DeepSeek V4 Flash 0731 ARC-AGI Results: 89% on ARC-AGI-1, 61.4% on ARC-AGI-2
On August 7, ARC Prize published verified ARC-AGI results for DeepSeek V4 Flash 0731. The post hit 763 points on Hacker News the same day, and the discussion mostly circled around one thing: the price. The scores were reached at two to four cents per task.
The scorecard
On ARC Prize's Semi-Private test sets with max effort reasoning:
| Benchmark | Score | Cost per task |
|---|---|---|
| ARC-AGI-1 | 89.0% | $0.02 |
| ARC-AGI-2 | 61.4% | $0.04 |
The model ships three reasoning levels: Max, High and Low. Max scores 89.0% on ARC-AGI-1 and 61.4% on ARC-AGI-2; High drops to 87.0% and 56.0%; Low sits at 84.0% and 46.0%. Halving the reasoning budget only costs a few points.
ARC-AGI-2 is considerably harder than the first version, and most models post ugly numbers there. 61.4% puts V4 Flash in the same neighborhood as GPT-5.6 Luna. One HN commenter put it plainly: "results comparable to gpt 5.6 luna but cheaper," and another noted the cost per task is roughly 2x cheaper than Luna. A popular comment did the math on Kimi K3: it was the talk of the town a month ago, and now you get comparable performance for 1/20th of the price.
What changed in 0731
DeepSeek moved the 0731 version of V4 Flash into public beta on July 31. The API call stays the same — model name is still deepseek-v4-flash. According to the official changelog, 0731 keeps the same architecture and parameter count as Preview and was only re-post-trained, with agent capabilities as the focus.
Official agent benchmark numbers:
| Benchmark | Score |
|---|---|
| Terminal Bench 2.1 | 82.7 |
| NL2Repo | 54.2 |
| Cybergym | 76.7 |
| DeepSWE | 54.4 |
| Toolathlon (verified) | 70.3 |
| Agent Last Exam | 25.2 |
| Automation Bench (Public) | 25.1 |
| DSBench-FullStack | 68.7 |
| DSBench-Hard | 59.6 |
DeepSeek says these results "far exceed V4-Pro-Preview". The model natively supports the Responses API, is adapted for Codex, and open weights were published on Hugging Face alongside the release.
Pricing and usage
API pricing: $0.14 per million input tokens, $0.28 per million output tokens, 1049K context. With the open weights, the 304B checkpoint fits on a single AMD MI300X (192GB HBM) without quantization or offloading — someone already measured ~168 tok/s single-stream decoding on it.
On August 6, DeepSeek posted a notice on its platform: overall API pricing will go up in the near future, with a significant increase expected, subject to official announcement. Most HN commenters guessed the hike would land on cache-hit pricing; others read it as a sign that the full V4 Pro release is coming. Planning usage before the change, or self-hosting the open weights, are the two common responses.
Takeaway
DeepSeek keeps tying capability and cost together. 61.4% on ARC-AGI-2 at $0.04 per task makes it hard to argue that cheap models can only handle simple tasks. Two things worth watching next: when V4 Pro ships, and how long this price advantage survives the increase.