Solar Pro 3
Language Deprecated Full comparison ↗ upstage/solar-pro-3 · by Upstage
· mixture-of-experts
Pricing — 1 offering(s)
Input tokens
- $0.15 / 1M tokens (input) Current 2026-03-23 → 2026-10-28
Output tokens
- $0.60 / 1M tokens (output)
Current
2026-03-23 → 2026-10-28
Same sourcing basis as the input tier.
Cached input tokens
- $0.015 / 1M tokens (cached input) Current 2026-03-23 → 2026-10-28
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Solar Pro 3 fits into a cost-aware routing setup
See how →Capability profile
How Solar Pro 3 rates across core capability dimensions, with the task-level evidence behind each rating.
Toggleable reasoning mode. Upstage reports "Improved Deep Reasoning" in the 260323 update but publishes limited standalone reasoning-benchmark numbers.
Upstage-reported gains in the 260323 update: SWE-bench 14.5 -> 28.6, Terminal-Bench 2 2.2 -> 10.1. Mid-tier; not a coding specialist.
Function calling and structured outputs supported. Upstage reports the Tau2-all agentic benchmark roughly doubling (36.0 -> 72.3) in the 260323 release — a large vendor-reported jump, no independent reproduction.
128,000 tokens (Upstage Console model docs).
Korean is the primary optimisation target — one of the stronger Korean-language models available via a hosted API — with English and Japanese also supported. Weaker outside those three.
12B active parameters (MoE) — fast and cheap to serve relative to its 102B total size.
$0.15 / 1M input, $0.60 / 1M output (Upstage Console; cross-confirmed on OpenRouter). Cheap for a hosted flagship with a first-party API and an on-prem option.
Ratings are estimated — limited independent data is available for this model.
Operator guidance
The default choice when Korean-language quality is a first-order requirement and a first-party hosted API or on-prem deployment is wanted — Upstage's core strength. For English-only work, general flagships and other cheap MoE models (GLM, Qwen, DeepSeek) are stronger on published English benchmarks. Not the same product as Solar Open 2 (open-weights only, no hosted API — see the registry note).
Use cases
- Korean-language production workloads (the model's primary design target) needing a hosted API or on-prem deployment
- Trilingual (Korean / English / Japanese) applications where a single model covering all three at low cost is valuable
- Agentic / function-calling workflows within a Korean-enterprise context, per the 260323 update's stated focus
Limitations
- Capability strength is concentrated in Korean; English/Japanese are supported but not the optimisation target, and other languages are weaker
- Console pricing and the full benchmark table are JS-rendered — a plain fetch of Upstage's docs returns only a marketing shell (see registry entry)
- Agentic/coding benchmark gains for the 260323 release are Upstage-reported, no independent reproduction
- Two dated snapshots share the `solar-pro-3` name; a caller pinning a specific date gets a fixed model, the alias tracks the latest