Upstage's Solar Pro 4 (solar-pro4) is a proprietary agentic API aimed at multi-document, multi-tool jobs that need to finish. It is a closed API (dedicated / on-prem via sales), OpenAI-compatible at https://api.upstage.ai/v1, and also available on SolarChat, OpenRouter, Hermes Agent, and Upstage Studio.

What is the Solar Pro 4 context window?

512K context and 128K max output. Reasoning is on by default, with an effort dial (high vs low) for latency.

How much does Solar Pro 4 cost?

List price: $0.30 input / $0.06 cached input / $1.20 output per 1M tokens. Launch promo: 90% off on Upstage Console and OpenRouter through Sep 10, 2026 23:59 UTC. Re-check before budgeting past that date.

Which languages does Solar Pro 4 support?

English, Korean, and Japanese (in and out).

Are Solar Pro 4 benchmark scores independently verified?

No. Selected scores are from the vendor table vs Solar Open 2. Asterisked rows (BrowseComp*, SWE-Bench Verified OpenHands*) use Upstage's in-house harness. Independent LLM Stats verification is pending.

Is Solar Pro 4 open weights?

No. Weights are closed. Compare to Solar Open 2 if you need self-host.

Back to blog

Model Release·Agentic

Solar Pro 4: Upstage's Agentic API for Finishable Work

Upstage Solar Pro 4 (solar-pro4): 512K/128K agentic API, $0.30/$0.06 cached/$1.20 per 1M, EN/KO/JA, 90% off through Sep 10, 2026 23:59 UTC. Vendor scores vs Solar Open 2; LLM Stats verification pending.

Sebastian Crossa

Co-Founder @ LLM Stats

Aug 6, 2026·2 min read

At a Glance

Org: Upstage
ID: solar-pro4 (OpenAI-compatible at https://api.upstage.ai/v1)
Type: Closed API (dedicated / on-prem available via sales)
Context / out: 512K / 128K
Languages: EN, KO, JA (in and out)
Reasoning: On by default; effort high vs low for latency
Also on: SolarChat, OpenRouter, Hermes Agent, Upstage Studio

Pricing

Item	Per 1M tokens
Input	$0.30
Cached input	$0.06
Output	$1.20

Promo: 90% off through Sep 10, 2026 23:59 UTC (Console + OpenRouter). Re-check before budgeting past that date.

Selected vendor scores (vs Solar Open 2)

Benchmark	Solar Pro 4	Solar Open 2	Note
Terminal-Bench v2.1	57.0	43.2	Vendor table
τ³-Banking	23.0	18.1	Vendor table
AA-LCR	71.0	62.7	Bare bench name; vendor-cited
GDPval-AA v2	38.8	31.4	Bare bench name; vendor-cited
GPQA Diamond	89.0	85.6	Vendor table
BrowseComp*	49.2	37.3	* in-house
SWE-Bench Verified (OpenHands)*	70.6	69.2	* in-house

Asterisked rows = Upstage in-house harness under the same internal environment. Do not treat as LLM Stats verified.

What's New

Positioned for longer commercial agent workloads vs Solar Open 2 (self-host general model).
Evidence-first stance in demos: grounded / not-in-document / mismatch style answers.
Agent Cookbook + work-agent recipes (Excel / Word / slides) alongside the API.

Caveats

Closed weights; compare to Solar Open 2 if you need self-host.
Several headline agent scores are third-party or in-house; label before promoting.
Promo pricing expires Sep 10, 2026 23:59 UTC.

Sources

https://www.upstage.ai/blog/en/solar-pro-4

Questions

Frequently Asked Questions

Upstage's Solar Pro 4 (solar-pro4) is a proprietary agentic API aimed at multi-document, multi-tool jobs that need to finish. It is a closed API (dedicated / on-prem via sales), OpenAI-compatible at https://api.upstage.ai/v1, and also available on SolarChat, OpenRouter, Hermes Agent, and Upstage Studio.
512K context and 128K max output. Reasoning is on by default, with an effort dial (high vs low) for latency.
List price: $0.30 input / $0.06 cached input / $1.20 output per 1M tokens. Launch promo: 90% off on Upstage Console and OpenRouter through Sep 10, 2026 23:59 UTC. Re-check before budgeting past that date.
English, Korean, and Japanese (in and out).
No. Selected scores are from the vendor table vs Solar Open 2. Asterisked rows (BrowseComp*, SWE-Bench Verified OpenHands*) use Upstage's in-house harness. Independent LLM Stats verification is pending.
No. Weights are closed. Compare to Solar Open 2 if you need self-host.