Upstage releases Solar Pro 4, scoring 42 on Intelligence Index

Upstage’s latest reasoning model, Solar Pro 4, just landed with a score of 42 on the Artificial Analysis Intelligence Index. For context, its predecessor, Solar Pro 3, scored 14. That’s a 3x jump in a single generation.

The South Korean AI startup, founded in 2020, is positioning Solar Pro 4 as a purpose-built tool for enterprise environments drowning in documents, multi-step workflows, and complex reasoning tasks.

The numbers behind the upgrade

Solar Pro 4’s benchmark performance tells a story of broad improvement rather than a single flashy headline number. On Terminal-Bench v2.1, a test designed to evaluate agentic task execution, the model scored 57%. Its long-context reasoning score (AA-LCR) hit 71%, reflecting better performance when processing lengthy documents. Multi-turn tool use, measured by the τ³-Banking benchmark, came in at 23%.

The most eye-catching figure might be the GDPval-AA v2 Elo rating for real-world work tasks: 1,277. The human baseline on that same evaluation sits at 1,000, meaning Solar Pro 4 is measurably outperforming the average human worker on the specific task categories being tested.

The model supports a context window of roughly 524K tokens. Solar Pro 4 also includes what Upstage calls a “reasoning-effort dial,” letting users balance depth of analysis against speed. Solar Pro 4 takes an average of 8.6 minutes per Intelligence Index task, compared to 6.0 minutes for Pro 3.

Solar Pro 4 reportedly handles hallucinations better by simply abstaining from generating a response when it lacks sufficient evidence.

Aggressive pricing and distribution

Upstage launched Solar Pro 4 with a promotional discount of 90% on API pricing, running until September 10, 2026. During the promotional window, input tokens cost $0.30 per million and output tokens run $1.20 per million.

The model is available through Upstage’s own API, OpenRouter, and through Nous Research’s Hermes Agent platform.

The language support is also worth noting: English, Korean, and Japanese.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.