DeepSeek-V4-Pro review, pricing and verdict

Teams already running agentic coding pipelines or self-hosting open weights get the clearest win: only 49 billion of the model's parameters activate per token, keeping inference cheap while still leading open-weight models on competitive coding benchmarks. It replaces closed frontier APIs for MIT-licensed weight access or lower output cost, not for teams needing image or voice input.

DeepSeek-V4-Pro, released April 24, 2026, is DeepSeek's flagship open-weight Mixture-of-Experts reasoning model, built for agentic coding, long-document analysis and cost-sensitive production use. It ships under the MIT license with three selectable reasoning depths and leads publicly ranked models on competitive-programming benchmarks against closed rivals from OpenAI, Google and Anthropic.

Provider: DeepSeek · Family: DeepSeek-V4

More about DeepSeek on HokAI

Context window: 1,048,576 tokens · Max output: 384,000

Input modalities: text, tool-calls · Output: text, tool-calls

About DeepSeek-V4-Pro

DeepSeek-V4-Pro is the flagship model from DeepSeek, a Chinese AI research company founded in 2023 by High-Flyer hedge fund CEO Liang Wenfeng. Released April 24, 2026 as a preview under the MIT license, V4-Pro is a Mixture-of-Experts transformer with 1.6 trillion total parameters and 49 billion activated per token. It was pre-trained on 33 trillion tokens using a multilingual corpus with emphasis on code, math, scientific text, and agentic execution traces. The model was trained on Huawei Ascend 950PR chips rather than Nvidia hardware, making it the first frontier-scale model to complete training on Huawei silicon. V4-Pro is the larger of two V4-family releases: the companion DeepSeek-V4-Flash uses 284 billion total parameters with 13 billion active, optimized for faster throughput at lower cost.

In Think Max mode (maximum reasoning effort), DeepSeek-V4-Pro trails Claude Opus 4.6 by only a fraction of a point on SWE-bench Verified, and leads Gemini 3.1 Pro. On GPQA Diamond (graduate-level scientific reasoning) it scores 90.1%, behind Gemini 3.1 Pro at 94.3% and trailing Claude on Humanity's Last Exam (37.7% vs Claude's 40.0%). For mathematics, it achieves 92% on MATH-500, 95.2% on HMMT 2026 February, and a perfect 120/120 on Putnam-2025. On LiveCodeBench it scores 93.5, and holds a Codeforces rating of 3,206, ahead of GPT-5.4 at 3,168 and Gemini 3.1 Pro at 3,052. The pattern: V4-Pro leads on competitive coding and math olympiad evaluations; closed models from Google and Anthropic maintain edges on the hardest scientific reasoning tasks.

DeepSeek-V4-Pro provides a 1,048,576-token (1 million token) context window with a maximum output of 384,000 tokens per request. The Think Max reasoning mode requires at least 384K tokens of context budget to operate at full capacity. The hybrid attention architecture (Compressed Sparse Attention plus Heavily Compressed Attention) achieves 1M-token processing at just 27% of the single-token inference FLOPs required by DeepSeek-V3.2, and uses only 10% of V3.2's KV cache at that length. HCA applies 128x compression to the sequence before dense attention, eliminating the sparse selection step that becomes expensive at scale. No independent third-party needle-in-haystack evaluation at 1M token depth has been published as of May 2026; efficiency figures are from DeepSeek's own benchmarks in the April 27, 2026 model card. Among current frontier models, Gemini 3.1 Pro also supports 1M tokens; GPT-5.4 defaults to 128K with a 1M preview tier; Claude Opus 4.6 caps at 200K with 64K max output.

At launch in April 2026, DeepSeek-V4-Pro is text-only: there is no native image, audio, or video input in the preview release. DeepSeek has signaled that multimodal support is in development, potentially as a V4 Vision extension or a DeepSeek OCRv3 integration. Text tool use and function calling are available through both the OpenAI ChatCompletions interface and an Anthropic-compatible API, with structured output supported across every current hosting provider. Parallel tool calls are supported, enabling multi-step agentic loops where the model issues multiple function calls in a single response. The three reasoning modes (Non-think for speed, Think High for logical analysis, Think Max for maximum depth) are toggled per request via API parameters. Computer use and web browsing are not available in the April 2026 preview.

Per-token pricing undercuts closed frontier rivals by several multiples on output cost, with a limited-time promotional discount active in mid-2026 and a standing free-token grant for new API accounts; exact tiers and worked cost examples are covered in the pricing section below rather than repeated here.

DeepSeek-V4-Pro is available through multiple API providers: DeepSeek's direct API, AWS Bedrock, Google Vertex AI, Azure AI Foundry, Fireworks, Together.ai, DeepInfra, Lightning AI, Nebius, SiliconFlow, and Novita. On DeepSeek's own API, generation speed is approximately 29.8 tokens per second; Fireworks leads at 169.9 t/s and Lightning AI at 162.4 t/s, a 5x speed differential across providers. Time to first token on DeepSeek's API averages 2.1 seconds; for Think Max mode the time to first answer token can exceed 2 minutes on complex tasks due to extended chain-of-thought reasoning. Open weights are on Hugging Face at deepseek-ai/DeepSeek-V4-Pro under MIT license, with both base and instruct variants available. Self-hosting requires the official FP4+FP8 mixed precision weights and a cluster of 4x H200 141GB GPUs as the minimum practical configuration. Community GGUF quantizations (Q4_K_M, Q2_K) are available via Unsloth, but V4-Pro at Q2_K still exceeds 400GB, making self-hosted V4-Pro impractical below a serious GPU cluster.

DeepSeek's published safety documentation for V4 is limited compared to Western frontier labs. After a two-stage post-training pipeline (SFT plus GRPO domain cultivation, then unified consolidation via on-policy distillation), V4 undergoes safety-focused alignment tuning including constitutional-style guidelines and multi-language safety alignment. The April 27, 2026 model card states that sensitive personal data, credit card numbers, and identification information are excluded from training data. No external red-team partners are disclosed. The MIT license enables any party to download and modify the weights, including removing safety training, a categorically different risk profile from API-only closed models. The direct DeepSeek API does not hold SOC 2 Type II, ISO 27001, or HIPAA certification; teams with data compliance requirements should deploy through AWS Bedrock or Azure AI Foundry.

Teams running agentic coding workflows (V4-Pro leads SWE-bench at 80.6% Think Max and Codeforces at 3,206), long-document analysis needing 1M token context, and cost-sensitive production workloads priced well under closed frontier APIs have the clearest use case for V4-Pro. Open-source teams requiring fine-tuning, private inference, or air-gapped deployment benefit from the MIT license and downloadable weights. Teams building vision or multimodal applications should not use V4-Pro at launch (text-only) and should use GPT-5.4, Gemini 3.1 Pro, or Claude Opus 4.6 instead. Voice-first applications are ruled out by the absence of audio I/O. For the hardest scientific reasoning tasks, Gemini 3.1 Pro (GPQA Diamond 94.3%) outperforms V4-Pro (90.1%) by 4.2 points.

DeepSeek-V4-Pro was trained on 33 trillion tokens, more than double V3's 14.8 trillion, with emphasis on long documents and agentic execution traces. Post-training uses two stages: independent domain-expert cultivation through SFT and GRPO, followed by unified model consolidation via on-policy distillation. The Muon optimizer (Momentum plus Orthogonalization) replaces standard AdamW; MoE expert weights use FP4 precision and most other parameters use FP8, reducing memory footprint. Training cutoff date has not been publicly disclosed. All training was executed on Huawei Ascend 950PR chips with no Nvidia hardware in the compute stack.

DeepSeek-V4-Pro launched April 24, 2026 after three delays spanning four months since the December 2025 preview window target. Compared to V3.2, V4 raises SWE-bench Verified from 67.8% to 80.6% (Think Max, +12.8 points), LiveCodeBench from 74.1 to 93.5 (+19.4 points), and extends the context window from 128K to 1M tokens. The legacy API aliases deepseek-chat (routing to V4-Flash non-thinking) and deepseek-reasoner (routing to V4-Flash thinking) are deprecated with full access removal on July 24, 2026 at 15:59 UTC. The V4 series marks DeepSeek's first model line trained entirely on Huawei Ascend chips. No stable (non-preview) release date has been announced as of May 2026.

Pricing

Standard: $1.74/$3.48 per 1M cache-miss input/output. Promotional rate (75% off, until 2026-05-31): $0.435/$0.87 per 1M. Cache-hit input $0.145/M standard ($0.003625/M promo). New accounts receive 5M free tokens. No batch API; prompt caching is the primary cost optimization lever.

What a real job costs

Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.

Key Features

• Three Reasoning Modes: Non-think (fast), Think High (logical analysis), and Think Max (maximum depth) selectable per API request. Think Max leads open-weight models on SWE-bench Verified and requires a 384K token context budget.
• 1 Million Token Context Window: 1,048,576-token context with 384K max output. Hybrid CSA+HCA architecture processes 1M tokens at 27% of V3.2's inference FLOPs and 10% of its KV cache.
• MIT License Open Weights: Full model weights (865GB at FP4+FP8 mixed precision) on Hugging Face at deepseek-ai/DeepSeek-V4-Pro. Zero restrictions on commercial use, fine-tuning, or redistribution.
• Tool Use and Function Calling: Parallel tool calls supported via OpenAI ChatCompletions and Anthropic-compatible interfaces. JSON mode confirmed across all eight API hosting providers.
• Multi-Provider API Access: Available through 8+ providers including AWS Bedrock, Google Vertex AI, Azure, Fireworks (169.9 t/s), and Together.ai, with up to 1.9x price variation across providers.

Pros

• Think Max mode and a Codeforces rating of 3,206 are the highest published open-weight coding results as of May 2026.
• MIT license with full weight download enables fine-tuning, private inference, and air-gapped deployment with no vendor dependency.
• Output tokens cost several times less than Claude Opus 4.7 or GPT-5.5 at comparable coding benchmark performance.

Cons

• Text-only at launch: no native image, audio, or video understanding in the April 2026 preview, blocking all multimodal use cases.
• Self-hosting requires 865GB VRAM and a multi-GPU cluster; community Q2_K GGUF still exceeds 400GB, making on-device deployment impractical.
• DeepSeek's direct API lacks SOC 2, ISO 27001, and HIPAA certification, requiring enterprise users to route through Bedrock or Azure.

Benchmarks

• MATH: 92% vendor-reported · 27 Apr 2026 — Competition maths problems, % solved.
• MMLU: 89% vendor-reported · 27 Apr 2026 — General-knowledge exam across 57 subjects, % correct.
• MMLU-Pro: 87.5% vendor-reported · 27 Apr 2026 — A harder version of the 57-subject knowledge exam, % correct.
• HumanEval: 90% vendor-reported · 27 Apr 2026 — Small programs that must pass hidden tests, % passing.
• LiveBench: 93.5% vendor-reported · 27 Apr 2026 — A rolling set of fresh questions that cannot have been in training data, % correct.
• GPQA Diamond: 90.1% vendor-reported · 27 Apr 2026 — PhD-level science questions that are hard to search for, % correct.
• SWE-bench Verified: 80.6% vendor-reported · 27 Apr 2026 — Real GitHub issues fixed end to end, % solved.
• Humanity's Last Exam: 37.7% vendor-reported · 27 Apr 2026 — Expert-written questions across many fields, % correct.
• AA blended price: $0.71/M cited: Artificial Analysis · 01 May 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
• Output speed: 30 tok/s cited: Artificial Analysis · 01 May 2026 — Median tokens written per second as measured by Artificial Analysis.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit DeepSeek-V4-Pro Official Page