Korean AI lab Upstage has released Solar Pro 4, scoring 42 on the Artificial Ana

Upstage's Solar Pro 4 improved most on agent and long-context work, and its knowledge score rose because it now declines more questions, according to Artificial Analysis's evaluation. The Korean lab's new proprietary reasoning model replaces Solar Pro 3 from April 2026 and scores 42 on the Artificial Analysis Intelligence Index, up from Solar Pro 3's 14. The evaluator says the largest gains are on agentic and long-context tests: Terminal-Bench v2.1 rose from 12% to 57%, the long-context test AA-LCR from 31% to 71%, and τ³-Banking from 9% to 23%. On GDPval-AA v2, which rates real-world work tasks against a human baseline of 1000 Elo, Solar Pro 3 scored 498 and Solar Pro 4 scores 1276.

The knowledge result moved for a different reason. Solar Pro 4's score on AA-Omniscience, a knowledge test whose score can fall below zero, went from -53 to -1, and Artificial Analysis attributes that to abstaining: the model attempts 41% of questions, against 92% for Solar Pro 3. Its hallucination rate fell from 88% to 24%, still higher than Command A+ (14%) and MiniMax-M3 (18%), while accuracy stayed at 19%. The model makes far fewer wrong guesses but answers no more questions correctly. Upstage's own launch post presents that behavior as intended, saying the model is built to "stop and say so when the evidence runs out."

The gains cost more time and money. Through Upstage's own API, prices double to $0.30 per million input tokens and $1.20 per million output tokens, from $0.15 and $0.60, and cache hits rise to $0.06 from $0.02. An average Intelligence Index task takes 9.5 minutes against 6.9 for Solar Pro 3, even though Solar Pro 4 writes about 17% fewer output tokens per task (43,000 against 52,000). At that price it trails models that score higher on the same index: Artificial Analysis notes that MiniMax's MiniMax-M3 has the same per-token price and scores 45, and DeepSeek's V4 Flash 0731 scores 52 at $0.14 and $0.28.