Terminal-Bench 2.1, blended API price per 1M tokens at a 3:1 input:output mix – Qwen sits at $2.48 and scores 86.6
Points of agentic capability per dollar:
• Qwen 3.8 Max: 35.0
• Kimi K3: 14.2 – 2.5x behind
• Claude Opus 5: 8.9 – 3.9x behind
• GPT-5.6 Sol: 8.0 – 4.4x behind
• Claude Fable 5: 4.2 – 8.3x behind
Sol and Opus 5 still score higher in absolute terms – 89.5 and 89.1. That's ~3 points, or 2-3 tasks (TB 2.1 is exactly 89 tasks, so points map almost 1:1 to tasks solved). You pay 4x for those
Against Fable 5 there's no tradeoff at all. Qwen scores 86.6 to Fable's 84.6 and costs 8x less. @Alibaba_Qwen said 3.8 Max trails only Fable 5 – on this benchmark it's ahead of it
The only model at the same price point is GLM-5.2 at $2.15 blended. It scores 77.9. Same money, 8.7 points apart
Caveat: 86.6 is Alibaba's own run, no independent number yet. Everything else here is @ArtificialAnlys – Terminus 2, E2B sandbox, pass@1 over 3 repeats. Kimi self-reported 88.3 and independently came in at 85.0, so vendor numbers in this class have been running ~3 points hot

📢Meet Qwen3.8-Max — our most capable model to date.
— Qwen (@Alibaba_Qwen) August 3, 2026
Next week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights to meet you all!🎉
Qwen3.8-Max, a new bar for coding and cowork at 2.4T parameters:
- Autonomous coding: 10+ days of… pic.twitter.com/e3YFj2hqcT