Long-document reasoning
Works across roughly 100k tokens of reports and filings in one session.
Solar Pro 4 launch · 90% off through September 10
The agentic model that finishes the job
Solar Pro 4 reads the documents, runs the tools, and produces the deliverable — then stops and says so when the evidence runs out.
512K context128K output tokensReasoning by defaultEnglish · 한국어 · 日本語OpenAI-compatible API
Solar Pro 4 finishes work because it was trained on finished work
Compared with Solar Pro 3, agent capability took a major step up — and the gains are largest on evaluations that resemble real work: long documents, terminal tasks, and multi-turn tool use.
Terminal tasks
Completes multi-step jobs in a live shell instead of only generating commands.
Multi-turn tool use
Finds the right policy in a large knowledge base and acts on it across a tool-calling conversation.
Answers on evidence
Says it cannot verify when a clause or number is not in the documents.
OfficeVerse, Upstage’s pipeline since Solar Open 2, synthesizes office tasks from real public data across 11 industry domains and 12 task types, grading each pass or fail on the final deliverable. Ko-GDPval, the Korean office-work benchmark, came from the same effort.
Strongest results where real work gets done
Solar Pro 4’s strongest results land on the evaluations that decide whether real work gets done.
Terminal-Bench v2.1
Completes multi-step jobs in a live shell, not just generating commands.
τ³-Banking
Finds the right policy in a large knowledge base and acts on it across a multi-turn, tool-calling conversation.
AA-LCR
Reasons across ~100k tokens of reports and filings at once, synthesizing answers scattered across multiple documents.
* Scores from Artificial Analysis (artificialanalysis.ai) as of August 2026; public listing of our result upcoming. Each is a several-fold step over Solar Pro 3.
From Excel to report to slides
A high benchmark score means little if the model cannot finish the job. We handed Solar Pro 4 an entire store-location analysis for Solarbean Coffee, a fictional coffee brand.
1
store-opening policy document
6
market-data files
3
prompts to screen ten candidate sites
3
deliverables produced in sequence
We are releasing seven work agents in the same family as the Solar Pro 4 Agent Cookbook — Excel workbook generation, evidence-based Word reports, and PowerPoint decks with charts and speaker notes, each with its system prompt, actual outputs, and pass criteria.
Find what’s there. Don’t invent what isn’t.
The most dangerous answer in real work is not a plainly wrong one — it’s an unverified claim stated as fact.
We gave Solar Pro 4 a fictional service agreement and a quotation, then asked ten questions. Half have answers in the documents. The other half are traps: clauses that don’t exist, a premise the contract contradicts, and an amount that differs between the two documents.
Solar Pro 4 is built to say it cannot verify rather than fill the gap. A trustworthy model isn’t the one with an answer every time, but the one that first checks whether an answer has grounds.
The Solar Pro 4 Agent Cookbook research agent links every sentence to a source and labels anything it cannot ground as unverified.
Which one, when
Upstage ships two current models with different jobs.
Solar Open 2Solar Pro 4
How you run itOpen-weights model you deploy yourselfAPI with no GPUs or serving infrastructure
Best fitOn-premises and data that must stay in the buildingLong, complex commercial agent workloads
Agent gap vs Solar Open 2BaselineTerminal +13.8 · BrowseComp +11.9 · GDPval +7.4
Knowledge, math & codingRoughly levelRoughly level
Korean professional evalsStrongSlightly ahead on law, medicine, and office work
Token pricing that fits repeated agent calls
Agent work never costs one call. What matters is the total cost of finishing the task, retries included.
API pricing
Input$0.30 / 1M tokens
Cached Input$0.06 / 1M tokens
Output$1.20 / 1M tokens
Launch promotion
90% off through September 10
To mark the launch, Solar Pro 4 is 90% off on Upstage Console (23:59 UTC) and OpenRouter. Swap the model name and endpoint in your existing API code and you’re running.
A higher completion rate and fewer bad tool calls mean fewer failed runs you pay for twice.
Getting started takes minutes
Solar Pro 4 reasons by default, and the response carries the reasoning trace. If your agent stack is OpenAI-compatible, change the endpoint and the model name — that’s the whole migration.
Upstage Console
Call it over the API — OpenAI-compatible, model name solar-pro4.
SolarChat
Try it in the browser. No install and no API key.
OpenRouter
Connect from the routing setup you already use.
Hermes Agent
Call Solar Pro 4 from Nous Research’s agent environment.
Upstage Studio
The no-code document agents run on Solar Pro 4.
Dedicated / on-premises
For organizations that keep data inside their own network.
Upstage operates Solar Pro 4 under SOC 2 and ISO 27001 certification.
Solar Pro 4 FAQ
Quick answers on capabilities, migration, and pricing.
What makes Solar Pro 4 different?+
Solar Pro 4 is built for agent work. It reads long documents, runs terminal tasks, chains multi-turn tool calls, and stops to say it cannot verify when the evidence runs out.
How much context does Solar Pro 4 support?+
Solar Pro 4 supports a 512K context with up to 128K output tokens, and handles English, Korean, and Japanese for both input and output.
How do I migrate from my current model?+
The endpoint is OpenAI-compatible. Change the endpoint and the model name (solar-pro4) in your existing API code — that is the whole migration.
How much does it cost?+
Solar Pro 4 is $0.30 per 1M input tokens, $0.06 per 1M cached input tokens, and $1.20 per 1M output tokens. To mark the launch, it is 90% off through September 10.
Solar Open 2 or Solar Pro 4 — which should I pick?+
Pick Solar Open 2 when you need a general open-weights model to deploy yourself. Pick Solar Pro 4 when you want to hand longer, more complex work to an API.
Watch whether it finishes the job
Don’t just count how many answers Solar Pro 4 gets right. Try it on a multi-step job and see whether the output can be handed to the next step without re-checking.