DeepSeek has pushed DeepSeek-V4-Pro to general availability, and the release notes read like a checklist of what developers have been asking for. The model combines a one million token context window with three selectable reasoning effort levels, adds native support for the OpenAI Responses API and ships with open weights under the MIT license. You can access it right now in the DeepSeek app and on the web by switching on Expert Mode, or call it through the API under the same model names you already use. This article walks through what the GA release changes, how the model scores on independent benchmarks and what the new peak and off-peak pricing means for your budget.
What the V4-Pro GA release adds
The headline upgrade is aimed at agents. DeepSeek describes major agent upgrades with strong production gains, which fits the design goal of the V4 series: long horizon workflows where the model plans, calls tools and writes code across many steps. Three changes matter most in practice.
First, flexible reasoning effort comes to both V4-Pro and V4-Flash. You pick low for simple tasks, high for daily agent workflows and max for complex problems. Second, the API now natively supports the OpenAI Responses API and is optimized for Codex with a one click setup, so teams with a toolchain built on OpenAI can swap endpoints with minimal friction. Third, the model names in the API stay unchanged, which means existing integrations keep working while the underlying model improves.
Architecture built for a million tokens
V4-Pro is a mixture of experts model with 1.6 trillion total parameters and 49 billion active per token. Its smaller sibling V4-Flash carries 284 billion parameters with 13 billion active. Both support a context length of one million tokens, roughly 1500 A4 pages of text.
The efficiency claims rest on a hybrid attention architecture that combines Compressed Sparse Attention and Heavily Compressed Attention. At a one million token context, V4-Pro needs only 27% of the single token inference FLOPs and 10% of the KV cache compared with DeepSeek-V3.2. That is the difference between a context window that looks good on a spec sheet and one you can afford to use in production.
Two more design choices stand out. Manifold-Constrained Hyper-Connections reinforce the residual connections between layers, which stabilizes signal propagation in a very deep network while preserving expressivity. Training ran with the Muon optimizer, chosen for faster convergence and greater stability across a pre-training run of more than 32 trillion tokens.
Post-training follows a two-stage recipe. DeepSeek first cultivates domain specific experts independently through supervised fine-tuning and reinforcement learning with GRPO. It then consolidates those specialists into one unified model through on-policy distillation, so a single deployment carries the coding, math and agentic skills that were trained separately.
How the three reasoning effort levels work
The reasoning dial is the most practical feature of the release. Low gives fast, intuitive responses for routine daily tasks and low risk decisions. High switches to conscious logical analysis that is slower but more accurate, and DeepSeek positions it as the default for daily agent workflows. Max pushes reasoning to its fullest extent for complex problem solving and planning, using a special system prompt and an extended thinking summary.
A few operational details are worth knowing. Some API providers expose the top tier as xhigh, which maps to the max reasoning mode. For local deployment in Think Max mode, DeepSeek recommends a context window of at least 384K tokens, because the model generates long chains of thought before answering. Independent testing by Artificial Analysis confirms the verbosity: the max effort variant produced 130 million output tokens during its evaluation suite, about 30% more than the median model.
Benchmark performance
Artificial Analysis gives V4-Pro running at max effort a score of 53 on its Intelligence Index, well above the median of 27 for comparable models and among the leading scores overall. The evaluation alone cost $604.51 to run, which hints at the sheer length of the reasoning traces involved.
On the public benchmarks DeepSeek reports, the pattern holds. V4-Pro resolves 80.6% of SWE-bench Verified issues, scores 90.1 on GPQA Diamond, 92.6 on GSM8K and 55.4 on the harder SWE-bench Pro set. DeepSeek positions V4-Pro-Max as the strongest open weights model available today, with coding performance at the top tier and a much smaller gap to leading closed models on reasoning and agentic tasks.
Speed is competitive. The model generates 81 tokens per second through DeepSeek’s own API against a median of 66 for its class, and time to first token sits at 1.85 seconds. For a reasoning model of this size, that throughput makes the max effort mode genuinely usable rather than purely aspirational.
Pricing and the new off-peak rates
V4-Pro sits at the higher end of its class on price. Through the DeepSeek API you pay $1.32 per million input tokens and $3.96 per million output tokens, above the medians for comparable open weights models. A blended rate that accounts for cache hits lands at $0.69 per million tokens.
Alongside the V4 lineup, DeepSeek is restructuring how you pay. The API moves to peak and off-peak rates, with off-peak prices set 50% lower than peak. The new pricing takes effect at 16:00 UTC on August 16, 2026. If your workloads include batch evaluation, nightly codebase analysis or large-scale information synthesis, scheduling those jobs in the off-peak window cuts the token bill in half without changing anything about your prompts.
Combined with the reasoning effort dial, you now control cost on two axes: how hard the model thinks and when it thinks. Few models at the frontier give you both levers at once.
Ways to run V4-Pro today
You have four realistic routes. The fastest is Expert Mode in the DeepSeek app and web interface, which exposes V4-Pro without any setup. For programmatic access, the DeepSeek API keeps the existing model names and adds the Responses API for Codex workflows. Aggregators such as OpenRouter also list the model, useful if you already consolidate billing there.
The fourth route is self-hosting. The weights are on Hugging Face under the MIT license, which permits commercial use. One caveat: the release ships without a Jinja chat template. DeepSeek provides a dedicated encoding folder with Python scripts that show how to encode OpenAI-format messages into model input strings and parse the output back. For local inference, the team recommends a temperature of 1.0 and top_p of 1.0.
Where V4-Pro leaves you
V4-Pro closes much of the remaining gap between open weights and the frontier, but the more durable shift is economic. A model that scores 53 on the Intelligence Index while letting you schedule heavy jobs at half price and dial reasoning down for easy tasks turns the AI budget from a fixed cost into a tunable one. The teams that benefit most will be the ones that rebuild their workloads around those two dials first.
