# DeepSeek V4 Flash 0731 — Detailed Ranking Research Based on live data from **Artificial Analysis** (the leading independent AI benchmarking platform, as of August 2026), here is a detailed breakdown of where **DeepSeek V4 Flash 0731 (Reasoning, Max Effort)** ranks relative to other major models. ## 1. Overall Intelligence Ranking (Artificial Analysis Intelligence Index v4.1) The model scores **50 on the Intelligence Index** (out of a scale where the top model scores 61). This places it in the **upper tier of all models** — roughly **#21 overall** (with ties) out of 590+ models tracked. **Models scoring ABOVE or EQUAL to DeepSeek V4 Flash 0731 (50):** | Score | Model | |-------|-------| | 61 | Claude Opus 5 (max) | | 60 | Claude Opus 5 (xhigh) / Claude Fable 5 | | 59 | GPT-5.6 Sol (max) / Claude Opus 5 (high) | | 58 | GPT-5.6 Sol (xhigh) | | 57 | Kimi K3 (max) | | 56 | Claude Opus 5 (medium) / GPT-5.6 Sol (high) | | 55 | GPT-5.6 Terra (max) | | 54 | Grok 4.5 (high) / GPT-5.6 Sol (medium) | | 53 | Claude Sonnet 5 (max) | | 52 | GPT-5.6 Terra (xhigh) | | 51 | GPT-5.6 Luna (max) / GLM-5.2 (max) / Muse Spark 1.1 / Claude Opus 5 (low) | | **50** | **DeepSeek V4 Flash 0731 (max)** ⬅ tied with **Gemini 3.5 Flash** and **Gemini 3.6 Flash** | **Key takeaway:** It is tied with Google's Gemini 3.5/3.6 Flash models, and sits just one point below the 51-tier (GPT-5.6 Luna, GLM-5.2, Muse Spark, Claude Opus 5 low-effort). It is the **highest-scoring DeepSeek model**, comfortably ahead of **DeepSeek V4 Pro (44)** and **DeepSeek V4 Pro (max, 44)**. ## 2. Ranking Within Its Class (Open Weights, Large >150B) Among the **101 open-weights models of its size class**, it ranks: - **Intelligence: #3 / 101** — a leading open-weights model - **Price: #22 / 101** - **Cache Hit Price: #1 / 101** (best in class) - **Verbosity: #36 / 101** ## 3. Agentic & Coding Performance The Intelligence Index v4.1 is built from 9 evaluations, and DeepSeek V4 Flash is particularly strong on agentic/coding workloads: - **GDPval-AA v2** (real-world agentic tasks): **1559 Elo** - **Terminal-Bench v2.1** (agentic coding & terminal use): **82.7%** - Strong on multi-step reasoning and tool-use benchmarks (𝜏³-Banking, SciCode) This makes it one of the top open-weights models for **agentic and coding workflows** specifically. ## 4. Pricing & Efficiency (Its Standout Advantage) | Metric | Value | Class Rank | |--------|-------|-----------| | Input price | $0.14 / 1M tokens | Competitively priced (median $0.43) | | Output price | $0.28 / 1M tokens | Competitively priced (median $1.20) | | Cache hit price | $0.003 / 1M tokens (**-98%**) | **#1 / 101** | | Context window | 1M tokens | — | ## 5. Technical Specifications - **Total parameters:** 284B (sparse Mixture-of-Experts) - **Active parameters:** 13B per token - **Context window:** 1M tokens - **License:** MIT (open weights) - **Reasoning model** (max effort variant) ## Summary of Its Position **DeepSeek V4 Flash 0731 ranks in the top tier of all AI models** — roughly #21 overall on the Intelligence Index (score 50), tied with Gemini 3.5/3.6 Flash and just below the 51-tier frontier models. Among **open-weights models it is #3 in its size class** for intelligence, making it one of the best open-weights models available. Its defining characteristic is **cost-efficiency**: it delivers frontier-adjacent intelligence at a fraction of the price of proprietary rivals (with the #1 cache-hit price in its class), making it arguably the **best intelligence-per-dollar** model among major open-weights options — particularly for high-volume, agentic, and coding workloads. The main trade-off is that it trails the absolute frontier (Claude Opus 5, GPT-5.6 Sol, Kimi K3) by ~10 points on raw intelligence, and its output is verbose (210M tokens generated during evaluation, 4/4 verbosity units).