Upstage Solar Pro 4 API 2026: 512K Agentic LLM

What Upstage Solar Pro 4 Is in 2026

Upstage is a Korean AI company that has been shipping its own Solar family of large language models since 2023, built around a specific thesis: enterprises do not want a chatbot, they want work finished. Solar Pro 4, launched in early August 2026, is the clearest expression of that thesis so far. It is positioned not as a general reasoning model but as an agentic one — a model you hand a multi-step job to and expect to see the deliverable, not a first draft.

The vendor's own framing is blunt about what that means: reading the documents, running the tools, producing the deliverable, and "stopping to say so when the evidence runs out." That last clause is the interesting one. Upstage's launch post devotes a full section to grounded refusal — the model is trained to answer "not in document" or "discrepancy between sources" rather than inventing a number that flows downstream unchecked. For anyone evaluating a model on document-heavy pipelines, that is a more useful property than another point of benchmark score.

Structurally, Solar Pro 4 is a commercial API model, distinct from Solar Open 2, Upstage's open-weights sibling that organizations deploy on their own hardware. Upstage's guidance is explicit: pick Solar Open 2 if you need a general model you can self-host, pick Solar Pro 4 if you want to hand longer, more complex work to an API. The two share a family lineage but not a deployment model, and confusing them is the most common mistake when reading Solar pricing pages.

Solar Pro 4 Pricing: The Promotion Is the Story Right Now

This is where the article has to be careful, because the price you see depends on the date you load the page. Upstage's pricing page ships a client-side script that holds the standard rates in the card markup and swaps in the active promotional phase. Extracting the figures from the page source rather than the rendered text gives you the full schedule:

So if you are reading this on September 10, 2026, you are on the last day of the 90%-off phase: $0.03 per million input tokens and $0.12 per million output, with cached input at $0.006. Tomorrow the discount steps down to 70% ($0.09 / $0.018 / $0.36), and on October 11 the standard rate of $0.30 / $0.06 / $1.20 applies. The developer documentation banner carried the same message verbatim during the run: "Solar Pro 4: 90% off through Sep 10 (UTC)."

The rest of the Solar line is priced below the flagship and has not moved:

A useful sanity check on the promotional rate: OpenRouter's public model API lists upstage/solar-pro4 with prompt at $0.03 per 1M and completion at $0.12 per 1M, matching the 90%-off phase exactly, alongside a 524,288-token context length. The two sources agreeing on a rate that is 90% below list is itself confirmation that the promotion is real and routed through, not a marketing overlay.

Document AI Pricing: Parse, Extract, Classify

Half of Upstage's product surface is not chat at all — it is document intelligence, billed per page rather than per token. If your workload is invoices, claims, or contracts rather than conversation, this is the pricing that matters:

Upstage bills Studio agents as the sum of their steps, and the pricing page gives worked examples so you can forecast without running a pilot: parsing invoices to structured text is $0.01 per page (Parse only); extracting vendor, amount, and date from receipts is $0.04 per page (Parse $0.01 + Extract $0.03); sorting mixed documents by type before extracting is also $0.04 per page (Parse $0.01 + Classify free + Extract $0.03). Every agent includes 10 free runs, no credit card required.

One billing detail worth pinning down before you build a business case: all prices shown are exclusive of 10% VAT. Korean VAT applies on top, and the commitment tier credits are shared across Console and Studio under one account. The tiers themselves are prepaid, not subscribed — Explore at $100+/month and $1,200+/year, Build at $500+/month and $6,000+/year, Scale at $5,000+/month and $60,000+/year, plus Enterprise at custom pricing. Bonus credits scale with commitment (+10%, +15%, +20% monthly; +30%, +35%, +40% yearly), and rate limits step up with them — Solar Pro 3 runs at 30, 50, or 80 requests/sec depending on tier, and Information Extract at 10, 20, or 40 requests/sec. Rate limits are tiered automatically by monthly purchase volume rather than negotiated.

Benchmarks: Where Solar Pro 4 Actually Moves

Upstage publishes a head-to-head between Solar Pro 4 and Solar Open 2, with asterisked benchmarks run in-house under a shared evaluation environment. The pattern in the numbers is more informative than any single score: the gain is concentrated in exactly the agentic and long-context evaluations the model was positioned for.

Read the third column as a map of intent. Terminal-Bench v2.1 at 57.0 means the model completes multi-step jobs in a live shell rather than merely emitting commands. AA-LCR at 71.0 means it reasons across roughly 100,000 tokens of reports and filings in one pass. t³-Banking at 23.0 measures finding the right policy in a large knowledge base and acting on it correctly across a multi-turn tool-calling conversation. Where the model does not move — AIME 2026 is fractionally behind Solar Open 2, and Japanese instruction-following is 0.9 behind — is equally worth knowing if your workload happens to be math-heavy or Japanese-instruction-sensitive.

Upstage attributes the agentic gains to OfficeVerse, its training pipeline since Solar Open 2, which synthesizes office tasks from real public data across 11 industry domains and 12 task types and grades each one pass or fail on the final deliverable — not on intermediate text quality. Ko-GDPval, its Korean office-work benchmark, came from the same effort. This is the difference between training against a rubric of "did the job get done" and training against a rubric of "does the output look plausible," and it is the most defensible explanation for why the terminal and long-document numbers moved so much more than MMLU-Pro did.

API and Developer Experience

The migration story is deliberately boring, which is a compliment. Solar Pro 4 is exposed through an OpenAI-compatible chat completions interface under the model name solar-pro4. If your stack already speaks the OpenAI protocol, the entire change is the base URL and the model string — there is no new SDK to learn, no bespoke request format, and no proprietary auth dance.

Two behavioural details matter more than the compatibility claim. First, Solar Pro 4 reasons by default, and the response carries the reasoning trace — so the mental model is closer to a reasoning model than to a classic completion endpoint, and you should expect to handle the trace in your client. Second, response speed is a dial rather than a fixed property: setting reasoning effort high favours deep analysis, while low targets real-time interaction at what Upstage describes as everyday-chatbot speed. For agent loops that call the model dozens of times per job, that dial is the single most useful cost and latency lever you have.

Distribution is wider than a single API endpoint. Upstage Console is the primary API surface; SolarChat offers a browser playground with no install and no key; OpenRouter carries it for teams already routing through a third-party gateway; Upstage Studio runs its no-code document agents on the model; and dedicated and on-premises deployment is available for organizations that keep data inside their own network. Upstage states that it operates under SOC 2 and ISO 27001 certification, and the models are purchasable through AWS, Azure, and Snowflake marketplaces — which matters if your procurement path runs through a cloud account rather than a new vendor relationship.

The Agent Cookbook and Work Agents

Alongside the model, Upstage shipped seven work agents in the same family as its Solar Pro 4 Agent Cookbook: Excel workbook generation, evidence-based Word reports, and PowerPoint decks with charts and speaker notes. Each ships with its system prompt, actual outputs, and pass criteria, and the cookbook includes a full English-language worked assignment — a fictional Austin site-selection review with company standards, raw data, and pass criteria — so you can run the same job yourself rather than taking a demo on faith.

The reference assignment is a useful benchmark of ambition. Given one store-opening policy document and six market-data files, the model screened ten candidate sites against the policy and produced three deliverables in sequence — an Excel workbook with PASS/FAIL marked by conditional formatting across five sheets, a review report, and a slide deck — in three prompts. The stated design goal is that a number computed in the workbook carries into the report unchanged. That continuity between deliverables, rather than any single artefact's quality, is what "agentic" is being asked to mean here.

Verdict: Who Should Use Solar Pro 4

Choose Solar Pro 4 if your workload is genuinely multi-step and document-heavy: contract review across several files, reconciling figures between reports, screening candidates against a policy, or any pipeline where a wrong value propagates into a downstream deliverable. The 512K context lets you load several contracts and data files into one session without splitting them, and the model's trained habit of flagging what it cannot verify is precisely the property you want where a hallucinated number is expensive. The promotional pricing makes the next few weeks an unusually cheap window to run a real evaluation on your own documents.

Be careful if your work is math-heavy or Japanese-instruction-sensitive — the benchmark table shows AIME 2026 and Japanese IFEval essentially flat or slightly down versus Solar Open 2, so the agentic uplift does not generalise automatically. And if you need a model to self-host for data-residency reasons, Solar Pro 4 is the wrong choice by design; Solar Open 2 is the one to look at.

Watch the calendar. The single most actionable fact in this review is that the 90%-off phase ends September 10, 2026 at 23:59 UTC, and the rate steps down to 70% off on September 11 before reverting to standard pricing on October 11. If you are going to run an evaluation, the same tokens cost roughly 3× more three days from now and 10× more five weeks from now. That is not artificial urgency — it is the published price schedule, and it is the kind of thing a per-million comparison site exists to make visible.

Fact-checked against Upstage's live pricing page, the Solar Pro 4 launch post on upstage.ai/blog, Upstage developer documentation, and OpenRouter's public model API on September 10, 2026.