DeepSeek launches V4 Pro flagship model with agent upgrades, price hike

DeepSeek has moved its 1.6-trillion-parameter flagship model, V4 Pro 0813, from preview into general availability, pairing the release with an open-source agent framework and a pricing overhaul that ends the company's era of ultra-cheap API access.

Agent Performance and Harness Release

The official V4 Pro release, available through the DeepSeek app's Expert Mode and via API, closes a testing period that began in April and delivers dramatic improvements on agent benchmarks. DeepSeek's vendor-reported scores show Terminal-Bench 2.1 climbing from 72.1 to 87.9, CyberGym jumping from 52.7 to 83.3, and DeepSWE leaping from 12.8 to 62.7 — all attributed to improved post-training on an unchanged architecture. On Terminal-Bench 2.1, V4 Pro now sits just 0.1 points behind Anthropic's Fable 5. These scores have not been independently verified.memeburn

Alongside the model, DeepSeek released Harness, an MIT-licensed open-source agent framework built on a plugin-first architecture. The project hit 20,000 GitHub stars in roughly one hour on August 13 — a pace Chinese and Western tech press called the fastest ever recorded on the platform — and had accumulated over 141,000 stars by August 17. Harness positions DeepSeek directly against Anthropic's Claude Code and OpenAI's Codex by giving developers a fully inspectable, self-hostable alternative for building coding agents.opensourceforu+2

Pricing Reset Reveals User Geography

Starting August 16 at 16:00 UTC, DeepSeek introduced peak and off-peak billing across its V4 model family. For V4 Pro, peak-hour output tokens rose from $0.87 to $3.96 per million — roughly a 4.5x increase — while off-peak rates sit at $1.98. Some token categories saw increases exceeding 1,100% from previous flat rates.buttondown+1

The peak windows — 9:00 to 12:00 and 14:00 to 18:00 Beijing time, with a gap during the typical Chinese lunch break — suggest the bulk of DeepSeek's paying users remain in Asia. Converted to UTC, these hours fall entirely outside U.S. business hours. Similarweb data from March 2026 showed 44% of DeepSeek's desktop visitors came from China, with the U.S. accounting for just 5.68%.wccftech+1

Competitive Position

Even at peak rates, DeepSeek remains cheaper than Western rivals — Anthropic's Fable 5 charges $50 per million output tokens. But the gap has narrowed considerably from the months when DeepSeek was the industry's disruptive force on cost. The company raised approximately $7.4 billion in its first outside funding round in June and is building a gigawatt-scale data center in Inner Mongolia.memeburn

The combination of competitive agent benchmarks, an open-source framework challenging proprietary coding-agent stacks, and a pivot toward sustainable pricing marks a strategic shift. DeepSeek is no longer positioning itself primarily as the cheapest option — it is building a full stack.opensourceforu+1