DeepSeek has officially launched DeepSeek V4 Promoving its flagship LLM out of preview and into general release across its app, web interface, and API. The new build, designated DeepSeek V4 Pro 0813, supersedes the preview version that had been available since April and arrives as the company continues to expand aggressively in a crowded and fast-moving AI landscape.
The release is centered on agentic capabilities, the growing category of AI tasks in which a model uses external tools, writes and executes code, and carries out multi-step workflows with little or no human supervision.
DeepSeek has positioned V4 Pro 0813 as a significant step up from its preview predecessor on precisely these fronts, releasing a wide set of benchmark results to back the claim.
DeepSeek V4 Pro Improves Sharper AI Agent and Coding Performance
According to figures published by DeepSeek, V4-Pro-0813 outperforms the April preview across nearly every benchmark the company tracks. The model scored 87.9 on Terminal Bench 2.1, up sharply from 72.1 for the preview version. It also posted scores of 61.5 on NL2Repo, 83.3 on Cybergym, and 62.7 on DeepSWE, three tests designed to evaluate coding-agent competence.
On broader agent benchmarks, the model scored 74.1 on Toolathlon-Verified, 25.7 on Agents’ Last Exam, and 31.8 on the public AutomationBench. DeepSeek’s internal tests showed similarly wide gains: 71.1 on DSBench-FullStack and 67.2 on DSBench-Hard, compared with 41.8 and 31.1, respectively, for the preview model.
DSBench-FullStack is described by the company as an internal full-stack development evaluation, while DSBench-Hard is a harder internal set of coding-agent problems.
Flexible Reasoning, Codex Support, and DSpark Decoding
DeepSeek V4 Pro 0813 introduces a three-tier reasoning_effort parameter: low for simple tasks, high for everyday agent workflows, and max for complex problems that call for deeper deliberation. This lets developers dial computational intensity up or down depending on the task at hand rather than relying on a single fixed reasoning mode.
Notably, the release does not ship with a Jinja-format chat template. Instead, DeepSeek has provided a dedicated encoding folder containing Python scripts and test cases that convert OpenAI-compatible messages into input strings and parse the model’s output, allowing developers to set the reasoning effort per request using the same workflow.
A key technical addition is DSpark, a new speculative decoding module built on top of the V4 Pro Preview architecture, designed to speed up inference. It can be enabled in vLLM through the speculative-config option, with DeepSeek’s reference example using seven speculative tokens and greedy draft sampling. In SGLang, the DSPARK speculative algorithm can be enabled without referencing a separate draft model, since the target and draft weights are drawn from the same checkpoint.
DeepSeek V4 Pro Price Increase Follows the Launch
For the first time, the company is introducing separate peak and off-peak rates, with off-peak pricing set at half the peak rate. For V4 Pro, cached input will cost 0.30 yuan per million tokens at peak and 0.15 yuan off-peak, uncached input will cost 9 yuan at peak and 4.5 yuan off-peak, and output will cost 27 yuan at peak versus 13.5 yuan off-peak. The revised pricing takes effect at 16:00 UTC on August 16, 2026.
The change marks a sharp reversal from the discounting DeepSeek used to drive V4 Pro usage during its preview phase, when the company offered a 75 percent promotional discount on the model through early May. Under the new peak pricing, V4-Pro output tokens are set to rise to roughly $3.96 per million tokens from a flat rate of about $0.87 per million previously.
DeepSeek V4 Pro Availability
DeepSeek V4 Pro 0813 is now available in Expert Mode through the DeepSeek app and website, as well as via the DeepSeek API using the existing model names. The company has released the repository and model weights under the MIT License and published API documentation to guide developers through the setup process for migrating from the preview build.
Context: A Company Under Pressure to Keep Its Edge
The official V4 Pro launch comes at a pivotal moment for DeepSeek. Its cheaper V4-Flash model, released earlier in August, performed better than many expected in independent testing, in some cases even outperforming the April preview of V4-Pro itself, an outcome that raised questions given that Pro is meant to be the company’s more capable, premium offering.
With V4-Pro now generally available, built-in Codex support, expanded reasoning controls, and a new speculative decoding system, DeepSeek is signaling its intent to compete not just on price but on agentic performance, an area increasingly seen as the next major battleground among frontier AI developers.