GLM 5.3 Flash Was Ox Alpha The WHOLE Time ($0.29 Full SaaS)
It's official: Ox Alpha — the anonymous model 221,000 people were using free — was GLM 5.3 Flash the whole time. Z.ai confirmed it, published the MIT weights, and put up benchmark charts showing it beating Claude Opus 4.8 on GDPval-AA (1773 vs 1582), Toolathlon and AutomationBench. So I did what I always do: I gave it the same one-prompt SaaS build I gave Qwen 3.8 Flash yesterday. The benchmarks and my build disagree. THE REVEAL • Ox Alpha = GLM 5.3 Flash, confirmed August 26 • 320B total parameters, 18B active — natively multimodal, 1M context • MIT weights on Hugging Face — but at 320B this is server hardware, not laptop hardware (Qwen 3.8 Flash Next is 125B/6B, which is why THAT one runs locally) • $0.15/M input, $0.50/M output — 50% launch promo until September 9 MY TEST — SAME PROMPT AS YESTERDAY'S QWEN BUILD • It built the full SaaS: auth, Stripe payments, credits, image generation — all working • Total cost: $0.29 on OpenRouter (the cache hit rate is excellent) • BUT: it couldn't read the OpenRouter docs on its own — I had to give it my Bright Data skill. Qwen didn't need that • More errors back and forth, less self-verification, and the design is nowhere near Qwen's • Verdict: a good model, a great price... and I still prefer Qwen 3.8 Flash The one genuinely new thing: visual self-verification. It's natively multimodal, so it can look at its own output and improve it. If that gets cheap enough, it changes how these models build. ⚔️ Watch the Qwen 3.8 Flash build this is compared against: [LINK TO YESTERDAY'S VIDEO] 📦 MY FREE BRIGHT DATA SKILL https://github.com/IncomeStreamSurfer/brightdata-scraper-studio-skill The scraping skill I used to feed it real documentation — one-click install, MIT licensed, link in the pinned comment. 🎁 BRIGHT DATA — $25 FREE CREDIT Enough to scrape tens of thousands of pages. It's what gives OpenCode, the DeepSeek Harness — and even Claude Code — real eyes on the internet. Use my link or code ISS25: 👉 https://brdta.com/iss25 💡 STEAL MY SAAS PROMPT The exact prompt that generates an entire SaaS project is shown in the video — new Convex environment, paste, tell it what to build. CHAPTERS 0:00 — Ox Alpha was GLM 5.3 Flash all along 0:15 — China's Flash family is taking over 0:33 — Cheaper than Opus 4.8, and open 0:49 — 320B/18B: too big for your machine 1:07 — Why Qwen (125B/6B) runs locally and this doesn't 1:30 — My test: the same one-prompt SaaS build 1:38 — Early verdict: not as good as Qwen 1:46 — It needed Bright Data to read the docs 2:10 — Natively multimodal — no extra image cost 2:19 — Visual self-verification is genuinely clever 2:53 — The DeepSeek copycat cycle 3:19 — Build problems: errors back and forth 3:59 — The receipt: $0.29 total on OpenRouter 4:17 — The result vs yesterday's Qwen build 4:42 — Testing Stripe payments (test card!) 5:24 — Checking the Convex data 5:37 — Steal my SaaS prompt 6:10 — Bright Data: $25 free credit 7:35 — Final verdict: I honestly prefer Qwen #GLM #OxAlpha #aicoding glm 5.3 flash, ox alpha, glm 5.3, ox alpha revealed, z.ai, glm vs qwen, qwen 3.8 flash, chinese ai models, open weights, mit license ai, ai coding agent, build saas with ai, opencode, openrouter, bright data, deepseek v4 flash, opus 4.8, gdpval, ai benchmarks, benchmarks vs reality, flash models, moe model, multimodal llm, visual self verification, cheap ai coding, ai coding 2026, vibe coding, one prompt saas, glm 5.3 flash review, mystery model
