DeepSeek V4 Pro: Rock-bottom Cost Per Task, But Trailing Kimi K3
Set Trending Topics as a preferred source on Google.
DeepSeek rolled out the generally available version of its top model this week, listed internally as DeepSeek-V4-Pro-0813. The model is accessible via app, web and API; according to the company, nothing changes about how it is called – the model name deepseek-v4-pro is enough. The move was announced with nothing more than a short note on the website, without an accompanying technical report or blog post of the kind other labs usually publish. The update follows April’s preview version and the smaller V4 Flash, which drew attention at the end of July mainly for its cost efficiency.
Price as the main argument
The announcement centres on agent capabilities, which DeepSeek says have been “significantly enhanced”, particularly in production environments. There are also two technical additions: the DeepSeek API now natively supports OpenAI’s Responses API format and has been specifically adapted for Codex, with a configuration script provided for setup. In addition, thinking effort can be controlled in three levels for both V4 Pro and V4 Flash – low, high and max – depending on the complexity of the task.
V4 Pro is priced at roughly 44 US cents per million input tokens and 87 US cents per million output tokens. Benchmarking service Artificial Analysis rates the input price as somewhat expensive against a median of 33 cents, while calling the output price moderate – the median there is $1.20. On top of that comes a 99 percent cache discount.
More decisive is the maths per completed task: around 6 US cents per run on the Artificial Analysis Intelligence Index. That puts V4 Pro slightly above OpenAI’s lightweight GPT-5.6 Luna at 5 cents, but roughly 93 percent below Moonshot AI’s Kimi K3, which costs 84 cents per task. Evaluating the full index cost $135 in DeepSeek’s case.
From 16 August, 16:00 UTC, DeepSeek is also changing its API pricing structure: a peak/off-peak model will apply, with off-peak hours charged at half the regular rate. Last week the company had already announced a “significant” price increase due to high demand, advising users to plan their usage accordingly.
Where the model falls behind
V4 Pro scores 53 points on the Artificial Analysis Intelligence Index – clearly above the median of 27 for comparable models, but, according to a report by the South China Morning Post, four points behind Terra, the mid-tier model in OpenAI’s GPT-5.6 line-up, and seven points behind Kimi K3. It is on par with Zhipu AI’s GLM-5.2, released in June.
On the Vals Index compiled by analysis firm Vals AI, V4 Pro came in twelfth, behind OpenAI’s previous-generation GPT-5.5 and well behind Kimi K3 and Anthropic’s Claude Opus 5. Two weaknesses stood out in particular: tasks in a sandboxed terminal environment and building complex financial models in Excel. On social media, developers also reported that the model stops prematurely during long coding tasks.
On speed, by contrast, V4 Pro performs above average: 78 output tokens per second versus an average of 65. The model is comparatively talkative, though – it produced 130 million tokens for the Intelligence Index, against a median of 100 million. The context window spans one million tokens.
Strength in cybersecurity
The verdict is considerably better in one specialist area. According to Belgian security firm Aikido Security, V4 Pro found more system vulnerabilities than any other model tested – including Claude Opus 5 and Alibaba’s Qwen 3.8. Aikido researcher Philippe Dourassov qualifies this by pointing to the poor precision of those hits, meaning the model also raises plenty of false alarms.
In parallel, DeepSeek is working on its own “harness” – a software framework that lets language models operate as autonomous agents, comparable to Claude Code. In early August the company invited open-source developers to beta-test the DeepSeek Harness, having already registered a WeChat account for the team in July.
