Upstage's Solar Pro 4 Jumps 28 Points to Beat Human Agents
Upstage's Solar Pro 4 jumps 28 points on the AI Intelligence Index, with massive gains on agentic tasks, long-context work, and hallucination control
- Upstage releases Solar Pro 4, scoring 42 on the Artificial Analysis Intelligence Index -- a 28-point jump over Solar Pro 3's 14.
- Biggest gains are on agentic tasks: Terminal-Bench jumps from 12% to 57%, long-context reasoning from 31% to 71%, and real-world agent Elo from 498 to 1276 (above the human baseline of 1000).
- Hallucination rate drops from 88% to 24%, but the model achieves this by abstaining on 59% of knowledge questions rather than improving raw accuracy.
- Priced at $0.30/$1.20 per 1M input/output tokens; 90% launch discount ($0.03/$0.12) runs on Upstage Console and OpenRouter through September 10.
- Supports 512K context, 128K output, English/Korean/Japanese, tool calling, and structured output; OpenAI-compatible API endpoint.
- Most benchmark numbers are self-reported -- no independent third-party evaluation has been published yet.
Korean AI lab Upstage has released Solar Pro 4, its new proprietary flagship reasoning model and a significant step up from Solar Pro 3. On the Artificial Analysis Intelligence Index, it scores 42 -- a 28-point jump from Solar Pro 3's 14 -- placing it alongside models like xAI's Inkling (42) and just behind MiMo-V2.5-Pro (43). The headline number is striking, but the real story is in what changed and why it matters for agent builders.
Built on Finished Work, Not Just Benchmarks
Solar Pro 4 is built to carry real work to the finish -- reading documents, running tools, producing deliverables -- and to stop and say so when the evidence runs out. The work you hand an AI rarely ends with a single question; it's reviewing a contract, reconciling numbers across files, and verifying intermediate results before passing them to the next step.
The key training innovation behind this is OfficeVerse, Upstage's data synthesis pipeline. Solar Pro 4 finishes work because it was trained on finished work. OfficeVerse synthesizes office tasks from real public data across 11 industry domains and 12 task types and grades each one pass or fail on the final deliverable. This is a meaningful departure from training on Q&A pairs or synthetic reasoning chains -- the model was optimized against whether the job got done, not whether individual steps looked plausible.
Where the Gains Are Biggest
The improvements are not evenly distributed. Solar Pro 4's biggest wins are on the tasks that define whether an agent is actually useful in production:
- Terminal-Bench v2.1: 57% (up from 12%) -- completing multi-step jobs in a live shell, not just generating commands
- AA-LCR (long-context reasoning): 71% (up from 31%) -- synthesizing answers scattered across ~100K tokens of documents
- GDPval-AA v2 (real-world agentic tasks): Elo of 1276 (up from 498, well below the human baseline of 1000) -- Solar Pro 4 now sits above the human baseline, slightly ahead of Qwen3.7 Max (1272) and MiMo-V2.5-Pro (1266)