South Korean AI lab Upstage published a blog post today (August 13), announcing the launch of the Solar Pro 4 model, a commercial LLM designed for complex AI agent tasks.
The Solar Pro 4 model has a 512K context window and a maximum output length of 128K tokens. It supports loading multiple contracts, reports, and data files in a single session. The official blog post did not disclose information about the number of parameters.
In terms of pricing, ITHome cited the blog post and provided the following information:
Input: $0.3 per 1 million tokens (Note: approximately 2 yuan at the current exchange rate)
Cache Read input: $0.06 per 1 million tokens (approximately 0.41 yuan at the current exchange rate)
Output: $1.2 per 1 million tokens (approximately 8.1 yuan at the current exchange rate)
For agent tasks, Solar Pro 4’s main improvements focus on evaluations that more closely resemble real-world work, including long documents, terminal tasks, and multi-turn tool calls. The three core results listed officially are: a score of 57 on Terminal-Bench v2.1, 23 on τ³-Banking, and 71 on AA-LCR.
Upstage used site-selection analysis for the fictional coffee brand Solarbean Coffee as a demonstration task. The input materials included one store-opening policy document and six market data files.
Solar Pro 4 screened the policies for 10 candidate sites within three prompts and generated three types of deliverables in sequence: an Excel workbook, a review report, and slides.
ITHome reviewed publicly available information and found that the model has appeared on several AI model evaluation platforms. @ArtificialAnlys (Artificial Analysis) is a leading independent international AI benchmarking and analysis organization specializing in unbiased performance and hosting-cost evaluations across multiple areas, including large language models (LLMs), video generation, and AI music.
@arena (Arena.ai) is currently one of the most authoritative platforms in the global AI industry for real-time head-to-head battles and evaluations of large models based on human preferences. It originated with the renowned LMSYS Chatbot Arena team and uses an Elo rating system similar to chess, ranking AI models worldwide through rigorous blind A/B testing by human developers.
The organization pointed out that @upstageai’s Solar Pro 4 is the first model from a South Korean lab to appear on the Agent Arena, Code Arena: WebDev, and Text Arena leaderboards.
On the Agent Arena leaderboard, the Solar Pro 4 model ranks 44th, at the same level as models including MiniMax M2.7 and Nemotron 3 Ultra.
