Solar Pro 4: Upstage's Agent Model for Korean, English, Japanese

Most model comparisons are written from an English-first point of view, and Upstage's Solar line does not fit that frame well. Solar Pro 4 is pitched as an agent model, but its distinguishing feature is that it is built for Korean, English and Japanese as first-class languages. If your users work in those languages, it belongs on your shortlist for reasons a general leaderboard will not show.

This guide covers what Upstage has published about the model, what it has not, how the price sits next to its siblings, and how to test it on your own work.

What Upstage has disclosed

Upstage released Solar Pro 4 on 6 August 2026 and calls it its flagship model, specialised for agentic use. The model page lists:

• A 512K-token context window and up to 128K output tokens.
• Chat, reasoning mode, structured outputs and tool calling. Catalogue listings also show parallel tool calls and reasoning-effort control.
• Korean, English and Japanese for both input and output.
• A February 2026 training data cut-off.
• Recommended use in document-based work, coding and business tasks, with a focus on agentic workflows.
• Availability through Upstage's console and API, with on-premises deployment by arrangement.

Upstage also states that API input is not used for training and is not stored unless service delivery requires it. For teams that need data handling in writing, that statement and the on-premises option are part of the product, not footnotes.

What has not been disclosed

The parameter count is not published, and neither is the architecture. That is a change from Solar Pro 3, which Upstage described as a 102B mixture-of-experts with 12B active. Upstage's Hugging Face organisation carries a Solar Pro 4 tokenizer but not the weights, so this is an API model; Upstage's recent open-weight releases are a separate line called Solar Open.

The model page lists no benchmark results either. Upstage says Pro 4 improves agent performance over Pro 3, which is a direction rather than a figure. Without vendor numbers or independent results to anchor on, your own evaluation carries all the weight. When numbers do appear, check whether they were run by Upstage on its own scaffold or by an independent evaluator.

Three languages, and why that is the real pitch

A model that is good in Korean is not just a model that knows more Korean words. Tokenisation, register, honorifics and domain vocabulary all affect whether output reads as native or as translated. Korean has been a core language across Upstage's recent Solar API models, and Pro 4 keeps Japanese alongside it.

There is also a cost angle. The same content can tokenise very differently across languages, so the effective price per document depends on how efficiently a model's tokenizer handles your language. Token costs by language explains how to measure this. Upstage publishes the tokenizer, so you can count tokens on your own Korean or Japanese text before spending anything on inference.

If your workload is English only, the language focus gives you no particular reason to choose Pro 4, and it has to compete on agent quality and price with a much larger field.

512K in, 128K out

The window is large enough to hold a substantial document set or a long agent transcript in one prompt. The 128K output cap is the more unusual figure: it allows long generated documents and long reasoning traces without truncation, which matters for report drafting and for agent steps that produce large patches.

Long output is also where cost goes. Output bills at four times the input rate here, and reasoning tokens count as output. A loop that lets the model write freely can spend far more than its input suggests. Set max_tokens per step to what the step actually needs, and keep reasoning off where it does not change the result.

Pricing next to its siblings

Upstage's model page lists Solar Pro 4 at $0.30 per million input tokens, $1.20 per million output and $0.06 per million cached input. Its pricing page runs promotional rates on top: $0.09 in and $0.36 out at the time of writing, which is also what OpenRouter lists, then $0.15 in and $0.60 out from 20 October with no end date shown. Check the rate on the route you actually use and confirm it against the cost reported in each response.

• Solar Mini 4: $0.10 in, $0.40 out, same 512K window and 128K output, 35B total with 3B active. See the Solar Mini 4 guide for when the smaller model is enough.
• Solar Pro 3: $0.15 in, $0.60 out, 128K window, retiring on 30 October 2026. The Solar Pro 3 guide has a migration checklist.

At list prices Pro 4 is three times Mini 4's price on both input and output; on the promotional rates the gap is narrower, so recheck it when you budget. The premium is worth paying when a failed run costs more than the tokens: long unattended loops, multi-document reasoning, or code that has to be right first time. It is not worth paying for classification, extraction or routing, where the smaller model is the sensible default.

How to test it on your own work

Since there are no published numbers to lean on, build the evidence yourself:

1. Collect twenty tasks from real work, with known good outcomes. Include at least a quarter in Korean or Japanese if those languages matter to you.
2. Run each on Pro 4 with reasoning off and on, and on Mini 4, on the same harness and the same tools.
3. For agent tasks, record turns to completion and total tokens, not only pass or fail.
4. Have a native speaker judge the non-English outputs blind, alongside a strong general model you already use.
5. Pin solar-pro4-260806 for the run, so the result describes a fixed model.

How to benchmark LLMs on your own work has the harness setup. If Pro 4 wins on your non-English tasks and ties elsewhere, route by language rather than choosing one model for everything.