MiniMax M3 vs Qwen 3.8 Max in Claude Code — Same Score, 54x the Tokens

Original author：Mohd MILLAT

Original source：https://www.youtube.com/embed/BC_T66bod7k

MiniMax M3 vs Qwen 3.8 Max in Claude Code — I gave both models the exact same build task and graded them automatically. Result: a perfect 20/20 tie on code quality... but MiniMax M3 finished 4x faster on 54x fewer input tokens. Here's the experiment: one prompt, two fresh folders, identical 14 acceptance criteria. MiniMax M3 ran on the official MiniMax token plan; Qwen 3.8 Max ran through my local router. A Python grader scored both builds on 20 checks — merge logic, win/lose detection, seeded RNG, renderer alignment, test quality — zero manual scoring. Both models scored 20/20. Both wrote 53 tests. Both passed every criterion. The difference is how they got there: MiniMax M3 took 6.3 minutes and ~230K input tokens. Qwen 3.8 Max took 25.3 minutes and ~12.4M input tokens. Same result, wildly different token bills. CHAPTERS (adjust timestamps to your cut): 00:00 MiniMax M3 vs Qwen 3.8 Max — the setup 01:30 The 2048 gauntlet: one prompt, two models 04:00 MiniMax M3 session in Claude Code (6.3 min) 09:00 Qwen 3.8 Max session in Claude Code (25.3 min) 14:00 Auto-grading: 20 criteria, zero manual scoring 16:30 The verdict: quality tie, efficiency isn't 18:00 Which one should YOU use? SETUP SHOWN IN THE VIDEO: • MiniMax M3: official token plan — api.minimax.io/anthropic, model MiniMax-M3[1m] Setup docs: https://platform.minimax.io/docs/token-plan/claude-code • Qwen 3.8 Max: local OpenRouter-compatible router, route qd/qmodel_38max • Grader + acceptance criteria: fully automated, source shown on screen • Claude Code version used for both sessions: 2.1.234 THE NUMBERS (from the session logs): MiniMax M3: 6.3 min · 82 turns · 230K input · 118K output · 20/20 Qwen 3.8 Max: 25.3 min · 127 turns · 12.4M input · 143K output · 20/20 Token ratio: 54x more input tokens for the same quality score. Qwen's edge worth mentioning: more thorough tests (116 assertions vs 90, 11/12 edge probes vs 7/12). MiniMax's edge: 4x faster wall-clock and dramatically cheaper per completed task — which is exactly what Artificial Analysis found independently (~$0.14 vs ~$1.13 cost per task). Fairness rules I followed: same prompt verbatim, same Claude Code version, no steering, no fixes — any intervention would have counted as a failure on camera. If you're deciding between these two for Claude Code: quality is a coin flip, your wallet is not. #minimaxm3 #qwen38max #claudecode Related searches covered: minimax m3 claude code setup, qwen 3.8 max claude code, minimax m3 vs qwen head to head, qwen 3.8 vs claude speed, minimax m3 benchmark real test, qwen 3.8 max coding review, claude code with third party models, ai coding agent comparison 2026. ⚠️ Disclosure: This series is paid promotional content — but testing was 100% unbiased and no company interfered with the results. Model access: MiniMax M3 came from MiniMax's official source — the MiniMax Token Plan (free credits, used for testing only). Qwen 3.8 Max came via Qoder's public promotion "1st Anniversary × Qwen3.8-Max Launch" (docs.qoder.com/events/qwen-max) — 800–2,000 free model calls, available to ANY new user. Qoder did not give us anything special; they don't know us and are not involved in this video. I already have a video showing how to claim the free calls — linked below. Claims open until Sept 3, 2026. Every test ran identically on both models and was graded automatically by a script. 🌍 Audio available in English (auto-dubbed) and Bangla (original) — switch the audio track in player settings. Subtitles available in most languages. ⚡ Uncut real videos — no edits, nothing staged. Use the chapters to skip ahead. 🔐 SECURITY NOTE Never share your API keys, and never paste them into videos, screenshots or public repos. Rotate a key immediately if it's ever exposed. 🇧🇩 বাংলাদেশি ভাইদের জন্য ভিডিওর অডিও default-এ English (auto-dub) থাকে — আমার আসল বাংলা voice শুনতে player settings → Audio track → "Bangla original" সিলেক্ট করুন। 🌐 A NOTE FOR VIEWERS AROUND THE WORLD Thank you for watching from outside Bangladesh — a large part of this channel's audience now comes from India, Pakistan, Indonesia and beyond. Comment in your own language; I'll do my best to reply. 🇧🇩 ধন্যবাদ! বাংলায় comment করুন। 🇹🇷 Teşekkürler! Kendi dilinizde yorum yapabilirsiniz. 🇵🇰 شکریہ! اپنی زبان میں کمنٹ کریں — میں جواب دوں گا۔ 🇮🇩 Terima kasih! Silakan berkomentar dalam bahasa Anda. 🇻🇳 Cảm ơn bạn! Hãy bình luận bằng tiếng Việt. 🇪🇬 شكراً! اكتب تعليقك بلغتك. 🇧🇷 Obrigado! Comente no seu idioma. 🇷🇺 Спасибо! Пишите комментарии на своём языке. 🇮🇳 धन्यवाद! अपनी भाषा में कमेंट करें — मैं जवाब दूंगा। 🎯 Subscribe for practical AI coding tutorials — no hype, real projects, honest pricing. 🌐 Website: https://millat.is-a.dev 💻 GitHub: https://github.com/codermillat 🤗 Hugging Face: https://huggingface.co/millat 💼 LinkedIn: https://linkedin.com/in/codermillat 🐦 X (Twitter): https://twitter.com/codermillat 👍 If this helped, hit Like, Subscribe, and turn on the notification bell.