guide · ai

Solar Pro 4

Solar Pro 4 (Upstage, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.

August 12, 2026 · By Alastair Fraser

--- title: “Solar Pro 4” description: “Solar Pro 4 (Upstage, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-solar-pro-4.png imageAlt: “A retro robot representing Solar Pro 4” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: - name: “Upstage — Solar Pro 4 launch blog” url: “https://www.upstage.ai/blog/en/solar-pro-4” - name: “Artificial Analysis — Upstage Solar Pro 4 benchmarks” url: “https://artificialanalysis.ai/articles/upstage-solar-pro-4” - name: “OpenRouter — upstage/solar-pro4 model page” url: “https://openrouter.ai/upstage/solar-pro4” - name: “Hugging Face — Upstage org page” url: “https://huggingface.co/upstage” - name: “Upstage Console — first-party API access” url: “https://console.upstage.ai” facts: - label: “Vendor” value: “Upstage” - label: “Released” value: “2026” - label: “License” value: “Proprietary API” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-grok-4.6 - abs-model-kimi-k3 - abs-model-lfm2.5-vl-3b - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-muse-spark-1.2 - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.7-flash - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update tags: - solar-pro-4 - ai-models - model-release draft: false --- --- title: “Solar Pro 4” description: “Solar Pro 4 (Upstage, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-solar-pro-4.png imageAlt: “A retro robot representing Solar Pro 4” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: - name: “Upstage — Solar Pro 4 launch blog” url: “https://www.upstage.ai/blog/en/solar-pro-4” - name: “Artificial Analysis — Upstage Solar Pro 4 benchmarks” url: “https://artificialanalysis.ai/articles/upstage-solar-pro-4” - name: “OpenRouter — upstage/solar-pro4 model page” url: “https://openrouter.ai/upstage/solar-pro4” - name: “Hugging Face — Upstage org page” url: “https://huggingface.co/upstage” - name: “Upstage Console — first-party API access” url: “https://console.upstage.ai” facts: - label: “Vendor” value: “Upstage” - label: “Released” value: “2026” - label: “License” value: “Proprietary API” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-grok-4.6 - abs-model-kimi-k3 - abs-model-lfm2.5-vl-3b - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-muse-spark-1.2 - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.7-flash - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update tags: - solar-pro-4 - ai-models - model-release draft: false --- ## What it is Solar Pro 4 is Upstage’s new flagship proprietary agentic LLM (large language model) — a closed-weights API product built in Seoul, Korea, for multi-step office work that chains tool calls, reads long documents, runs terminal commands, and produces a finished deliverable. It is the sibling of Upstage’s separate open-weights model, Solar Open 2, but the two are not interchangeable: Solar Pro 4 is commercial and hosted, Solar Open 2 is free to download and self-host. The important framing is that Solar Pro 4 is not an upgrade in the “bigger context, smarter chat” sense. Upstage positions it as a category play for “production agents that finish the work,” and the launch copy is blunt about the design goal: > “Solar Pro 4 is built to carry real work to the finish, reading the documents, running the tools, producing the deliverable, and to stop and say so when the evidence runs out.” — Upstage, 2026 For the ABS operator, the simple read is a Korean-built, text-only model covering English, Korean, and Japanese, scoring a mid-tier 42 on the Artificial Analysis (AA) Intelligence Index, priced aggressively with a 90% launch discount, and openly accessible through OpenRouter today. ## Specs at a glance Upstage keeps the architecture and parameter count under wraps for Solar Pro 4 — that is the trade you make for a proprietary model. What is disclosed: | Attribute | Value | |---|---| | Parameters | Not disclosed (proprietary); Solar Pro 3 was MoE (mixture-of-experts), 102B total / 12B active | | Context window | 512K tokens (Upstage); OpenRouter reports 524K; AA lists 384K effective | | Max output | 128K tokens (Upstage) / 256K (AA) | | Modalities | Text input + text output only — no vision, no audio | | Languages | English, Korean, Japanese (input + output) | | Reasoning dial | Adjustable — “high” for deep analysis, “low” for real-time | | Knowledge cutoff | Not stated | Upstage trained the model on OfficeVerse, an internal pipeline that synthesizes office tasks across 11 domains and 12 task types. The context and language story is summed up in the launch post: > “Solar Pro 4 supports a 512K context with up to 128K output tokens … It handles English, Korean, and Japanese for both input and output, and response speed is a dial: set reasoning effort high for deep analysis, low for real-time interaction.” — Upstage, 2026 ## Benchmarks: how it actually scores The headline number is the AA Intelligence Index of 42, up from Solar Pro 3’s 14 — a 28-point jump that repositions the model from also-ran to mid-pack. Artificial Analysis puts it plainly: > “Korean AI lab Upstage has released Solar Pro 4, scoring 42 on the Artificial Analysis Intelligence Index, a significant increase from Solar Pro 3’s 14.” — Artificial Analysis, 2026 Where Solar Pro 4 earns its keep is agentic and long-document work. Terminal-Bench v2.1 climbs from 12 to 57, long-document reasoning (AA-LCR) from 31 to 71, and multi-turn tool use (τ³-Banking) from 9 to 23. GDPval-AA v2 Elo more than doubles to 1277 — above the human baseline of 1000 and ahead of Qwen3.7 Max (1272) and MiMo-V2.5-Pro (1266). Upstage’s own framing of the agentic gains: > “Terminal tasks (Terminal-Bench v2.1): 57 … Multi-turn tool use (τ³-Banking): 23 … Long-document reasoning (AA-LCR): 71” — Upstage, 2026 It also posts strong graduate-level marks — GPQA Diamond at 89.1%, SWE-Bench Verified at 70.6% (just above Solar Open 2’s 69.2%), and Ko-GDPval at 87.3%. But the ceiling shows: Humanity’s Last Exam (HLE) is 29.2% and research-level physics (CritPt) is a weak 5.4%. A telling detail on hallucination: AA-Omniscience Accuracy is essentially unchanged at 18.9% (vs 19%), while the hallucination rate drops from 88% to 24% — but largely because the model now attempts only 41% of questions (vs 92%) and abstains more. The honesty improvement comes from refusing, not from knowing more. ## Strengths 1. A real agentic leap over Solar Pro 3. Terminal-Bench 12→57 and GDPval Elo 498→1277 (above the human baseline) are several-fold gains, not marginal tweaks. This is a genuine category repositioning toward “production agents.” 2. Evidence-aware by design. The abstention behavior ships with grounded citations, “not-in-document” flags, and discrepancy reporting — exactly the discipline you want from any agent that touches real data. 3. 512K context with 128K output and explicit multi-document workflow support. The launch demo chains Excel→Word→PPT in one session. 4. Aggressive launch pricing. At $0.30 in / $1.20 out (list) with 90% off through Sept 10, 2026, it is effectively free to evaluate at scale. English/Korean/Japanese coverage is a bonus if you ever expand to Korea or Japan. ## Weaknesses 1. Proprietary, with no weights and no Hugging Face (HF) card. You cannot self-host, fine-tune, or audit it. If your stack has any “models must be open-weights or self-hostable” rule, Solar Pro 4 is excluded by definition — Solar Open 2 is the open alternative from the same lab. 2. Slower than its predecessor despite emitting fewer tokens. It takes 8.6 minutes per AA Index task vs Solar Pro 3’s 6.0. At a 42 Intelligence Index it is middle-of-the-pack, and cheaper-or-better options exist — DeepSeek V4 Flash 0731 reasons at 52 for $0.14/$0.28. 3. The accuracy gain is mostly refusal. Because the hallucination drop comes from attempting fewer questions, workflows that require an answer (not an abstention) lose coverage. 4. Limited frontier-reasoning headroom. CritPt at 5.4% and HLE at 29.2% mean it is not your top pick for deep research-grade reasoning. ## Where it fits for ABS operators The case for trying it is purely economic right now. The 90% launch discount runs through Sept 10, 2026 (23:59 UTC), dropping effective price to ~$0.03 in / $0.12 out per 1M tokens — competitive with DeepSeek and MiniMax-M3. The cost of evaluation is negligible, so it belongs in the model-eval queue for research, agentic, and long-document tasks. Concretely, ABS’s research-digest crons — operator-discovery-weekly, hermes-improvement-checker, and meta-ads-daily-stats-cron — all touch long documents and multi-step data, which is precisely Solar Pro 4’s target. It could replace GPT-4-class models on those flows at a fraction of the cost. The Korean/Japanese pass-through is free upside for any multilingual social or customer research; nothing else at this price covers both. Note the provenance: Nous Research / Hermes Agent is currently the #1 consumer of Solar Pro 4 on OpenRouter (43.4B tokens), so this evaluation comes from the same vantage point as one of the model’s heaviest real users. The clean decision rule: Solar Pro 4 is closed; Solar Open 2 is open. Reach for Solar Pro 4 when you want a cheap hosted agentic model you can call today; reach for Solar Open 2