DeepSeek V4 Pro just got benchmarked by NIST’s CAISI against GPT-5. The verdict: roughly the same intelligence, about 8 months behind frontier closed models, but cheaper than GPT-5.4 mini on 5 of 7 tested benchmarks — anywhere from 53% less expensive to 41% more.
That’s the whole story. China’s open-weight champ is again forcing closed-source labs to defend their pricing.
A 1.6T-parameter open-weight model
V4 Pro is Mixture-of-Experts: 1.6T total parameters, 49B activated, 1M-token context. The hybrid attention architecture uses about 27% of the inference FLOPs and 10% of the KV cache that V3.2 needed at million-token scale. Pre-trained on 32T+ tokens with the Muon optimizer.
Real benchmark numbers: 93.5% on LiveCodeBench, 90.1% on GPQA Diamond, 80.6% on SWE-Bench Verified. Coding and reasoning are the strong suits. Where it still trails: ARC-AGI-2 (46% vs GPT-5.5’s 79%) and PortBench software engineering (44% vs 78%).
API access and what to build with it
Open weights are on Hugging Face. API runs on the DeepSeek platform at $1.74 per 1M input tokens and $3.48 per 1M output, with three configurable reasoning modes — dial cost vs depth per request.
The obvious use case: long-horizon agent workflows where context blows up and per-token cost decides whether the project ships. Repo-scale code review, multi-doc legal pipelines, support agents that read full chat history. Anywhere GPT-5 makes sense but the bill doesn’t.
Every DeepSeek release validates open weights for production and squeezes another margin point out of OpenAI and Anthropic.
You Might Also Like
- Gpt oss 120b Openai Finally Goes Open Source and its Worth the Wait
- Deepseek v4 Just Dropped and the Open Source ai World is Losing its Mind
- Glm 5 Just Dropped and its the Open Source Model Nobody saw Coming
- Moonshine Open Weights stt the Tiny Speech Model That Punches way Above its Weight
- Evo 2 Just Dropped and it Might be the Most Important Open Source Model Youve Never Heard of
Top AI Product 
Leave a comment