Upstage (Solar)
by Upstage
Getting-started guide
Turns messy PDFs and scans into text that software can read — priced by the page, and built for people who write code.
Last reviewed 2026-09-09 · for Upstage (Solar)
You sign in to Upstage's console in a browser and, by Upstage's own description, drop a PDF or scan into its playground and press run. What returns is the same document as structured text, with headings, tables and paragraphs marked up. Adjust an option, run it once more, then download what comes back.
Two products share one name here. Solar is Upstage's family of models — the AI "brains" that do the actual thinking — and it is not the part of Upstage that reads your documents. That work belongs to separate products called Document Parse, Document OCR and Information Extract, which take a PDF, a photographed form or a slide deck and hand back clean structured text.
Upstage is a Korean company, now one of Korea's larger AI firms, and it sells both halves the same way: as an API — a way for programs to talk to each other without a person clicking — charged per page for documents and per chunk of text for Solar. There is no consumer app and no monthly subscription.
It also publishes some of its models as files anyone can download and run on their own hardware, under a licence Upstage wrote itself. The document engines are not published that way; those you rent.
A layer called Upstage Studio sits on top, where you assemble the document steps into a workflow from a template rather than by writing code.
Upstage is sold to developers, and you will get further with it if you have someone who can call an API. That is a fact about the product, not about you — if what you actually want is to ask questions of one PDF and read the answers in a browser, Adobe's Acrobat AI Assistant on this site does that, and does not involve an account with a spending balance.
Signup is self-serve — no sales call, no demo booking. Upstage's console sign-in page currently advertises free credit for new accounts; the pricing page doesn't repeat that offer and no expiry is published anywhere public, so treat it as current rather than guaranteed. Separately, the pricing page says every Studio agent includes ten free runs with no credit card required.
Beyond that you're buying credits — the platform's unit of spend, drawn down as you go. Document Parse is priced per page, and the enhanced setting costs about three times the standard one; plain text extraction is cheaper still, and pulling named fields out of a form costs more. The sign-up credit covers more ordinary pages than most people trying this out will get through.
The console, the documentation, the pricing page and the help centre are all in English and prices are quoted in US dollars. No regional restriction on signup is stated on any public page — though that is an absence of a rule rather than a published promise.
Every document you send leaves your machine and is processed on Upstage's infrastructure, which its privacy policy says runs on US cloud providers. What happens to it next depends on which tier you are on — but check the date on the terms before you rely on any of this. The version that draws the line sharply does not bind anyone yet. Upstage's terms of service take a new version on 21 September 2026, and that version's Article 22 says that for paid services Upstage will not use your inputs or outputs for service improvement or model training without separate consent, while for free services the data may be used for service improvement and for AI research and development, training included; it also tells free-service users not to input personal data at all. What governs a reader today is the version before it, which Upstage dates 1 July to 20 September 2026, and we could not retrieve its text from Upstage's public pages to confirm whether it splits free from paid the same way. Assume the free tier may be trained on until you have read the version in force on the day you sign up.
.hwp and .hwpx, the Korean word-processor formats almost nothing outside Korea handlesUpload a crooked phone photo of a printed form or receipt with a stamp across it, then compare the table it returns against the original page.
Why this one: Document Parse ships against exactly these cases: Upstage's published fix list names 90 and 180 degree rotations, tilted scans, and tables that continue across a page break merged back into one.
Send a page thick with tables through the standard setting, then the same page through the enhanced setting, and put the two outputs side by side.
Why this one: The tripling from standard to enhanced is one Upstage publishes but never explains in words, so four cents of your own documents buys the answer cheaper than guessing wrong across ten thousand pages.
Tweak: Do it again on a plain page of body text. If the two outputs are identical there, you've learned the setting only matters for a subset of your documents.
Run one scanned page through Document OCR, then the same page through Document Parse, and read both outputs.
Why this one: Upstage prices these nearly seven times apart because one returns the words and the other returns the words plus where they sat. Most people over-buy here.
Tweak: Use a page whose meaning lives in its layout — an invoice, a lab result, a timetable. That is where the cheap option falls down.
Open Upstage Studio's agent library, choose a template matching your document type, and run it against ten of your own pages.
Why this one: The ten free runs per agent make this the only part of the platform you can judge before payment details exist anywhere. Studio bills as the sum of the steps an agent runs, so the same test shows what one real workflow costs per document.
Tweak: Note which steps the agent chained together. That list is the shape of the job you would otherwise be writing code to do.
Take the structured output from a parsed invoice, paste it into the chat AI you already use, and ask it for the total, the due date and the vendor's tax number.
Why this one: Document Parse's stated output formats are HTML and Markdown, chosen specifically so a language model can read them, and pasting the result somewhere neutral is the fastest test of that claim. Because the models and the parser are sold separately, it also tells you whether you need Upstage's models at all.
Tweak: Ask a question whose answer lives inside a table cell rather than a sentence. That is where a bad parse shows up.
Want the full picture first? See the Upstage (Solar) overview for pricing, the free tier and related tools.