Last verified
KOREA'S FIRST AI UNICORNFOUNDED 2020SOLAR SERIESOPENAI-COMPATIBLE APIDOCUMENT AI

Upstage Solar API Pricing

The South Korean lab behind the Solar models, founded in October 2020 by Sung Kim — a former HKUST professor who led Naver's Clova team. Upstage pairs a compact, cost-efficient LLM line with a strong document-AI business, and in 2026 became Korea's first generative-AI unicorn.

Models priced here
3
Solar Pro 4 / 3 / 2
Founded
2020
Yongin, South Korea
Flagship list
$0.30/M
Solar Pro 4 input
Charged today
$0.03/M
90% off to Sep 10
Valuation
>$679M
Series C-1, 2026
Cached input
-80%
Solar Pro 4 cache hit
§ 01 / LINEUP

The full roster.

Side-by-side →
§ 02 / SHELF

All side-by-side.

Methodology →
Model Input /M Output /M Cached Context Max output Vision Tools Tier
Solar Pro 4 FLAGSHIP $0.30 $1.20 $0.06−80% Frontier
Solar Pro 3 $0.15 $0.60 $0.015−90% Active
Solar Pro 2 LEGACY $0.15 $0.60 $0.015−90% Legacy
§ 03 / PRICE CURVE

Pricing across the lineup.

How Upstage priced 3 models · AUG 26.

oldest → newest →
$2$1.5$1$0.5 $0.6 $0.15 $0.6 $0.15 $1.2 $0.3 Solar Pro 2 AUG 26 Solar Pro 3 AUG 26 Solar Pro 4 AUG 26
Input · newest $0.3/M
Output · newest $1.2/M
Each point is a model at its listed $/M price.

The Upstage story so far.

Company milestones and the Solar line. Sourced from Upstage's own newsroom, its pricing and docs pages, and Wikipedia.

AUG 16 · 2026
AI//COST adds Upstage — Solar Pro 4, Pro 3 and Pro 2 priced from Upstage's own rate card, with the flagship's 90% launch discount recorded against its list price
PRICING
AUG · 2026
Solar Pro 4 ships as an agentic model for tool calling and terminal work, reasoning on by default, launched at 90% off until Sept 10
RELEASE
JUL · 2026
Solar Open 2 announced — an open-weight agent model, with an Upstage AI agent embedded in the Daum portal
RELEASE
2026
Korea's first generative-AI unicorn — Series C-1 of ₩180B (~$122.3M) led by Sequoia Capital, with Hyundai Motor and Kia joining, pushing valuation past ₩1 trillion
CORPORATE
2024
$72M Series B raised to expand globally — at the time the most-funded South Korean AI software company
CORPORATE
DEC · 2023
Solar 10.7B tops the Hugging Face Open LLM Leaderboard, built with Upstage's Depth-Up Scaling technique
RELEASE
OCT · 2020
Upstage founded in Yongin by Sung Kim, previously an HKUST professor and head of Naver's Clova team
CORPORATE
§ 04 / ACCESS

Where to get it.

Methodology →
§ 04 / BEST FOR

Which Upstage for what.

More scenarios →
If you want the agentic flagship while it is 90% off
Solar Pro 4
Profile →
If you are budgeting for after September 11, 2026
Solar Pro 3
Profile →
If you need document parsing and OCR rather than chat…
Document AI, priced separately
Upstage pricing →
If you need a published context window before you commit…
Look elsewhere for now
Compare models →
§ 06 / BACKGROUND

The company behind it.

www.upstage.ai →

Upstage is a South Korean AI company founded in October 2020 by Sung Kim, a former professor at the Hong Kong University of Science and Technology who previously led Naver's Clova team. It is headquartered in Yongin, south of Seoul.

The company made its name twice over. First with document AI — parsing, OCR and extraction — which remains a large part of the business and is priced separately from the chat models. Then with Solar: in December 2023 Solar 10.7B topped the Hugging Face Open LLM Leaderboard, built using Upstage's own Depth-Up Scaling (DUS) technique, which grows a model by duplicating and re-stacking layers rather than training a larger one from scratch. The result was a compact model punching well above its parameter count — which is still the line's pitch.

Funding has escalated quickly: a $72M Series B in 2024 made it the most-funded South Korean AI software company at the time, and a ₩180 billion (~$122.3M) Series C-1 led by Sequoia Capital — with Hyundai Motor and Kia participating — pushed its valuation past ₩1 trillion and made it Korea's first generative-AI unicorn.

For buyers, the practical picture is this. The API is OpenAI-compatible, so trying it costs a base-URL change. Prices are low against Western frontier labs and roughly in the range of the cheaper Chinese labs, without the procurement questions a China-based vendor raises for some teams. The catch worth naming: Upstage does not publish a context window for the Solar models on any first-party page we could find, and published rate limits are modest at 100 requests/min — both matter more than the headline price if you are sizing a production workload.

§ 07 / COMPETITORS

Other frontier labs.

All providers →
PRICE PEER

Zhipu (Z.ai / GLM)

Similar price band with a far broader lineup, free tiers and open weights. But it is China-based and on the US Entity List, which is a procurement question Upstage does not raise.

$0.60/M GLM-4.7
CHEAPEST PEER

DeepSeek

Cheaper still off-peak and MIT open weights, but it moved to peak/off-peak billing in August 2026, so its cost now depends on the hour. Upstage's rate is flat.

Peak / off-peak
FRONTIER LEADER

Anthropic

Leads coding and reasoning by a wide margin, with published context windows and mature enterprise terms. But an order of magnitude more expensive per token.

Claude lineup
OPEN-WEIGHT PEER

Mistral

EU-native open weights with GDPR-first residency — the closest analogue to Upstage's compact-model pitch outside Asia, with a wider catalogue.

EU residency
OPEN-WEIGHT GIANT

Alibaba (Qwen)

The Qwen (Tongyi Qianwen) model family from Alibaba Cloud's Tongyi Lab. First released in 2023, Qwen is the most-downloaded open-weight model family in the world — most tiers ship under Apache 2.0 on Hugging Face and ModelScope, while the proprietary Max tier is served through Alibaba Cloud's Model Studio.

33 models · from $0.03/M
CHINA CONSUMER LEADER

ByteDance (Doubao)

The Doubao model family from ByteDance — parent of TikTok and Douyin. Served through the Volcano Ark (火山方舟) platform on ByteDance's Volcano Engine cloud, Doubao powers China's most-used consumer AI app and undercuts most frontier labs on price.

14 models · from $0.0223/M

Frequently asked.

Practical questions about Solar pricing, the launch discount, and what Upstage does not publish.

Q · 01 Which Solar model should I start with? +
Solar Pro 4 — it is the agentic flagship and, while the launch discount runs, also the cheapest thing on the page at $0.03/$0.12. Just size your budget on its list price of $0.30/$1.20, because that is what applies from September 11, 2026. If you want a rate with no expiry attached, use Solar Pro 3 at $0.15/$0.60.
Q · 02 Why does the flagship cost less than the mid tier right now? +
Because of the discount, and it is temporary. Upstage prints "90% off until 2026-09-10 (UTC). Standard pricing applies from 2026-09-11 (UTC)" beside Solar Pro 4. Until then the flagship bills $0.03/$0.12 against Solar Pro 3's $0.15/$0.60; on September 11 the ordering flips and Pro 4 becomes twice the price of Pro 3.
Q · 03 What context window do the Solar models have? +
Upstage does not publish one. We checked its pricing page and its API documentation on August 16, 2026 and found no context figure on any first-party surface. Aggregators quote numbers, but those are the hosts' claims rather than the vendor's, so this catalogue shows nothing rather than repeat them. It is the single biggest gap in Upstage's published specs.
Q · 04 How hard is it to migrate to Upstage? +
Mechanically, easy. The API is OpenAI-compatible: point the OpenAI SDK at https://api.upstage.ai/v1 and pass a Solar model name. The real constraint is throughput — published rate limits are 100 requests/min and 50,000 tokens/min, which is modest next to the major labs and worth checking against your peak load before committing.
Q · 05 Are reasoning tokens billed extra? +
They are billed, at the output rate. Solar Pro 4 reasons by default, and Upstage's docs count reasoning tokens inside usage.completion_tokens — the same bucket as the visible answer. The response exposes usage.completion_tokens_details.reasoning_tokens so you can see how much of a bill is thinking.
Q · 06 Is Upstage only an LLM company? +
No — document AI is arguably its core business. Parsing, OCR and information extraction are sold on the same pricing page under a separate tab, with their own units, and are not covered by the per-token rates here. The Embed model, meanwhile, is scheduled for end of service on December 31, 2026, so do not build new retrieval on it.