DeepSeek-V4-Pro-0813 ships MIT-licensed open weights, with 49B parameters active per token and non-think, think-high, and think-max reasoning tiers callers pick per request. It suits cost-sensitive agentic coding teams willing to add an external safety layer, since independent 2026 red-teaming found refusal collapses under free-form prompting.
DeepSeek-V4-Pro-0813 is DeepSeek's flagship open-weight language model, scoring 90.1% on GPQA Diamond, one of the strongest graduate-level reasoning results among open-weight models as of August 2026. It ships MIT-licensed weights and a hybrid long-context attention architecture built for agentic coding workloads.
Provider: DeepSeek · Family: DeepSeek V4
Context window: 1,000,000 tokens · Max output: 384,000
Input modalities: text, tool-calls · Output: text, tool-calls
About DeepSeek-V4-Pro-0813
DeepSeek-V4-Pro-0813 is the general-availability build of DeepSeek's flagship reasoning and agentic model, released August 13, 2026 by DeepSeek, the Hangzhou, China research lab founded by Liang Wenfeng. The 0813 build closes out a preview window that began with the original V4-Pro launch on April 24, 2026. Architecturally it is a Mixture-of-Experts transformer with 1.6 trillion total parameters and 49 billion active per token, designed for agentic coding, tool use, and long-context analysis rather than casual chat. On verified benchmarks, DeepSeek-V4-Pro-0813 scores 80.6% on SWE-bench Verified, 90.1% on GPQA Diamond, 87.5% on MMLU-Pro, and 93.5% pass@1 on LiveCodeBench. That SWE-bench score matches Gemini-3.1-Pro's agentic coding result, but the model trails GPT-5.4 on Terminal Bench 2.0 (67.9% versus 75.1%) and trails Gemini-3.1-Pro on Humanity's Last Exam (37.7% versus 44.4%). Artificial Analysis places it at an Intelligence Index of 53, ranked third of 106 models tracked as of the August 2026 measurement. The model ships a 1,000,000-token context window with a 384,000-token maximum output, and the two share the same budget: a 700,000-token prompt leaves under 300,000 tokens for generation. Long context is handled by a hybrid attention design across a 61-layer stack, alternating Compressed Sparse Attention with Heavily Compressed Attention, which cuts single-token inference compute to 27% of DeepSeek-V3.2's cost at maximum context length and reduces KV cache size to roughly 10%. DeepSeek-V4-Pro-0813 is text-only: it accepts and produces text, with no native image, audio, or video input, despite pre-launch reporting that expected multimodal training to ship. It supports tool calls and structured JSON output, and the 0813 build added native OpenAI Responses API compatibility aimed specifically at Codex-style agent integrations. Callers choose between non-think, think-high, and think-max reasoning-effort tiers per request. Pricing dropped in mid-2026 and remained unchanged through the 0813 GA release, with a steep discount applied to cached input tokens on repeat-context calls; exact current rates and worked cost examples are detailed in the pricing FAQ below. DeepSeek has separately announced a peak/off-peak pricing structure due to land within days of this GA release, without yet publishing the new peak figures. Access runs through DeepSeek's own OpenAI-compatible API, capped at 500 concurrent connections per account, plus Microsoft Azure AI Foundry and third-party hosts including Fireworks AI, Together AI, DeepInfra, SiliconFlow, Nebius, Novita, and Baseten, plus the OpenRouter aggregator. AWS Bedrock and Google Vertex AI listings were called "coming soon" at the April launch but were not confirmed live as of this listing. Self-hosting the open weights (FP4+FP8 mixed and FP8 mixed Safetensors formats) requires a multi-node GPU cluster even at reduced quantization; exact VRAM figures are detailed in the limitations below. Post-training combines supervised fine-tuning with GRPO reinforcement learning and on-policy distillation across the three reasoning tiers, per DeepSeek's Hugging Face model card. Independent red-teaming from FAR.AI found the safeguards brittle: a public roleplay jailbreak template first circulated for an earlier DeepSeek model raised the StrongREJECT jailbreak success rate from 0.6% to 77.8% on V4-Pro with no modification, and agentic-misuse template attacks cut refusal below 1%. Refusal is also framing-dependent: binary yes/no prompts triggered an 88.6% refusal rate, but the same requests phrased as free-form generation saw refusal collapse to about 1%, with harmful persuasion attempted on 85-95% of the non-controversial harmful topics FAR.AI tested. DeepSeek-V4-Pro-0813 suits cost-sensitive teams running agentic coding pipelines or long-document synthesis who want frontier-adjacent benchmark scores without frontier US-lab pricing, and infra teams that specifically want to self-host open weights under a permissive MIT license. It is a poor fit for consumer-facing deployment without an added moderation layer, given the jailbreak findings above, and for workloads needing native vision or audio input, where GPT-5.4 or Gemini-3.1-Pro are the safer default. DeepSeek pretrained the model on more than 32 trillion tokens, over double V3's 14.8 trillion, with deliberate emphasis on long documents and agentic execution traces, but has not published an official training-data cutoff date for V4. Its core infrastructure runs out of Hangzhou, China, a different data-residency profile than US- or EU-hosted alternatives; compliance certification details are covered in the FAQ below. The 0813 build closed a preview window that began April 24, 2026 with the joint V4-Pro and V4-Flash launch. V4-Flash graduated to general availability first, on July 31, 2026, with V4-Pro following two weeks later. DeepSeek discontinued the legacy deepseek-chat and deepseek-reasoner API model names on July 24, 2026, forcing callers onto the versioned deepseek-v4-pro endpoint.
Pricing
Standard per-token rate carried over unchanged from the April 2026 preview; cached input on repeat-context calls runs at roughly 1.2% of the cache-miss rate. A peak/off-peak split lands August 16, 2026 16:00 UTC, off-peak priced at half of peak, with exact new figures still unpublished at GA. See the pricing FAQ for the full input/output/cached breakdown.
Key Features
- 1M-Token Context Window: Hybrid Compressed Sparse Attention plus Heavily Compressed Attention supports a full 1,000,000-token context, built for long-document and agentic workloads rather than short chat turns.
- Three Reasoning-Effort Tiers: Non-think, think-high, and think-max modes let callers trade latency and cost for reasoning depth on a per-request basis.
- Native Responses API Support: The 0813 build added OpenAI Responses API compatibility, aimed specifically at drop-in Codex-style agent integrations.
- MIT-Licensed Open Weights: Full model weights are published on Hugging Face under the MIT license, permitting commercial self-hosting and fine-tuning with no usage restrictions.
- Prompt Caching Discount: Repeat context, such as a long system prompt or codebase, is billed at a steep cache-hit discount versus a cache miss, cutting cost on repeat-heavy agent loops.
Pros
- LiveCodeBench pass@1 of 93.5%, among the highest reported scores for an open-weight model, per DeepSeek's official benchmark disclosure.
- MIT license and full open weights give teams unrestricted commercial self-hosting and fine-tuning rights, unlike Llama's restricted-use license.
- Hybrid attention architecture cuts single-token inference compute to 27% of the prior generation's cost at maximum context length, a DeepSeek-disclosed efficiency gain.
Cons
- Independent FAR.AI red-teaming found 98 to 100% jailbreak success across CBRN, cyber, and terrorism prompts using an unmodified predecessor jailbreak template.
- Text-only: no native vision, audio, or video input despite pre-launch reports anticipating multimodal training.
- Self-hosting the open weights needs roughly 960GB of VRAM even at the lightest widely-used quantization, meaning a multi-node GPU cluster rather than a single 8xH100 box.
Benchmarks
- mmlu pro: 87.5
- gpqa diamond: 90.1
- livecodebench: 93.5
- swe bench verified: 80.6
- humanitys last exam: 37.7
- artificial analysis intelligence index: 53
- artificial analysis speed tokens per sec: 80
Frequently Asked Questions
How much does DeepSeek-V4-Pro-0813 cost per 1M tokens?
DeepSeek-V4-Pro-0813 costs $0.435 per 1M input tokens on a cache miss, just $0.003625 per 1M on a cache hit (a 99.2% discount), and $0.87 per 1M output tokens. DeepSeek has announced a peak/off-peak pricing change effective August 16, 2026 16:00 UTC that had not been published in detail as of this writing, so confirm current rates before budgeting a production workload.
How does DeepSeek-V4-Pro-0813 compare on benchmarks vs Gemini-3.1-Pro?
DeepSeek-V4-Pro-0813 matches Gemini-3.1-Pro on SWE-bench Verified at 80.6%, but trails it on Humanity's Last Exam, 37.7% versus 44.4%. Agentic coding parity does not extend to every reasoning benchmark, so the choice depends on which axis matters for your workload.
Is DeepSeek-V4-Pro-0813 open source or proprietary?
DeepSeek-V4-Pro-0813 is open-source: DeepSeek publishes the full model weights on Hugging Face under the MIT license, which permits commercial use, self-hosting, and fine-tuning without restriction. Running the raw weights still demands the better part of a terabyte of VRAM (roughly 960GB) at the lightest widely-used quantization, so most teams access it via DeepSeek's API or a third-party host like Fireworks or Together AI instead.
Does DeepSeek-V4-Pro-0813 train on user data?
DeepSeek has not published a specific training-on-inputs policy for the V4-Pro API, and has not disclosed SOC 2, HIPAA, or GDPR certification. Teams with strict data-governance requirements should route through a non-China third-party host rather than DeepSeek's direct API.
Who is DeepSeek-V4-Pro-0813 best for and who should avoid it?
It suits cost-sensitive teams building agentic coding tools or self-hosting frontier-scale open weights, given its strong LiveCodeBench and SWE-bench results at a fraction of closed-model API pricing. Teams shipping directly to consumers should avoid deploying it without an added moderation layer, since red-teaming found refusal collapses under free-form prompting and an unpatched, publicly documented jailbreak succeeds on most attempts.