DeepSeek V4 Pro, and the Computer It Needs

What is DeepSeek V4 Pro?

DeepSeek V4 Pro is DeepSeek's flagship model, and on August 13, 2026 it left preview as build DeepSeek-V4-Pro-0813, rolling out across the DeepSeek app, web, and API at once. The API name didn't change; anyone already calling deepseek-v4-pro woke up on the GA build.

The specs are the same shape as the preview: a 1-million-token context window, up to 384K output tokens, thinking and non-thinking modes, tool calling, JSON output, and both OpenAI- and Anthropic-compatible endpoints. What changed is the post-training. DeepSeek aimed this release squarely at agents, and said so plainly, calling out production environments rather than chat.

One thing V4 Pro doesn't do: images. There is no vision input, which rules it out for screenshot-driven work and anything that starts with a PDF scan.

DeepSeek V4 Pro benchmarks, and who measured them

DeepSeek published ten scores with the GA release. The agent numbers are the interesting ones: Terminal Bench 2.1 at 87.9, DeepSWE at 62.7, NL2Repo at 61.5, CyberGym at 83.3, Toolathlon-Verified at 74.1, and HLE at 42.7 without tools and 60.0 with them. Against the preview build, DeepSeek reports +15.8 on Terminal Bench and +49.9 on DeepSWE.

Those are vendor-reported, and worth reading as such. DeepSeek did release the agent framework it benchmarked with, though: DeepSeek Harness, MIT-licensed and shipped as a developer preview, so the harness half of the setup is at least inspectable. Independent replication is still pending.

The open weights are a separate story. DeepSeek publishes V4 Pro on Hugging Face under an MIT license at 1.6T parameters, but those weights were last updated on June 22, 2026 and the changelog announces 0813 as an API and app update, not a new checkpoint. If you self-host, you are not running the GA build.

DeepSeek V4 Pro pricing, and what changes on August 16

As published, V4 Pro runs $0.435 per million input tokens on a cache miss, $0.003625 on a cache hit, and $0.87 per million output tokens. On August 16, 2026 at 16:00 UTC, DeepSeek moves the whole V4 family to peak and off-peak billing: off-peak lands at $0.66 input and $1.98 output per million, peak at $1.32 and $3.96, with peak hours running 01:00 to 04:00 and 06:00 to 10:00 UTC. Off-peak is half of peak, which is the point of the split.

It is still cheap next to a frontier Western model, and it is also a price rise plus a clock you now have to think about. Developers on Hacker News noticed within hours, and the ones who had quietly standardised on V4 Pro were the least happy about it.

A model is only half of an agent

The most repeated argument in that thread had nothing to do with benchmarks: the harness carries as much weight as the model. Same weights, different scaffolding, different outcome, and the gap between two mainstream coding harnesses can be worth more than the gap between two model tiers. DeepSeek clearly agrees, or it wouldn't have open-sourced its own harness on the same day.

That is the part nobody sells you with an API key. An agent model that can run for hours needs a machine that stays up for hours: a real file system, a browser that keeps its session, a terminal that remembers what it did last turn, and somewhere for the output to land.

Where MoClaw fits

MoClaw is that machine. It's a managed cloud computer with a persistent disk, a browser, a shell, scheduling, and 50-plus skills already wired in, and DeepSeek runs on it as a platform-managed model, so there is no DeepSeek account, no API key to rotate, and no peak-hour arithmetic. You spend plan credits; MoClaw handles the provider side.

Today that platform-managed DeepSeek is V4 Flash, which is what MoClaw's Fast tier runs. If you want to put DeepSeek to work on a real computer right now, start a task on Fast, and watch this page for V4 Pro.