Model Details
DeepSeek V4 Pro
Quick answer: DeepSeek V4 Pro is DeepSeek's July 2026 proprietary API model scoring 83.4% BrowseComp, 96.7% AIME 2026, 89.3% Global MMLU-Lite, 76.5% IFBench, and 57.0% SimpleQA. Input priced at $0.435/MTok.
At a Glance
Where DeepSeek V4 Pro leads
- 83.4% BrowseComp — near top-of-market for web research
- 96.7% AIME 2026 — near-perfect competition mathematics
- 89.3% Global MMLU-Lite — excellent multilingual academic knowledge
- 57.0% SimpleQA — strong factual accuracy
- $0.435/MTok input — very affordable for this capability tier
Where it lags
- No open weights (unlike V3, R1 which were MIT licensed)
- 30.7% LHTB — moderate long-horizon agentic tasks
- 27.6% Agents Last Exam — limited on agent benchmark
Best for: Cost-effective frontier reasoning via API; web research pipelines (high BrowseComp); multilingual academic/knowledge tasks.
What DeepSeek V4 Pro Is
DeepSeek V4 Pro is DeepSeek's July 2026 proprietary API model — a departure from the open-weight approach of V3 and R1. The "Pro" variant in V4 sits below V4 Pro Max (the flagship) and above V4 Flash variants in the V4 family hierarchy.
The 83.4% BrowseComp is competitive with Kimi K2.6 (83.2%) and near GPT-5.5 (84.4%). The $0.435/MTok input pricing is a distinctive advantage — this capability tier at sub-$1/MTok is unusual among frontier models.
Specifications
| Field | Value |
|---|---|
| Organization | DeepSeek |
| License | Proprietary (API only) |
| Release date | July 2026 |
| Modality | Text only |
| Input price | $0.435/MTok |
Pricing
| Tier | Price per MTok |
|---|---|
| Input | $0.435 |
Available via DeepSeek API (api.deepseek.com).
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| BrowseComp | 83.4% | Benchgen evaluation | 2026-07 |
| AIME 2026 | 96.7% | Benchgen evaluation | 2026-07 |
| Global MMLU-Lite | 89.3% | Benchgen evaluation | 2026-07 |
| IFBench | 76.5% | Benchgen evaluation | 2026-07 |
| SimpleQA | 57.0% | Benchgen evaluation | 2026-07 |
| LHTB | 30.7% | Benchgen evaluation | 2026-07 |
DeepSeek V4 Pro vs Alternatives
| Model | BrowseComp | AIME 2026 | Price (input) | License |
|---|---|---|---|---|
| DeepSeek V4 Pro | 83.4% | 96.7% | $0.435/MTok | Proprietary |
| DeepSeek V4 Pro Max | — | — | Higher | Proprietary |
| Kimi K2.6 | 83.2% | 96.4% | Free (open) | Apache 2.0 |
| GPT-5.5 | 84.4% | — | Much higher | Proprietary |
DeepSeek V4 Pro vs Kimi K2.6: near-equal BrowseComp (83.4% vs 83.2%) and AIME (96.7% vs 96.4%). K2.6 is Apache 2.0 open-weight; V4 Pro is API-only but well-priced. For API use: V4 Pro at $0.435/MTok. For self-hosting: K2.6.
Related Models & Benchmarks
- Compare: DeepSeek V4 Pro Max · Kimi K2.6 · GPT-5.5
- Benchmarks: BrowseComp · AIME 2026 · IFBench
Frequently Asked Questions
What is DeepSeek V4 Pro?
DeepSeek's July 2026 proprietary API model scoring 83.4% BrowseComp, 96.7% AIME 2026, 89.3% Global MMLU-Lite. $0.435/MTok input.Is DeepSeek V4 Pro open-source?
No — unlike DeepSeek V3 (MIT) and R1 (MIT), V4 Pro is proprietary API-only. DeepSeek changed to a closed model approach for the V4 generation.Specs from DeepSeek's V4 Pro release (July 2026) and Benchgen evaluations. Last updated 2026-07-24.