Upstage's Solar Mini 4 scores 24 on AA Intelligence Index, but costs ~5x GPT-6 Luna per task
ArtificialAnlys · x · 2026-10-01
- Korea's Upstage released Solar Mini 4, a proprietary reasoning model with a reported 35B total / 3B active parameters, setting a new Pareto point on the Artificial Analysis Intelligence Index for models under 3B active parameters.
- It scores 24 — 16 points above its predecessor Solar Pro 3, 6 above Qwen3.6 35B A3B (same 3B active), and 1 above the 55B-active Nemotron 3 Ultra.
- Long-context reasoning is a strength: 83% on AA-LCR v1.1 (matching MiniMax-M3 and GPT-6 Luna max), 48% on SciCode. Agentic coding is weak: 1% on Terminal-Bench 4.0, 22% on AutomationBench-AA.
- Fast output (208 tokens/s vs GPT-6 Luna max's 152) but very verbose — 88k output tokens per task (7.1 minutes), and roughly 5x GPT-6 Luna's cost per task despite similar per-token pricing. Low knowledge accuracy (AA-Omniscience -11) but a 64% non-hallucination rate.
- Specs: 1M context, 262k max output, text-only, Feb 2026 knowledge cutoff, $0.10/$0.40 per 1M input/output tokens, closed weights.
Related event: Upstage Launches Solar Mini 4, Scoring 24 on Intelligence Index(3 posts)→
More from Models
- Users report Muse claims fixes without checking anything — ivan_bezdomny · 2026-10-01
- Dev claims 'Sonnet 5.5' hits 120-130tps with 100% pass rate on his end-to-end app benchmark — julianharris · 2026-10-01
- GPT-6.1 Sol sees highest demand ever as OpenAI doubles serving speed — pvncher · 2026-10-01
- GPT-6 Astra cracks 217-year-old cipher to Napoleon's general in 6 hours — arthurcolle · 2026-10-01
- OpenAI's 20x-to-10x quota cut is a push from Astra to Sol, not a better deal — CtrlAltDwayne · 2026-10-01
- Gemini 4 Argon tops APEX-Agents at 82.2% Pass@1, first model to break 80%, but burns 2.6M tokens per run — xennygrimmato_ · 2026-10-01