DeepSeek V4.1 Flash matches Gemini 3.1 Pro on ARC-AGI-2 at ~1/6 the cost
teortaxesTex · x · 2026-10-23
ARC Prize's verified analysis shows DeepSeek V4.1 Flash 1017 scaling behavior: ARC-AGI-1 rises 97.2%→98.4% (26k→34k reasoning tokens), ARC-AGI-2 rises 75.6%→79.2% (108k→134k tokens/task, $0.16/task at Max).
Key comparisons from the verified leaderboard:
- Gemini 3.1 Pro (High), Feb 19, 2026: 77.1% at $0.962/task — the earliest ordinary LLM within one point of DeepSeek, roughly 8 months earlier, at 6× cost
- GPT-5.4 Pro (XHigh), Mar 5: 83.3% at $16.41/task (excludable as a Pro/Deep Think-style compute-heavy config)
- GPT-5.5 (XHigh), Apr 23: 85.0% at $1.87/task — first verified conventional LLM above 78.1%, at 12× cost
- DeepSeek V4.1 Flash 1017 (Max), Oct 17: 78.1% at $0.16/task
The takeaway: DeepSeek matched February-level frontier accuracy eight months later at a fraction of the price, showing open-weight models keep closing the capability gap while undercutting closed models on inference cost.
More from Venture
- Arbitrage idea: take faceless YouTube gigs on Fiverr, outsource via Xianyu with an agent — yangyi · 2026-10-23
- Stripe's Link adds purchase protection as AI agent "fat-finger" insurance debate heats up — jeff_weinstein · 2026-10-23
- VC Yang Yi: don't build AI apps now, the frontier period belongs to infra plays — yangyi · 2026-10-23
- Indie RAG eval auditor lands zero paid gigs despite reproducible findings — Additional-Ladder-24 · 2026-10-23
- Everyone may be right about AI but wrong about the investment, argues AI infra skeptic — sebpaquet · 2026-10-23
- SF Compute's Evan Conrad: AI startups hit credit limits before power limits — AI Engineer · 2026-10-23