Value-ranked with real ratings and real prices — because a cheap model that fails the task is the most expensive one.
Live data · updated 2026-09-15 18:18 · sources: LMArena + OpenRouter
glm-5.3-flash: arena-top-20 quality at $0.25/1M output — the best deal in AI right now.
| Model | Output $/1M | Arena rating | Value | License |
|---|---|---|---|---|
| glm-5.3-flash | $0.25 | 1588 | ★★★★★ | MIT |
| qwen3.8-flash-next | $0.47 | 1635 | ★★★☆☆ | qwen-community-1.0 |
| gemini-3.8-flash-high | $3.75 | 1493 | ★☆☆☆☆ | Proprietary |
| muse-spark-1.3-max | $4.25 | 1645 | ★☆☆☆☆ | Proprietary |
| muse-spark-1.2 (xHigh) | $4.25 | 1500 | ★☆☆☆☆ | Proprietary |
| qwen3.8-max-0902 | $6 | 1681 | ★☆☆☆☆ | Proprietary |
| grok-4.6-high | $6 | 1596 | ★☆☆☆☆ | Proprietary |
| gemini-omni-1.1-flash | $9 | 1515 | ★☆☆☆☆ | Proprietary |
| kimi-k3-max | $15 | 1674 | ★☆☆☆☆ | Kimi K3 license |
| gpt-5.6-sol-xhigh (codex-harness) | $20 | 1604 | ★☆☆☆☆ | Proprietary |
As of 2026-09-15 18:18, the #20 model on LMArena (muse-spark-1.2 (xHigh), 1500) would have topped the board a year ago — and it costs $4.25 per million output tokens. The value tier now clears bars that frontier models set last year.
Long agentic runs, subtle instruction following and tool-call reliability still separate the top 5 from the pack. For interactive chat and batch transforms, value models are fine; for autonomous agents, measure failure cost before celebrating the price.