「Korean Model of Light」 Solar Pro4 Smart Index surges to 42 points, but still can't beat DeepSeek V4Flash

According to TrendForce's Beating Monitor, South Korean AI company Upstage has unveiled its flagship inference model, Solar Pro 4. In the latest Artificial Analysis assessment, its AI score has surged from 14 points in the previous generation Solar Pro 3 to 42 points, with the main improvement coming from the Agent and long-context capabilities.

In particular, Terminal-Bench v2.1 has jumped from 12% to 57%, while long-context inference has risen from 31 to 71. GDPval-AA v2 has also increased from 498 Elo to 1277, surpassing the human baseline of 1000. Solar Pro 4 has become a representative model from South Korea making a splash on the global stage.

However, when faced with DeepSeek V4 Flash 0731, it still falls short. Within the same set of Artificial Analysis AI scores, DeepSeek has scored 52 points, which is 10 points higher than Solar Pro 4. DeepSeek processes approximately 122 tokens per second, whereas Solar Pro 4 handles around 65 tokens, nearly half as efficient as the former.

The price gap is even wider. Solar Pro 4 costs $0.30 and $1.20 per million tokens for input and output, respectively, while DeepSeek only requires $0.14 and $0.28. Solar's output price exceeds that of DeepSeek by four times.

At least based on this comprehensive evaluation set and API cost-effectiveness, the "Light of Korean Models" has not yet caught up to DeepSeek.