Video summary
The mystery is solved... and the answer is 40x cheaper than Claude
Main summary
Key takeaways
Main technological/product analysis: “Ox Alpha” (an anonymous Frontier-grade model)
Origin & identity
- An anonymous model called Ox Alpha appeared on Open Router, following several other anonymous “animal” models previously traced to China.
- Early evidence included:
- Stack trace / error code matches to Zoo’s API
- A tokenizer fingerprint matching Zoo/GLM-style GLM series tokenization
- A privacy warning in fine print stated prompts were retained by a mysterious entity, but adoption still surged due to performance/cost.
Performance & adoption
- In its first 6 days, it reportedly processed 42 trillion tokens, becoming the most popular model on Open Router and doubling the next closest (DeepS).
- At peak, it accounted for ~one-third of Open Router’s weekly traffic.
- Reports included people routing Claude via Ox Alpha to dodge rate limits.
- Some hype/marketing spread:
- A viral screenshot claimed an 80% result on a “DeepS.WE” benchmark; it was later corrected to about 58%.
- Speculation suggested it might be a stolen Claude checkpoint.
Official reveal
- Jepu later confirmed (Aug 26) that Ox Alpha = GLM 5.3 Flash.
- Described characteristics:
- Natively multimodal Mixture-of-Experts (MoE)
- ~320B total parameters
- MIT-licensed weights released on Hugging Face the same day
Hardware / compute note
- The “stealth run” allegedly used only 100,000 Chinese-made chips, framed as a competitive/strategic point.
Pricing / cost comparison
- Ox Alpha switched from fully free samples to a paid endpoint with a pricing page.
- Cited costs:
- 15 cents per million input tokens
- 50 cents per million out tokens
- With a 50% discount through Sep 9, it was claimed to be up to ~40× cheaper than Claude.
Known behavior / limitations
- Benchmarks reportedly “held up” under outside scrutiny, but the host characterizes it as:
- Slow
- Prone to excessive output (“loves to yap”)
- Occasionally stuck in doom loops
Practical “tutorial/test” use case: modernizing a legacy web app
Test project
- The speaker tested Ox Alpha by giving it a “Horse” project:
- A 2016 Angular.js app
- Last commit ~10 years ago
- Goal: modernize the stack without losing humor/vibes.
What it did (feature demonstration)
- Detected outdated technologies and outlined a modernization plan (noted as very verbose, with extensive “narrating”).
- Built a design system using raw CSS, including:
- Styled UI elements (e.g., purple slot backgrounds)
- Rainbow gradient buttons
- Used vision capability to:
- Detect a CSS overflow bug on mobile
- Diagnose and fix it using the “min-width: 0” workaround (“cursed min width equals zero trick”)
- Add enter animations for video cards
- Preserved existing humor and added UX elements like toast notifications.
Coding style / modernization approach
- Replaced a legacy stack with a simplified implementation:
- HTML + CSS + vanilla JavaScript
- The output was described as resembling a coherent legacy-style codebase, rather than an alien rewrite.
Vision/video understanding test
- The speaker provided a random horse video and asked the model to:
- Add it to the app
- Generate a title, description, and comments aligned with the video content
- Claimed approach:
- Used FFmpeg to break the video into frames
- Performed frame analysis
- Reported results:
- Correctly captured the full context
- Produced a relevant description
- Matched humor style to the platform
Additional tool covered (sponsor): Exa for AI-agent search
What Exa is
- A search engine for AI agents with more than web results:
- 350M research papers
- Financial data
- Info on 1B people and 70M companies
- Weekly source updates
Integration / example
- The sponsor example involved hooking Exa into Polyhip to retrieve recent SEC filings.
- Exa purportedly converts filings into structured JSON in one shot.
- Exa provides citations/source links to reduce “cheating” accusations.
Agent usage
- Plugin compatible with Claude, ChatGPT, and other agent systems (via native plugins).
- Mentions Exa Agent for complex queries spanning multiple datasets.
Offer
- Free trial and $20 credits via link (sponsor callout).
Main speakers/sources (as stated)
- Speaker/host: narrator of “The Code Report” (no personal name given in subtitles)
- Source/confirming entities:
- Jepu (confirmed Ox Alpha = GLM 5.3 Flash on Aug 26)
- Zoo / GLM / GLM series (inferred from tokenizer + API matching)
- Exa (video sponsor)