DeepSeek just took its flagship AI model out of testing, gave away the software that turns it into an autonomous worker, and made it more expensive to run at the same time.
The Chinese AI lab released the production version of V4 Pro, ending a preview period that had run for months. DeepSeek says the model now handles multi-step tasks, like writing and testing code on its own, far better than before, and it points to benchmark scores that beat Anthropic's Claude on several agent tests. There is a catch: those numbers come from DeepSeek's own testing, on software it has not released yet, and nobody outside the company has independently checked the finished version. An independent leaderboard run by Vals AI puts the new model second among open models, just behind a rival called Kimi K3, not first.
The more interesting move is the software DeepSeek gave away for free. It open-sourced Harness, the toolkit that turns a language model into an agent that can use files, run code, and complete tasks across many steps without a person guiding every move. This is the same category of product OpenAI and Anthropic sell as Codex and Claude Code. By making its version free and open, DeepSeek lets any company build its own AI agent on top of DeepSeek's cheap models instead of paying a Western provider for both the model and the agent software.
That generosity comes with a price increase attached. Starting this week, DeepSeek is charging different rates depending on the time of day, with usage during Chinese business hours costing roughly double the off-peak rate. The bigger hit is on cache pricing, the discount companies get when an AI system rereads the same file or document it already processed. That discount is shrinking sharply, which matters most for exactly the kind of repetitive, file-heavy agent work Harness is built to do.
This is DeepSeek's second pricing move in three months, after cutting prices in May. Cache rates are now higher than they were before that cut. The timing lines up with reports that DeepSeek is preparing for a stock listing in China at a valuation near seventy four billion dollars, and needs to show investors it can turn its enormous usage into real revenue rather than just market share.
For any business already running workloads through DeepSeek to save money, the lesson is simple: check your bill before assuming the old math still holds. DeepSeek remains far cheaper than paying for Claude or GPT directly, but the gap is narrowing, and the free agent software is likely the bigger long-term opportunity here. It lowers the cost of building your own AI assistant, even if the model underneath it keeps getting a little pricier.