A June 7 head-to-head evaluation by RuntimeWire found DeepSeek V4 Pro outperforming OpenAI’s GPT-5.5 Pro on a precision benchmark, scoring 38.0 to 33.0 across four fresh text-generation tasks (RuntimeWire). The result echoes a pattern that has defined the Chinese AI lab’s trajectory since its founding: DeepSeek’s models rarely lead on every aggregate benchmark, but they consistently win on efficiency and task-specific discipline at a fraction of the cost.
The benchmark used four tasks generated on the fly so neither model could prepare. DeepSeek V4 Pro won on three: a log-redaction task where it used a single correct regex while GPT-5.5 Pro introduced ordering bugs; an instruction-following exercise where DeepSeek adhered precisely to the prompt while GPT-5.5 Pro added unnecessary details; and a schema-matching task where DeepSeek followed the required JSON structure exactly while GPT-5.5 Pro broke the format. Both models performed correctly on the fourth task, a messy-orders-to-JSON conversion.
What DeepSeek V4 Pro actually is
Released April 24, 2026, under an MIT license, DeepSeek V4 Pro is a 1.6-trillion-parameter Mixture-of-Experts model with 49 billion activated parameters per token and a 1-million-token context window. It uses a new hybrid attention mechanism (Compressed Sparse Attention combined with Heavily Compressed Attention) that reduces inference FLOPs by 27% and KV cache by 90% compared to the previous V3.2 architecture at full context length (Simon Willison; DeepInfra; OpenRouter).
It is available in three reasoning modes: non-thinking, think high, and think max. The think-max mode achieved a LiveCodeBench score of 93.5 — the highest ever recorded, ahead of Gemini 3.1 Pro at 91.7 and Claude Opus 4.6 at 88.8 — and a Codeforces ELO rating of 3206, also the highest on record (DeepInfra).
The cost difference is the headline
The per-task cost in RuntimeWire’s evaluation was approximately $0.10 for DeepSeek V4 Pro compared to approximately $22 for GPT-5.5 Pro, a ratio of roughly 1 to 200. On OpenRouter, DeepSeek V4 Pro costs $0.44 per million input tokens and $0.87 per million output, while GPT-5.5 Pro costs $5 and $30 respectively. Directly from DeepSeek’s API, the pricing is identical at $0.435 input and $0.87 output (DeepSeek API Docs; Simon Willison).
One developer reported consuming 16 million tokens on DeepSeek V4 Pro in a single day and spending $0.47. The same volume on GPT-5.5 Pro would have cost over $100 (TheTesseraPress, June 8).
Where it still trails
The NIST Center for AI Standards and Innovation (CAISI) published an official evaluation on May 1 putting DeepSeek V4 Pro’s performance in context. On broad capability benchmarks, GPT-5.5 Pro still leads: SWE-Bench Verified (81% for GPT-5.5 vs 74% for DeepSeek V4 Pro), GPQA Diamond (96% vs 90%), and PortBench internal security evaluation (78% vs 44%) (NIST).
CAISI estimates that DeepSeek V4 Pro performs similarly to OpenAI’s GPT-5 tier, lagging the frontier by roughly eight months. Its notable weakness is a high hallucination rate on knowledge-intensive tasks — it scored 94% hallucination on AA-Omniscience, as it rarely chooses to abstain even when uncertain (DeepInfra).
The open-source factor
DeepSeek V4 Pro is released under the MIT license, making it freely available for download, modification, and deployment. That open access has practical implications: organizations that cannot or will not route sensitive code through US-based API endpoints can run DeepSeek V4 Pro on their own hardware. The 49-billion-parameter active footprint means it can be deployed on fewer GPUs than equivalent dense models, though the full 1.6-trillion-parameter checkpoint requires significant infrastructure.
The V4 Pro sits alongside V4 Flash (284 billion total parameters, 13 billion active, $0.14 per million input tokens) as DeepSeek’s first two-tier lineup and the first architecture change since V3. Both are labeled as “preview” releases, with legacy API aliases scheduled to retire on July 24.
The precision benchmark win is a narrow one — a single test of instruction-following on four tasks. But combined with the cost ratio, the coding-benchmark leadership, and the open license, it tells a story that the broader industry is grappling with. The frontier of AI capability is no longer exclusively American, and the gap in deployment cost is not closing.