DeepSeek V4 Flash Tops ARC-AGI Cost-Performance, Costing a Quarter of GPT-5.6
DeepSeek V4 Flash impresses on ARC-AGI, redefining the performance-cost Pareto frontier with top scores at minimal inference cost. ARC Prize official validation shows leading semi-private task scores, with cost only a quarter of GPT-5.6, breaking the convention that higher performance implies higher cost, prompting industry rethink on agent deployment costs and inference mechanisms.
Confirmed
- ARC Prize official validation: DeepSeek V4 Flash scores 61.4% on ARC-AGI-2 semi-private tasks at $0.04/task, and 89.0% on ARC-AGI-1 at $0.02/task (cited by @petrusenkomax and @GregKamradt).
- Cost advantage: @GregKamradt notes cost is a quarter of GPT-5.6; @togethercompute finds in DeepSWE benchmark that at equal budget, DeepSeek V4 Flash costs about a third of GPT-5.6 and solves more tasks.
- Counterintuitive cost curve: Developer @DeArgonaut's tests on DeepSeek V4 (0731) confirm that higher scores come with significantly lower per-task inference cost. @GregKamradt cites official technical analysis explaining that models in highest inference mode are actually cheaper than high mode because more reasoning tokens lead to more derailment.
- Solving agent cost barrier: @Teknium and @MaziyarPanahi believe this disruptive cost-performance will break the funding bottleneck for running AI agents. @MaziyarPanahi adds that in real clinical handoff workflow tests, the model runs fast and is cost-controllable at scale.
Unconfirmed
- Specific test environment, hardware configuration, and detailed calculation of inference costs have not been fully disclosed.
Why it matters
- DeepSeek V4 Flash achieves top reasoning performance at extremely low cost, potentially driving down AI inference costs significantly, impacting developer model selection and industry pricing strategies. Its counterintuitive 'higher performance, lower cost' phenomenon offers a new perspective on LLM inference cost curves.
2026-08-08 ~ 2026-08-09 · 9 related posts
- Episode 1: DeepSeek-V4-Flash Architecture Leaked with Million-Token Context(2026-07-31, 3 posts)
- Episode 2: Unsloth Releases Quantized DeepSeek V4 Flash 0731 for Local Deployment(2026-07-31, 6 posts)
- Episode 3: DeepSeek on Huawei Ascend Beats OpenAI in Inference Profitability(2026-08-01, 2 posts)
- Episode 4: DeepSeek-V4-Flash Excels in Frontend Coding with Unmatched Cost-Performance(2026-08-01, 3 posts)
- Episode 5: DeepSeek V4 Preview: Flash to Introduce Four-Level Reasoning Effort(2026-08-01, 2 posts)
- Episode 6: DeepSeek Drastically Reduces Training Compute Costs Across Models(2026-08-01, 2 posts)
- Episode 7: DeepSeek V4-Flash API Public Beta Launches with Major Agent Upgrades(2026-08-01, 5 posts)
- Episode 8: DeepSeek V4-Flash Costs 105x Less, But Stability and Benchmark Overfitting Questioned(2026-08-01, 17 posts)
- Episode 9: DeepSeek V4 Flash Sparks a Wave of Local Deployment Tests(2026-08-01, 26 posts)
- Episode 10: DeepSeek V4 Flash Quantization Tests Show Major Speed Gains Without Quality Loss(2026-08-04, 2 posts)
- Episode 11: Single RTX 5090 Runs DeepSeek 1M Context(2026-08-04, 2 posts)
- Episode 12: DeepSeek-V4-Flash Tops Cost-Efficiency with Ultra-Low Running Costs(2026-08-05, 4 posts)
- Episode 13: DeepSeek V4 Flash Benchmarks Leak, Outperforming Pro(2026-08-05, 3 posts)
- Episode 14: DeepSeek V4 Flash Tops ARC-AGI Cost-Performance, Costing a Quarter of GPT-5.6(2026-08-08, 9 posts)
- Episode 15: DeepSeek V4 Flash Passes 22 Coding Tests on Dual DGX Spark Cluster(2026-08-10, 2 posts)
Primary sources
- [source] DeepSeek V4 Flash Dominates ARC-AGI Cost-Performance Frontier at 1/4 the Cost — GregKamradt · 2026-08-08
- ARC-AGI Insight: More Reasoning Tokens Lead to Wrong Hypotheses; GPT-5.6 Max is Cheaper — GregKamradt · 2026-08-08
- Teknium: DeepSeek Flash Pricing Will Eliminate AI Agent Spending Woes — Teknium · 2026-08-08
- Testing DeepSeek v4 flash: Price-performance ratio solves AI agent spending — MaziyarPanahi · 2026-08-08
- DeepSeek V4 Flash Scores Impressively on ARC-AGI Semi-Private Tasks — petrusenko_max · 2026-08-09
- [source] DeepSeek V4 tested: ARC-AGI reasoning costs drop despite higher performance — DeArgonaut · 2026-08-09
- [source] DeepSeek V4 Flash Solves More Tasks Than GPT-5.6 Luna at One-Third the Cost — togethercompute · 2026-08-09
2 near-duplicate retellings: yacineMTB · teortaxesTex