DeepSeek V4 Pro Matches GPT-5 on MMLU-Pro, Costs 90% Less Per Token

📊 Benchmark Results Challenge Frontier Labs

DeepSeek released V4 Pro on May 18, 2026, and independent evaluations confirm it achieves near-identical reasoning scores to OpenAI's GPT-5 on MMLU-Pro (87.3 vs 87.1), MATH-500 (94.2 vs 94.8), and HumanEval (96.1 vs 96.6). The 685-billion parameter model uses a Mixture of Experts architecture with 256 experts and top-8 routing, activating roughly 37 billion parameters per forward pass.

The pricing differential is stark: DeepSeek charges $0.14 per million input tokens and $0.56 per million output tokens through its API, compared to $1.75 and $14.00 respectively for GPT-5. Several enterprise customers including Perplexity AI and Poe have already switched portions of their inference workloads to V4 Pro, citing indistinguishable quality at dramatically lower cost.

🧬 Architecture and Training Details

The model was trained over 58 days on a cluster of 2,048 NVIDIA H800 GPUs, using 14.8 trillion tokens of curated data. DeepSeek introduced Multi-Head Latent Attention (MHLA) in the V4 series, reducing KV-cache memory requirements by 56% compared to standard multi-head attention while maintaining accuracy. The context window extends to 256K tokens natively.

DeepSeek credits its training efficiency to a novel curriculum learning strategy that phases in coding, math, and multilingual data in later training stages, avoiding the quality dilution that plagues many open models trained on uniform corpus mixtures throughout. The V4 Pro also introduces native tool-use training, allowing it to be deployed for agentic tasks without fine-tuning.

🔓 Open-Source Release and Ecosystem Impact

The model weights are available on HuggingFace under Apache 2.0, with quantized versions from 4-bit to 8-bit already available via the llama.cpp and vLLM ecosystems. Community fine-tunes appeared within 72 hours, including chat-optimized variants and domain-specific adaptations for legal and medical applications.

Analysts at SemiAnalysis project that V4 Pro could capture 18-22% of the enterprise API inference market by Q4 2026 if DeepSeek maintains its pricing strategy. This release continues the trend of open-weight models closing the gap with proprietary frontier systems, following Meta's Llama 4 release in April.