DeepSeek V4 Pro matches GPT-5.2 at 17x lower cost

An open-source challenger lands at the frontier

DeepSeek's V4 Pro, released on 24 April 2026, has matched OpenAI's GPT-5.2 to within 3% on FoodTruck Bench, an independent agentic business simulation benchmark, while costing roughly 17 times less per token. The Chinese lab's flagship now sits fourth on the FoodTruck leaderboard behind Claude Opus 4.6, GPT-5.2 and Grok 4.3, but the price gap is what has analysts talking.

The numbers behind the headline

GPT-5.2 is priced at $1.75 per million input tokens and $14 per million output tokens. DeepSeek V4 Pro charges $0.435 and $0.87 respectively, producing the roughly 17x cost advantage on output-heavy agentic workloads. Cached-token pricing falls as low as $0.145 per million tokens with six API providers already offering the model.

Architecture and efficiency

V4 Pro is a 1.6 trillion-parameter Mixture-of-Experts model that activates only 49 billion parameters per token, a 30:1 compression ratio between total and active parameters. A hybrid attention mechanism combining Compressed Sparse Attention and Heavily Compressed Attention requires only 27% of the inference FLOPs and 10% of the KV cache versus its predecessor V3.2 at a one-million-token context. The model is released open-source under an MIT licence on Hugging Face.

Important caveats

Independent evaluations have tempered the enthusiasm. The CAISI evaluation found V4 Pro's actual performance aligns more closely with the original GPT-5, released roughly eight months earlier, rather than GPT-5.2 as DeepSeek's own benchmarks suggest. The model also exhibits a 94% hallucination rate on the AA-Omniscience benchmark, meaning it almost always generates an answer rather than abstaining when it lacks knowledge. DeepSeek itself acknowledges V4 falls marginally short of top closed-source models by roughly three to six months of development.

Why it matters for enterprise

Despite the gaps, industry observers view the cost-performance ratio as transformative for enterprise AI. V4 Pro is now the largest open-weights model available, surpassing Kimi K2.6 at 1.1 trillion parameters, and positions DeepSeek to capture cost-sensitive agentic workflows where sustained, affordable inference matters more than absolute frontier performance.