DeepSeek V4 just dropped, and the price tag is insane
DeepSeek made everyone wait four months for V4, and now we know why.
The preview release lands as two models. V4-Pro is a 1.6 trillion parameter MoE with 49 billion active, trained on 33 trillion tokens. V4-Flash is the efficient sibling at 284 billion total and 13 billion active. Both come with a 1 million token context window. Both are MIT licensed. V4-Pro is the largest open-weight model anyone has ever shipped.
The numbers where it counts for agents: SWE-bench Verified 80.6. Terminal-Bench 2.0 67.9. LiveCodeBench 93.5. Codeforces rating 3206. MRCR 1M at 83.5. This is frontier territory, maybe three to six months behind GPT-5.5 and Opus 4.7 on the hardest evals, but indistinguishable on plenty of agent workloads.
Now the part that reshapes the market. V4-Flash is $0.14 per million input tokens and $0.28 per million output. V4-Pro is $1.74 in and $3.48 out. Compare that to Opus 4.7 at $5 and $25, GPT-5.5 at $5 and $30. Inference cost on a long-running agent is about to drop ten to twenty times for anyone willing to switch backends.
The technical spine is a hybrid attention architecture combining Compressed Sparse Attention and Heavily Compressed Attention. At 1M context V4-Pro uses 27 percent of the FLOPs and 10 percent of the KV cache of V3.2. FP4 for experts, FP8 for everything else. This is agent-era infrastructure, long context is finally cheap enough to actually use.
For anyone building agents with Chinese open weights as a fallback, the calculation just changed again. For anyone building with closed-source APIs, the pricing pressure is real.
Link: https://chat.deepseek.com and https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
← Back to all articles
The preview release lands as two models. V4-Pro is a 1.6 trillion parameter MoE with 49 billion active, trained on 33 trillion tokens. V4-Flash is the efficient sibling at 284 billion total and 13 billion active. Both come with a 1 million token context window. Both are MIT licensed. V4-Pro is the largest open-weight model anyone has ever shipped.
The numbers where it counts for agents: SWE-bench Verified 80.6. Terminal-Bench 2.0 67.9. LiveCodeBench 93.5. Codeforces rating 3206. MRCR 1M at 83.5. This is frontier territory, maybe three to six months behind GPT-5.5 and Opus 4.7 on the hardest evals, but indistinguishable on plenty of agent workloads.
Now the part that reshapes the market. V4-Flash is $0.14 per million input tokens and $0.28 per million output. V4-Pro is $1.74 in and $3.48 out. Compare that to Opus 4.7 at $5 and $25, GPT-5.5 at $5 and $30. Inference cost on a long-running agent is about to drop ten to twenty times for anyone willing to switch backends.
The technical spine is a hybrid attention architecture combining Compressed Sparse Attention and Heavily Compressed Attention. At 1M context V4-Pro uses 27 percent of the FLOPs and 10 percent of the KV cache of V3.2. FP4 for experts, FP8 for everything else. This is agent-era infrastructure, long context is finally cheap enough to actually use.
For anyone building agents with Chinese open weights as a fallback, the calculation just changed again. For anyone building with closed-source APIs, the pricing pressure is real.
Link: https://chat.deepseek.com and https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
Comments