If your stack already speaks the OpenAI API, using Solar Pro 4 is a two-line change: point the client at the Upstage endpoint and set the model name to solar-pro4. Everything else — tool calling, streaming, reasoning — is what you already send.
What you get from the API
- 512K context with up to 128K output tokens in a single call.
- Reasoning by default: the response carries the reasoning trace, with reasoning effort adjustable from real-time interaction to deep analysis.
- Multi-turn tool calling designed for agent loops, not just single function calls.
- English, Korean, and Japanese for input and output.
First call checklist
- Get an API key from the Upstage Console.
- Point your OpenAI-compatible client at the Upstage base URL.
- Set the model name to
solar-pro4. - Stream if your UI needs it; the response shape is unchanged.
{
"model": "solar-pro4",
"messages": [
{
"role": "user",
"content": "Read the attached policy and tell me which clause applies."
}
]
}Using the long context
Load the whole document set instead of retrieving fragments first: several contracts, filings, or data files can sit in one session. Two habits make the difference:
- Put the stable material first and the question last, so repeated calls hit the cached-input rate ($0.06 per 1M tokens).
- Budget for output. With up to 128K output tokens, a single call can return a full deliverable rather than a summary.
Tool calling in an agent loop
Define tools the same way you do today, then let the model chain calls across turns: read a record, look up a policy, take an action, verify the result. This is the shape the τ³-Banking score measures; the benchmark guide explains what the number means, and Solar Pro 4 vs Solar Pro 3 covers what changes when you migrate an existing agent.
Where to run it
Besides the API, Solar Pro 4 is available through Upstage Console, SolarChat, OpenRouter, Hermes Agent, and Upstage Studio, as well as dedicated or on-premises deployments. To try the model in the browser first, sign in on the home page — new accounts start with free credits. Pricing for production use is on the pricing page.
FAQ
Is the endpoint really OpenAI-compatible?
Yes — the endpoint and the model name are the migration. Existing OpenAI-compatible code runs with those two changes.
How large can a single request be?
A 512K context with up to 128K output tokens; adjust your chunking accordingly. See the model overview for the capability list.
Does the API return reasoning?
Reasoning is on by default and the response carries the trace. If your parser is strict, account for it before you ship.
What does an agent run cost?
It depends on tokens, not calls: input is billed at $0.30 per 1M tokens, cached input at $0.06, and output at $1.20. The pricing guide works through an example.
