South Korean AI company Upstage shipped Solar Pro 4, a closed commercial LLM, on August 11, 2026. The new generation is no longer pitched as "bigger, stronger, better on general benchmarks" — it is framed around "agents that don't silently fail", with retraining and evaluation rebuilt around long documents, terminal tasks, multi-turn tool use, and refusal-to-fabricate behaviors.

A block of agent-task scores

Solar Pro 4 ships with a 512K context window and a 128K output ceiling, supports English, Korean, and Japanese, and reasons by default. Upstage's blog lists these agent scores from the Artificial Analysis evaluation system:

  • Terminal-Bench v2.1 (multi-step tasks in a live shell): 57
  • τ³-Banking (multi-turn tool use to find the right policy in a knowledge base): 23
  • AA-LCR (reasoning across roughly 100k tokens of long documents): 71

Upstage's own side-by-side against Solar Open 2, on the same evaluation environment: Terminal-Bench v2.1 +13.8, BrowseComp +11.9, GDPval-AA v2 +7.4, AA-LCR +8.3. On GPQA Diamond, MMLU-Pro, LiveCodeBench, AIME 2026, the two models are roughly even. Artificial Analysis gives Solar Pro 4 a composite index of 42 — Upstage describes this as more than triple Solar Pro 3 and ahead of Nvidia's Nemotron 3 Ultra at 38 and Google's Gemini 3.5 Flash-Lite at 37.

In an August 20 piece in The New Stack, Upstage's US CEO Kasey Roh summarized the pitch: "Save the frontier models for the frontier problems; we built the workhorse." Her concrete cost comparison: a document fact-checking task running 300K input plus 15K output tokens costs "roughly $1 per task on premium frontier pricing versus $0.10 on Solar Pro 4."

Training core: synthesizing finished work with OfficeVerse

Solar Pro 4 is trained on work synthesized by Upstage's in-house OfficeVerse pipeline. OfficeVerse generates 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. Upstage states that Solar Pro 4 was trained and validated on work in the same shape it takes in the real world. The same pipeline produced Ko-GDPval, Upstage's Korean office-work benchmark.

This is the philosophical opposite of benchmark-focused training. Many closed models top public leaderboards while still failing in production on documents: "the model handles rows 1 through 200 just fine, then somewhere past that point it starts skimming — dropping rows, or silently filling a cell by pattern-matching from earlier rows instead of reading the actual value." Roh named this failure mode in The New Stack interview: "That failure mode is precisely what we trained against."

"Cannot verify" is the headline deliverable of this generation

Solar Pro 4 draws three explicit verdicts when answering document questions: grounded (the clause exists and gets a citation), not-in-document (the clause does not exist), mismatch (two documents differ on the same number — flagged as a discrepancy). When the evidence is missing, the model says "cannot verify" rather than producing a language-shaped answer from a learned prior.

For regulated industries — financial services, insurance, manufacturing, supply chain — this matters more than another leaderboard tick. Once a model invents a number or clause that isn't in the source document, it flows downstream unchanged, and eventually someone reconciles it manually. Solar Pro 4's design choice is to make traceability the default.

Pricing and availability

Solar Pro 4 ships on Upstage Console, OpenRouter, Hermes Agent (Nous Research's agent environment), and Upstage Studio. Pricing: $0.30 per 1M input tokens, $0.06 per 1M cached input tokens, $1.20 per 1M output tokens. Upstage runs a launch promotion at 90% off list price from August 11 through September 10, 23:59 UTC.

The New Stack cites Upstage's figures: Solar Pro 4 crossed 370 billion tokens consumed on OpenRouter within a week of listing. Upstage's existing partners include AWS and AMD; the company opened a San Jose US headquarters in 2025.

A take

Solar Pro 4 is not pitched as "bigger than Sonnet or GPT" or "the next round of model capability arms race." It is a bet that, after several rounds of leaderboard saturation on SWE-Bench, Terminal-Bench, and AA-LCR, what actually decides whether an agent runs in production is whether the model can hold accuracy through hundreds of tabular rows, dozens of contract pages, and hundreds of tool calls without slipping into silent failure. That is an engineering problem, not a scaling problem. For developers everywhere — Chinese, Korean, US, European — Solar Pro 4 is a useful signal to watch: the next constraint in agent deployment is not parameter count but silent-failure rate.