Summary
Matthew Berman covers the release of Alibaba’s Qwen 3.8 Max, a 2.4-trillion-parameter open-weights model positioned as competitive with leading closed-source frontier models from OpenAI and Anthropic. Berman walks through benchmark comparisons including Terminal Bench (Qwen 3.8 Max at 86.6 vs. Fable’s 84.6), Sweep Pro (67.7 vs. Fable’s 80), and multimodal reasoning — where Qwen dominates — alongside document and office intelligence categories. He cautions that benchmarks can be gamed and notes that Kimi K3, despite strong numbers, underperforms Fable on real tasks.
A significant portion of the video focuses on pricing economics. Qwen 3.8 Max is available on OpenRouter at $2 per million input tokens and $6 per million output tokens, compared to GPT-5.6 Soul at $5/$30 and Fable at $10/$50. Berman argues the more meaningful metric is cost-per-task-completed — referencing Artificial Analysis data showing Qwen 3.7 Max at $1.28 per task versus Kimi K3 Max at roughly $0.80-0.90 — and notes Qwen 3.8 Max has not yet been benchmarked on that framework.
Berman also highlights Alibaba’s demonstration of the model’s research reproduction capability, where Qwen 3.8 Max reproduced academic paper results from scratch and independently generated 18 improvement ideas across four iterations — framing this as a step toward autonomous AI research. He positions both Qwen 3.8 Max and Kimi K3 as evidence that Chinese open-source labs are converging on the same 2-3 trillion parameter frontier scale as leading U.S. closed-source models.
📺 Source: Matthew Berman · Published August 04, 2026
🏷️ Format: Review
