Solar Pro 4: Upstage's Flagship Model and the Pricing Clock Still Running On It

Solar Pro 4 is the flagship text model from Upstage, the Korean AI company behind the wider Solar line, and the most useful thing to know about it is not on its spec sheet. It is that the price you pay for it this week is a promotion with an expiry date. Upstage's own pricing page shows Solar Pro 4 running at $0.09 per million input tokens and $0.36 per million output tokens through October 20, 2026, after which it steps to $0.15 and $0.60. The model itself is not new — Upstage released it on August 6, 2026, and wrote it up on its blog the following week. What follows is the reference entry for it: what Solar Pro 4 is, how it differs from the two neighbours it gets confused with, Solar Mini 4 and Solar Open 2, what Upstage does and does not publish about its size and licence, and what independent measurement says as distinct from what the launch material claims.

What Solar Pro 4 is, in the vendor's own terms

Upstage's model card describes Solar Pro 4 as its flagship model, "specialized for agentic use", particularly strong at document-based work, coding, and business-related tasks, with what the version notes call "substantially improved agent performance compared to Solar Pro 3". That is the vendor's framing and it is worth taking at face value as framing: it tells you which workloads Upstage built the model for, not how well it performs on them.

• Vendor — Upstage, the Korean AI company that publishes the Solar family
• Model identifier — solar-pro4-260806, with the alias solar-pro4
• Released — 2026-08-06
• Stated position — flagship; the card's own phrase is "specialized for agentic use"

The model identifier is worth a second look, because Upstage bakes the release date into it: 260806 is 6 August 2026. The company uses the same convention across the line, which means the age of any Solar model is readable from its name rather than from a changelog. That is a small convenience, and it is the kind of thing a reference page should record once so nobody has to reconstruct it later.

Where Solar Pro 4 sits in the Solar line

The Solar family is not one model in three sizes. It is three different bargains, and the axis that separates them is licence and size rather than benchmark position. Solar Open 2 is the member with published weights — a 250-billion-parameter mixture-of-experts you can take away and run yourself. Solar Mini 4 is the small, cheap tier. Solar Pro 4 is the flagship, and it is the one you cannot have.

• Solar Pro 4 — flagship, closed: no published parameter count, no published weights, list pricing of $0.30 per million input tokens and $1.20 output
• Solar Open 2 — the open-weights member of the family, published as a 250-billion-parameter mixture-of-experts, which is the one you can host yourself
• Solar Mini 4 — the small tier, listed at $0.10 per million input tokens and $0.40 output against Pro 4's $0.30 and $1.20

Read those three lines as a placement rather than a scoreboard. The question the family poses is not which model is better; it is whether your workload needs weights you control, a bill you can keep small, or the top of the range — and Upstage has deliberately made those three different products rather than three price points on one product.

There is a deadline attached to the older end of this line. Upstage's pricing page lists Solar Pro 3, Solar Pro 2, and Solar Mini as reaching end of service on 2026-10-30 (KST). If you are still running one of those, the migration decision is being made for you, and the three models above are where the traffic goes.

Size and architecture: what Upstage does not say

The parameter row on Solar Pro 4's model card reads "Undisclosed parameters". There is no architecture paper, no published configuration, no mixture-of-experts ratio, no active-parameter count. The honest answer to "how big is Solar Pro 4" is that the vendor does not publish it, and no amount of reading around the question will produce a number that Upstage has not released.

That absence has practical consequences. You cannot estimate memory requirements from a parameter count, you cannot reason about latency from first principles, and you cannot compare the model to an open-weights alternative on a per-parameter basis. What you can do instead is compare measured behaviour, which is what the later section on benchmarks does. The one classification that does exist comes from Artificial Analysis, whose provider row lists Solar Pro 4 as Proprietary — but that is a licence classification, not a size one, and it should not be read as a parameter estimate.

Licence and weights

Upstage publishes no weights for Solar Pro 4 and no licence text governing them. If a downloadable checkpoint is a requirement in your evaluation, this model does not meet it, and no amount of capability comparison changes that. Solar Open 2 is the family member built for that requirement.

• Weights — not published for Solar Pro 4
• Licence — Upstage publishes none for this model's weights; Artificial Analysis's row lists the model as Proprietary
• On-premises — listed by Upstage as an available platform, which means a commercial deployment arrangement rather than a file you can pull

That last line is the distinction people get wrong. "On-premises" appearing on a model card is a statement about where Upstage will let a paying customer run the model, not a statement that the model is open. Those are different commitments, and only one of them is on offer here.

Context window and output ceiling

The model card gives Solar Pro 4 a 512K context length and up to 128K output tokens. The context figure is the better established of the two, because it is the one an independent source also reports: Artificial Analysis's row for the model lists a 512k context. The 128K output ceiling is vendor-only — the tracker does not carry an output limit in that row, so there is no second source to check it against.

• Context window — 512K, per Upstage's model card, and independently listed as 512k by Artificial Analysis
• Maximum output — up to 128K tokens, vendor-stated with no independent confirmation
• Training data cut-off — February 2026, per the vendor

The cut-off date is the one with the longest tail. A February 2026 training window means anything that happened after that is outside the model's parametric knowledge, which matters most in exactly the agentic and document-heavy work Upstage positions this model for: an agent that reasons well but does not know what changed in the last eight months still needs its retrieval and tool calls to be right.

Modalities and languages

Solar Pro 4 is a text-in, text-out model. The card does not list image, audio, or video input, and the capabilities it does list are chat, structured outputs, tool calling, and reasoning, each marked Supported. A reasoning mode is available, and the independent measurement backs up that it is doing real work rather than being a label: Artificial Analysis times a full response at 28.23 seconds, of which 20.97 seconds is reasoning. Roughly three-quarters of the wall clock is spent thinking rather than emitting tokens.

The language claim is narrower than a lot of readers will assume. Upstage lists English, Korean, and Japanese, each marked as both input and output, and claims no others. For a Korean lab that is a deliberately short list, and it is the line that should decide the question for anyone routing non-English work: if your traffic is in French, Portuguese, or Arabic, this model's card does not cover it.

Vendor claims and independent measurement, kept separate

These are two different kinds of number and they should not be averaged together. Upstage's launch material relays Artificial Analysis figures for three benchmarks — Terminal-Bench v2.1 at 57, τ³-Banking at 23, and AA-LCR at 71. Those come to you through the vendor, dated to August 2026, and they carry the vendor's framing of which evaluations matter. I have not reproduced them.

Read separately, Artificial Analysis's own Upstage page, on October 13, 2026, shows the Solar Pro 4 row reading as follows:

• Intelligence Index — 28, under index version 4.3.2
• Coding Index — 20
• Agentic Index — 20
• Output speed — 95 tokens per second
• Latency — 2.03 seconds to first chunk, 28.23 seconds end to end
• Function calling and JSON mode — both marked supported
• Blended price — $0.22 per million tokens

One figure needs flagging because it has moved. A value of 42 has circulated as this model's Artificial Analysis Intelligence Index. The live row reads 28. If you have seen 42 attributed to Solar Pro 4, that number has been revised, and 28 is what the tracker shows now.

Two further cautions on the same row. Artificial Analysis lists no cost-per-task value for Solar Pro 4, so per-task economics — the figure that matters most for agentic workloads with long reasoning traces — are not independently published. And there is no head-to-head between Solar Pro 4 and the frontier closed models run at matched effort on a single harness, so nothing here supports a ranking claim in either direction.

There is also a mismatch worth naming between the two independent numbers and the vendor's own pricing. Artificial Analysis's blended price of $0.22 does not track Upstage's promotional ladder, so it is not the rate you will pay this week. Treat it as a like-for-like comparison figure against other models on the same tracker, not as a quote.

How you get it, and what it costs

The model card's available-platforms line lists the Upstage Console, the Playground, and On-premises. The API entry states that data is not collected, and Upstage says inputs are not used for training and are not retained beyond what delivering the service requires. The Playground is free to try with no API key, which is what Upstage's own page says today.

On price, the two vendor pages do not agree, and the disagreement is worth understanding rather than resolving. The model card carries a standing rate. The pricing page carries a dated ladder that has been revised as the promotion moved.

• Through October 20, 2026 — $0.09 input / $0.018 cached input / $0.36 output per million tokens
• From October 21, 2026 — $0.15 / $0.03 / $0.60
• Model card standing rate — $0.30 / $0.06 / $1.20, a level the ladder never reaches
• Tax — all prices shown are exclusive of 10% VAT

If a budget depends on this, the pricing page is the one to read, because it is the page carrying the dated steps. If inference reaches you through a router rather than a direct contract, the promo clock matters for a second reason: OrcaRouter passes provider list price through at 0% markup, so a vendor price step lands on your invoice the day the vendor makes it, with nothing to renegotiate.

Can you run Solar Pro 4 on your own hardware?

Not in the sense of downloading it. Upstage publishes no weights for this model, so there is no checkpoint to fetch and nothing to self-host. The on-premises option on the model card is a commercial deployment arrangement with Upstage, which is a different thing from an open release: it means Upstage will run the model somewhere you choose, under terms you agree with Upstage. If self-hosting is the actual requirement, Solar Open 2 is the model in this family that satisfies it.

Why do two Upstage pages show different prices for Solar Pro 4?

Because one is a standing rate card and the other is a promotional schedule that has been revised more than once. The model card's $0.30 and $1.20 is the standard rate; the pricing page carries the dated ladder, which currently runs at $0.09 and $0.36 through October 20, 2026 and then steps to $0.15 and $0.60. Neither page is wrong, but only one of them tells you what changes and when, and that is the one to plan against.

Is Solar Pro 4 actually good at Korean and Japanese, or is that marketing?

The defensible claim is coverage, not quality. Upstage marks English, Korean, and Japanese as input-and-output languages on the model card and claims no others, so the fact of support is settled. How well the model performs in those languages rests on Upstage's own evaluation material, because there is no independent, language-stratified benchmark for Solar Pro 4 that I can point you to. Treat "supports Korean and Japanese" as established and any claim that it leads in them as unproven.

Two dates worth keeping

October 20, 2026, when the promotional rate steps up. October 30, 2026 (KST), when Solar Pro 3, Solar Pro 2, and Solar Mini go end of service. Everything else about the model is consistent across the vendor's card and the independent tracker — a 512K context, text in and text out, three languages, weights you cannot have — and the two questions left open are the ones Upstage has not answered: how large the model is, and under what licence the weights are held. Until those appear, the independent picture is a single row on a single tracker reading 28, and that is the number to use.