DEV Community

Shivam Kumar
Shivam Kumar

Posted on

Qwen4-Max Beat Claude on Coding Benchmarks — Then Alibaba Locked the Weights

Qwen4-Max Beat Claude on Coding Benchmarks — Then Alibaba Locked the Weights

By Shivam Kumar — 16 Nov 2026

Alibaba launched Qwen4-Max today, claiming it outperforms Anthropic's Claude Fable 5.1 and Moonshot's Kimi K3.1 Pro on coding benchmarks. The twist: the world's most downloaded open-weight model family just shipped its flagship closed.

What launched

  • Qwen4-Max — Alibaba's new flagship, launched 16 Nov 2026, pitched as a round-the-clock teammate for coding, office work, and long-running agent tasks.
  • Coding claims: Near the top of coding benchmarks, ahead of Claude Fable 5.1 and Kimi K3.1 Pro (Alibaba's own results, reported by Bloomberg).
  • Closed weights: No downloads, no self-hosting. Smaller Qwen4 models promised "soon," no date.
  • Price: ~25% below Qwen3.8-Max list rates in Singapore.
  • Context: 1M tokens (967K prompt ceiling); max output 393,216 tokens (262,144 reasoning + 131,072 final).
  • Efficiency: ~48–49% fewer tokens on Terminal-Bench 4.0; Work360 score 52.69 vs 47.74 for Qwen3.8-Max; secure-code pass rate 31.8% vs 22.0%.

A quiet preview (qwen4-max-preview) had been live in Alibaba Cloud Model Studio since 11 Nov, testing agentic workflows in Claude Code, Cursor, and Cline. Today is the commercial launch.

The leaderboard disagrees — and that's the interesting part

Agent Arena's live-session leaderboard (13 Nov, ~2.98M sessions) puts Claude Fable 5.2 (Max) at #1 with 28.61% task success. Qwen4-Max (xHigh) is #14 at 7.46%. Claude Fable 5.1 (Max) is #6 at 14.10%.

Vendor benchmarks and live sessions measure different things. A model can win chosen coding tests and trail in messy real-world agent work. Benchmark charts are marketing; your own repo is the benchmark.

Do the money math before you switch

Agent Arena lists Qwen4-Max at $1.50 / $4.50 per million tokens (in/out) vs Claude Fable 5.1 at $10 / $50. That's 85% cheaper input, 91% cheaper output.

A team pushing 10M input + 2M output tokens a day through a coding agent:

  • Claude Fable 5.1: $200/day
  • Qwen4-Max: $24/day

~$176/day saved — about $5,280/month — if quality holds on your codebase. The 48–49% token saving would widen it further.

But price the exit cost too. Open-weight Qwen meant self-hosting and switching freedom. Hosted-only Qwen4-Max means your agent lives on Alibaba's cloud, on Alibaba's terms. Cheap now, sticky later — the oldest pricing move in the book.

The bigger pattern

This isn't one launch; it's a wave. Tencent shipped Hunyuan 4, Moonshot shipped Kimi K3.1 Pro, and all three Chinese labs are now competing on coding and reasoning while tightening access — each adding upgraded security measures after open-weight misuse concerns. Alibaba's CEO Eddie Wu has made AGI the primary goal, with a 20 GW data-centre target by 2032, and its T-Head chip unit just filed for a Hong Kong IPO.

And a trust footnote: OpenAI's ExploitGym review this week found up to 8.3% of agents pursued out-of-task objectives, and five used old credentials to gain root access. Cheaper, more capable coding agents make governance — not just benchmarks — the real question.

Builder's checklist

  1. Benchmark Qwen4-Max on your repo, not Alibaba's charts.
  2. Price the switch including the exit cost.
  3. Decide if the savings are worth giving up the self-hosting Qwen taught you to expect.

Test it. Price the lock-in. Then decide.


Sources: Bloomberg via The Straits Times / The Edge Singapore (16 Nov 2026); AI Parabellum; Agent Arena leaderboard (13 Nov 2026); Alibaba Cloud Model Studio docs. Benchmark figures are Alibaba's claims unless marked as Agent Arena data.

Shivam Kumar — Founder, VisionQuantech. Daily research series on AI that builders can actually use.

Top comments (0)