Upstage (Solar) (Upstage (Solar))

💰 价格与方案

🔧 API 与开发者体验

• •OpenAI-Compatible Base URL: 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 migration is the base URL plus the model string — no new SDK, no bespoke request format.
• •Reasoning by Default: Solar Pro 4 reasons by default and the response carries the reasoning trace, so the mental model is closer to a reasoning model than a classic completion endpoint. Handle the trace in your client.
• •Adjustable Reasoning Effort: Response speed is a dial rather than a fixed property: high reasoning effort favours deep analysis, low targets real-time interaction at what Upstage calls everyday-chatbot speed. For agent loops calling the model dozens of times per job, this is the main cost and latency lever.
• •Context Window: 512K-token input context with up to 128K output tokens — enough to load several contracts, reports, and data files into a single session without splitting them. OpenRouter lists the context at 524,288.
• •Languages: English, Korean, and Japanese for both input and output. Japanese-specific instruction-following (IFEval JA 81.6) is marginally behind the open-weights sibling (82.5).
• •Distribution: Upstage Console (primary API), SolarChat (free browser playground, no key), OpenRouter (upstage/solar-pro4), Hermes Agent, Upstage Studio (no-code document agents), and dedicated / on-premises deployment.
• •Compliance & Procurement: Upstage operates under SOC 2 and ISO 27001 certification; models are also available on AWS, Azure, and Snowflake marketplaces for buyers who procure through a cloud account rather than a new vendor relationship.

📊 基准测试（Solar Pro 4 对 Solar Open 2）

🎯 适用场景与结论

• •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 removes the need to split document sets, and the model's trained habit of flagging what it cannot verify is exactly the property you want where a hallucinated number is expensive.
• •Be careful if…: your work is math-heavy or Japanese-instruction-sensitive — AIME 2026 and Japanese IFEval are essentially flat or slightly down versus Solar Open 2, so the agentic uplift does not generalise automatically. And if you need to self-host for data-residency reasons, Solar Pro 4 is the wrong choice by design; look at Solar Open 2.
• •Watch the calendar: The 90%-off phase ends September 10, 2026 at 23:59 UTC and steps down to 70% off on September 11 before reverting to standard pricing on October 11. The same tokens cost roughly 3x more three days later and 10x more five weeks later. That is the published price schedule, not artificial urgency.
• •Work agents & cookbook: Upstage shipped seven work agents alongside the model (Excel workbook generation, evidence-based Word reports, PowerPoint decks with speaker notes), each with its system prompt, actual outputs, and pass criteria, plus a full English worked assignment so you can run the same job yourself rather than trusting a demo.