DeepSeek Launches V4-Pro and Open-Source Claude Code Rival, Then Hikes API Prices

DeepSeek spent the last year winning developers over with cheap, capable models. Now it wants to charge them more for the privilege while chasing a much bigger target: the software layer that lets AI agents actually do work instead of just answering questions.

The Chinese AI lab launched the official version of DeepSeek-V4-Pro on Thursday, an updated flagship model built for agentic workloads, according to VentureBeat. Alongside it came DeepSeek Harness v0.1, an open-source "agent harness" released under the MIT license that competes directly with Anthropic's Claude Code.

V4-Pro is now live across DeepSeek's web interface, mobile app and API, with native support for OpenAI's Responses API and integration with Codex, VentureBeat reported. DeepSeek is building compatibility with its rivals' ecosystems rather than walling itself off.

The price hike nobody's celebrating

Starting at 16:00 UTC on Sunday, Aug. 16 (12 p.m. ET), DeepSeek is scrapping its flat API pricing for a peak and off-peak structure, according to VentureBeat. Even the discounted off-peak rates will be substantially higher than what developers pay today.

DeepSeek built its reputation as the low-cost disruptor that spooked Silicon Valley and rattled markets back when it first emerged. Raising prices, even with a better model to justify it, is a real shift in posture. Developers who came for the bargain rates now have a decision to make before Sunday.

A harness where everything is a plugin

DeepSeek Harness, also called dsh, is built on a framework called Cordis with one guiding idea: "Everything is a plugin," according to VentureBeat. That covers models, tools, skills, sessions, sandboxes, filesystems, loops, orchestration and user interfaces. Instead of a fixed coding-agent product, developers can swap and extend nearly every component.

The New Stack, which also covered the release, confirmed the plugin-first architecture as the project's central design philosophy. The code is available now on GitHub for download.

DeepSeek is upfront that this isn't ready for prime time. The repository is explicitly labeled a developer preview and warns, in its own words, that "THERE WILL BE COMPATIBILITY-BREAKING CHANGES," according to VentureBeat. That's a meaningful caveat for any enterprise team thinking about building production systems on top of it right now.

The sustained bet behind the launch

According to reporting from briefs.co, DeepSeek formed a dedicated "DeepSeek Harness Team" ahead of this launch, with Tencent verifying an official WeChat account for the unit and job listings already posted to build it out. That points to a sustained organizational bet, not a one-off release. DeepSeek announced via WeChat on the evening of Aug. 12 that V4-Pro now handles more complex agentic work independently, briefs.co reported.

The timing lines up with genuine demand. A ZDNET survey of 138 developers who use AI coding tools found three out of four use Claude Code, with about a third using OpenAI's Codex and 22% using both. Engineers at Meta, Palo Alto Networks and Nvidia told ZDNET they value Claude Code specifically because it reads an entire codebase before editing and handles context across large, complex projects. That's the territory DeepSeek is now trying to enter with an open, modular alternative.

DeepSeek isn't alone in racing toward agents, either. Writer Inc. announced its own flagship model, Palmyra X6, this week, claiming its agent platform now runs 52% cheaper and 48% faster with a 10% quality improvement, according to SiliconAngle. Writer says X6 outscored Claude Opus 4.8, GPT-5.5 and Gemini 3.1 on its evaluations, at lower token prices than Anthropic and OpenAI charge. Whether those enterprise-facing benchmarks hold up outside Writer's own testing is an open question, since none of the competing labs have responded to the specific comparison.

Enterprises spent $143 billion on cloud infrastructure in the second quarter of 2026 alone, with providers like Meta now monetizing excess compute capacity to each other. Nick Patience of The Futurum Group told CIO Dive that "power and physical capacity, not model quality, are the scarce assets right now." That's the backdrop against which DeepSeek's price hike makes more sense. Compute isn't getting cheaper industry-wide, no matter how efficient any one lab's model gets.

For developers, the question is straightforward: Is V4-Pro plus Harness good enough to justify sticking around after DeepSeek's prices jump Sunday, or does the pricing shift push more of them toward Claude Code and Codex, the tools ZDNET's survey shows they already trust?