Upstage Launches Solar Pro 4 AI Model with Strong Coding and Document Processing Capabilities

Upstage officially launched its latest AI model, 'Solar Pro 4', on August 6, specifically optimized for AI agents. This release comes just five months after the company introduced Solar Pro 3, signaling rapid development in their AI offerings.

Solar Pro 4 is a closed model, similar to offerings from OpenAI and Anthropic, meaning its specific parameters are not publicly disclosed. However, it boasts an impressive context length of 512,000 tokens, equivalent to three to four novels. This extended context window allows for stable processing of extremely long-form content, such as complex coding projects and detailed reports. The model supports multiple languages, including Korean, Japanese, and English, and is adept at tasks like chat, reasoning, structured output, and tool invocation, making it particularly suitable for document and coding-related office work.

Technically, Solar Pro 4 leverages Upstage's proprietary 'OfficeVerse' pipeline, which has significantly enhanced its coding and task performance. This methodology involves synthesizing and analyzing office tasks across 11 industries and 12 types of work based on real-world data. As a result, the model shows a dramatic improvement in accuracy on the terminal-bench v2.1 test, jumping from 12% in Solar Pro 3 to 57%. Tests related to tool invocation within conversations (τ³-banking) also saw a substantial increase, rising from 9 to 23. Furthermore, its long-document reasoning capability (AA-LCR) improved from 31 to 71, a roughly 23-fold increase. These enhancements have led to a significant rise in its artificial analysis intelligence score on the Artificial Analysis index, from 14 to 42.

Compared to Upstage's open-weight model, Solar Open 2, Solar Pro 4 generally exhibits superior performance across benchmarks, web-based tasks, and mathematical capabilities. While Solar Open 2 can be downloaded and run on private systems or used as a base for new AI applications, Solar Pro 4, like GPT and Gemini, is accessed online via API for real-time AI functions and automation. Pricing for Solar Pro 4 is also competitive, with API token costs set at $0.3 per million tokens for input and $1.2 per million for output. This pricing is comparable to Chinese AI companies like DeepSeek and significantly lower than models like GPT-5.6 Luna, making it a cost-effective option for long-form content analysis and document processing.