DeepSeek V4 Pro reached general availability on August 13, 2026, and it is the most open frontier model available: 1.6T total parameters, 49B active per token, a 1,000,000 token context window, and weights published under the MIT license. This guide covers the official API pricing including peak and off-peak rates, the architecture changes that make the context window practical, and what it takes to run the weights yourself.
The model profile is on DeepSeek V4 Pro, and the wider picture is in Frontier AI Models 2026.
Official API Pricing
DeepSeek changed its pricing structure with the V4 lineup. From 16:00 UTC on August 16, 2026 there are peak and off-peak rates, and the off-peak rate is 50% lower than the peak rate, which makes scheduled batch work materially cheaper.
| Model route | Input | Output | Context |
|---|---|---|---|
deepseek/deepseek-v4-pro | $0.96 / 1M | $1.91 / 1M | 1,048,576 |
deepseek/deepseek-v4-pro-0813 | $1.32 / 1M | $3.96 / 1M | 1,048,576 |
Those figures are the OpenRouter list prices for the routes, and they are the most convenient way to compare the two snapshots. The 0813 snapshot is the general-availability release; the plain route is the cheaper line.
Three API features landed with the release and are worth knowing about before you write integration code:
- Reasoning effort levels. Low for simple tasks, high for daily agent workflows, and max for complex tasks, selectable per request rather than per model.
- Native OpenAI Responses API support. This is what makes the one-click Codex setup work; if you already speak the Responses API, you can point an existing client at DeepSeek.
- Peak and off-peak pricing. Off-peak rates are 50% lower, so a job that can wait for the window should wait for the window.
Architecture: Why a 1M Token Context Is Practical
The model card explains the efficiency work behind the long context, and it is the reason the context window is usable rather than theoretical.
- Hybrid attention. Compressed Sparse Attention and Heavily Compressed Attention are combined. In the 1M-token setting, DeepSeek-V4-Pro requires only 27% of the single-token inference FLOPs and 10% of the KV cache compared with DeepSeek-V3.2.
- Manifold-Constrained Hyper-Connections. A change to the residual structure that improves signal propagation stability without changing model expressivity.
- Muon optimizer. Used for faster convergence and greater training stability during pre-training.
- Pre-training scale. More than 32T diverse, high-quality tokens, followed by domain expert post-training and on-policy distillation to consolidate the specialisations.
The headline capability claim is the Pro-Max reasoning mode, which DeepSeek positions as the strongest open-source model for coding benchmarks and agentic tasks, narrowing the gap with leading closed models. Treat that as the vendor's claim; the model card is the primary source.
Running the Weights Locally
The weights are published on Hugging Face under the MIT license, with the transformers library named in the model metadata and a technical report available at arxiv.org/abs/2606.19348.
A realistic checklist before you plan capacity:
- Budget the memory first. At 1.6T parameters, a single GPU is not an option. Plan on multi-node inference or a large-memory host, and confirm the parameter precision you plan to serve.
- Use a mixture-of-experts aware engine. Community practice for MoE checkpoints is a dedicated serving engine such as vLLM or SGLang rather than a plain
transformersloop; the activation split of 49B active per token is what makes MoE serving worthwhile. - Check the long-context path. The 1M token window is the headline feature, so validate KV-cache behaviour and throughput at 100K and above rather than only at short context, and confirm that your hardware and engine combination keeps the compressed-attention gains.
- Decide between serving and proxying. For most teams the API is cheaper than the hardware. The hosted routes above already expose both snapshots, so self-hosting only pays off when you need the weights on your own infrastructure, or when the API path is not available in your region.
If you want a hosted route with a free allowance while you test, OpenRouter and Together AI both expose open models through an OpenAI-compatible API, and Hugging Face Inference Providers covers supported models from one account.
DeepSeek V4 Pro Against the Field
On price per token, DeepSeek V4 Pro sits below Grok 4.7 ($0.96 and $1.91 per million against $2.00 and $6.00) and far below closed alternatives. Where it differs is in access: MIT weights mean the model can be run, audited and modified on your own infrastructure, which no closed frontier model allows. The trade-off is operational, not technical: you own the serving stack.
Frequently Asked Questions
How much does DeepSeek V4 Pro cost?
$0.96 per million input tokens and $1.91 per million output tokens on the standard route, or $1.32 and $3.96 on the 0813 general-availability snapshot, with a 1,048,576 token context. DeepSeek's own platform also applies peak and off-peak rates, where off-peak is 50% lower.
Can I download the DeepSeek V4 Pro weights?
Yes. The weights are published on Hugging Face under the MIT license, with a model card and technical report. Serving them needs multi-node capacity because of the 1.6T parameter count.
What is the context length of DeepSeek V4 Pro?
One million tokens, implemented with hybrid compressed attention that uses 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2 in that setting.
Does DeepSeek V4 Pro work with Codex?
Yes. The release added native OpenAI Responses API support and was optimized for Codex with a one-click setup, so an existing Responses API client can point at DeepSeek directly.
Is DeepSeek V4 Pro better than Grok 4.7?
That depends on what you are optimising. DeepSeek V4 Pro is far cheaper per output token and the only option in this comparison whose weights you can run yourself. Grok 4.7 publishes a stronger cost-per-task benchmark table for long-horizon coding. The official benchmark comparison for Grok is covered in the Grok 4.7 pricing guide.
