legacy code [4:07] base of its own. But I really wanted to [4:08] test its vision skills, so I fed it a [4:10] random horse video and asked it to add [4:12] it to the app and create a title, [4:14] description, and comments that are [4:16] relevant to what's in the video. It [4:17] decided to use FFmpeg to break down the [4:20] video into a series of frames, analyze [4:22] those frames, and it did a really good [4:24] job. It got the full context of what [4:26] happens in the video, nailed the [4:27] description, and even aligned the humor [4:29] with the rest of the platform. And [4:31] another tool worth checking out is Exa, [4:33] the sponsor of today's video. It's a [4:35] search engine built for AI agents that [4:37] gives you a lot more than just web [4:38] results. It includes 350 million [4:41] research papers, financial data, and [4:43] info on over 1 billion people and 70 [4:45] million companies, with new sources [4:47] added every week. I hooked it up to [4:49] PolyChat to let my users gamble on the [4:51] stock market in even more degenerate [4:53] ways, like betting on which tech CEOs [4:55] will sell company stock right before [4:57] their next earnings call. And Exa was [4:58] able to pull the recent SEC filings and [5:01] convert it into structured JSON in one [5:03] shot. It also cites all of its sources, [5:05] so users can't accuse me of cheating [5:07] anymore. You can plug directly into [5:09] Claude, ChatGPT, or your other favorite [5:11] agent using their native plugins. Or try [5:13] Exa agent to get answers to more complex [5:16] queries that need to pull from multiple [5:18] data sets simultaneously, which is a lot [5:20] more powerful than what LLMs can do on [5:22] their own. Try Exa today for free and [5:24] get $20 in credits at the link below. [5:27] This has been the Code Report. Thanks [5:28] for watching, and I will see you in the [5:30] next one.