.jpg?w=1024)
DeepSeek just released its strongest AI model yet, aiming directly at rivals with better performance and much lower pricing.
The Chinese artificial intelligence startup on Thursday launched the official version of its flagship model, DeepSeek-V4-Pro-0813, across its web interface, mobile app, and API. Designed to excel at autonomous AI agent tasks and software engineering, the release delivers capabilities competitive with leading models like Moonshot AI's Kimi K3, but at a fraction of the cost.
The general-availability release marks a substantial upgrade over the preview version introduced in April. According to DeepSeek, the new V4 Pro achieved a score of 62.7 on the DeepSWE benchmark for real-world software engineering, up sharply from 12.8 in preview testing. It also recorded an 87.9 on Terminal Bench 2.1.
Independent evaluations highlight DeepSeek's technical progress alongside its low pricing. Independent benchmarking firm Artificial Analysis awarded the reasoning version of V4 Pro a score of 53 on its Intelligence Index. While that trails Moonshot’s Kimi K3 score of 60 and Anthropic’s Claude Opus 5 score of 63, DeepSeek's pricing significantly undercuts its rivals.
On third-party platforms like OpenRouter, V4 Pro is listed at $0.435 per million input tokens and $0.87 per million output tokens. In contrast, Moonshot’s Kimi K3 costs $14 to $15 per million output tokens, while U.S. competitors such as Anthropic’s Claude Fable 5 run as high as $50 per million.
Open-source tooling and looming price hikes
DeepSeek also released DeepSeek Harness (dsh) v0.1 under the open-source MIT license, providing an alternative to Anthropic's Claude Code. Built on a modular architecture in which "everything is a plugin," the tool lets developers orchestrate autonomous coding agents across local file systems and development environments.
Alongside these launches, DeepSeek announced an overhaul to its API billing. The company will replace flat rates with peak and off-peak pricing, raising API costs between 50% and 1,100% depending on model type and usage hours. During peak hours, V4 Pro output will climb to $3.96 per million tokens, a sharp jump from current rates, but still below Western alternatives.
What this means for developers and businesses
DeepSeek’s strategy is becoming more nuanced. The company is raising API prices, but it is also trying to convince developers that V4-Pro delivers enough capability to replace much more expensive models in production systems.
That could matter for startups and enterprise teams running large-scale AI workloads, where inference costs often determine which model gets deployed. A model that is slightly weaker on some benchmarks but dramatically cheaper to operate can still be the better business decision.
The tradeoff is that DeepSeek is no longer competing only on ultra-low pricing. By moving V4-Pro into a premium tier while keeping it well below many U.S. competitors, the company is testing whether it can build a sustainable business rather than simply win a price war.
Related reading: For more on DeepSeek’s broader ambitions, see how the company is pushing into the AI chip race as China works to strengthen its position in the global semiconductor market.