DeepSeek's V4-Flash-0731 Beats Bigger Models on Agent Tasks at $0.08 per Million Tokens
DeepSeek's V4 Flash 0731 jumps 10 points on the intelligence index with pure post-training, beating its own Pro model at a fraction of the cost
- DeepSeek V4 Flash 0731 is now in public beta via the DeepSeek API, scoring 50 on the Artificial Analysis Intelligence Index -- a 10-point jump.
- The upgrade is post-training only: same 284B total / 13B active MoE architecture, no new weights, just re-post-trained for agentic tasks.
- Agent benchmarks nearly doubled in some cases: DeepSWE jumps from 7.3 to 54.4, Cybergym from 38.7 to 76.7, Terminal-Bench 2.1 to 82.7.
- Priced at $0.14/$0.28 per 1M input/output tokens with a 98% cache hit discount ($0.0028/M), undercutting GPT-5.6 Luna even after OpenAI's 80% price cut.
- The model is text-only and currently API-only; open weights are expected to be released in the coming weeks.
- V4-Pro official release is teased as coming soon; the V4-Pro API and consumer app remain on existing builds for now.
DeepSeek has pushed the official API for DeepSeek-V4-Flash-0731 into public beta, and the headline number is hard to ignore: a 10-point jump on the Artificial Analysis Intelligence Index, from 40 to 50, achieved without changing a single weight in the base model. DeepSeek-V4-Flash-0731 keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained. That distinction matters enormously for what it implies about where capability gains are actually coming from.
The company says it has made a substantial upgrade to the model's agentic capabilities, with benchmark scores now well ahead of even its own V4-Pro-Preview. The Flash model, the smaller sibling in the V4 family, has effectively leapfrogged the Pro preview on agent tasks. This update releases a notable signal: the performance of the official Flash version in multiple agent benchmark tests has approached or even exceeded the level of the V4-Pro preview version.
What changed under the hood
According to DeepSeek's official technical report, V4-Flash has a total of 284 billion parameters and 13 billion activated parameters; V4-Pro has a total of 1.6 trillion parameters and 49 billion activated parameters. The two models differ by nearly an order of magnitude in scale, yet the smaller one now wins on agent tasks. If Flash can bring agent capabilities close to the level of Pro through post-training, it means that for specific tasks, model scale is not the decisive factor -- the weight of training methods and data quality is rising.
The V4 architecture itself is worth understanding. DeepSeek-V4-Flash is a fine-grained Mixture-of-Experts language model that uses a 43-layer all-MoE backbone with 256 routed experts plus one shared expert per block and top-6 routing. MoE (Mixture-of-Experts) means only a subset of the model's parameters activate per token -- here, just 13B out of 284B total -- keeping inference costs low. The V4 series uses a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to dramatically improve long-context efficiency; in the 1M-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2.