Source: MarkTechPost — 2026-08-03
Summary
Alibaba's Qwen3.8-Max moved from bare preview to full general availability on August 3, 2026, this time with a benchmark sheet attached — the July 20 preview had shipped with no numbers at all. The model is a 2.4-trillion-parameter Mixture-of-Experts system with roughly 95B active parameters per token, a 1-million-token context window, and native text/image/video input. On OSWorld-Verified, a benchmark for AI agents operating a real computer desktop, it scores 86.1 — ahead of both GPT-5.6 Sol Max (83.2) and Fable 5 (85.0). It's priced at $2 input / $6 output per million tokens, with an OpenAI- and DashScope-compatible API.
Key Takeaways
- Architecture: 2.4T total parameters, ~95B active per token (roughly 4% activation ratio), 1M-token context, native multimodal (text, image, video) input.
- OSWorld-Verified (agentic computer-use benchmark): 86.1 — beats GPT-5.6 Sol Max (83.2) and Fable 5 (85.0).
- Terminal-Bench 2.1: 86.6, ahead of Claude Opus 4.8 and Fable 5 (both 84.6), but behind GPT-5.6 Sol Max's 88.8.
- On SWE-bench Pro and FrontierSWE, Qwen3.8-Max trails Fable 5 by a wide margin (67.7 vs 80.0, and 73.5 vs 88.8 respectively) — the agentic-computer-use lead doesn't carry over to software-engineering benchmarks.
- Overall public leaderboard rank: #46 of 215, score 60.9/100 — a reminder that a strong showing on one benchmark category doesn't mean top-tier across the board.
- Pricing undercuts most frontier proprietary models at $2/$6 per million tokens, with open weights reportedly following within the week.
Reel Script
Hook (~18s, 40 words): Alibaba just released benchmark numbers for its newest flagship model, and it's claiming something specific: better than GPT-5.6 and Fable 5 at actually operating a computer — clicking, typing, navigating apps — not just answering questions.
Core Concept (~75s, 175 words): Qwen3.8-Max is a Mixture-of-Experts model — think of it as a company with 2.4 trillion dollars of total specialist talent on payroll, but for any single task, only about 95 billion dollars' worth of specialists actually get called into the room. That's why a model this large can still run at a competitive price: you're not paying to activate the whole company, just the relevant department. The headline number is OSWorld-Verified, a benchmark that doesn't ask a model to write an essay — it drops an AI agent into a real desktop environment and scores whether it can complete tasks the way a human would, by clicking buttons and navigating real software. Qwen3.8-Max scores 86.1 there, edging out both GPT-5.6 Sol Max and Fable 5. But — and this matters — that lead doesn't hold everywhere. On pure software-engineering benchmarks like SWE-bench Pro, it falls well behind Fable 5.
Hands-On (~60s, 140 words): Here's the actual scoreboard worth screen-capturing: OSWorld-Verified — Qwen3.8-Max 86.1, Fable 5 85.0, GPT-5.6 Sol Max 83.2. Terminal-Bench 2.1 — GPT-5.6 Sol Max 88.8, Qwen3.8-Max 86.6, Opus 4.8 and Fable 5 tied at 84.6. Now flip to SWE-bench Pro — Fable 5 jumps to 80.0, Qwen3.8-Max drops to 67.7. Same model, wildly different standing depending on which task you're measuring. That's the real lesson: "beats GPT-5.6 and Fable 5" is true for one narrow category — desktop agent tasks — and false for general software engineering. Pricing is $2 per million input tokens, $6 per million output, through an OpenAI-compatible API, so swapping it into an existing agent stack is mostly a base-URL change.
Takeaway (~25s, 55 words): Don't take "Qwen beats GPT-5.6" headlines at face value — check which specific benchmark that claim is riding on before you swap your agent stack over. For desktop-automation agents specifically, Qwen3.8-Max is now a legitimate cheaper alternative worth benchmarking against your own workload.