Skip to content

DeepSeek and Alibaba Release Frontier Models with Dramatically Reduced Costs

Bottom line: Chinese AI models reach the performance level of Western frontier models at 1–10 times lower operating costs while leveraging domestic chip infrastructure.

Chinese companies DeepSeek and Alibaba have launched new AI models that compete in performance with US systems but operate at a fraction of the cost. Both rely on proprietary hardware infrastructure and thereby circumvent US trade restrictions.

DeepSeek transitioned its V4 model to general availability on July 19, 2026, three months after an initial preview. On the SWE-bench Verified benchmark for software development tasks, V4 achieved a score of 80.6 percent, approaching the performance level of Claude Opus 4.8 and GPT-5.6 Sol. Simultaneously, Alibaba unveiled the open-weight model Qwen 3.8 with 2.4 trillion parameters, which according to the company differs only marginally from the proprietary model Fable 5. Chinese startup Moonshot had shortly before released an open-weight model called Kimi K3 with 2.8 trillion parameters.

Pricing is critical for infrastructure teams: DeepSeek V4 Pro costs $0.87 per million tokens during off-peak hours, while Anthropic’s Fable 5 costs $50 per million tokens — a price advantage by a factor of 57. Alibaba is offering Qwen 3.8 in preview at one-tenth its regular rates. For CTOs, this represents a substantial reduction in inference costs when operating frontier models, provided latency requirements and data sovereignty permit.

Both companies are strategically reducing their dependence on US hardware. DeepSeek is optimizing V4 for Huawei’s Ascend processors instead of Nvidia chips and is developing its own inference chip in parallel. Alibaba operates its infrastructure with the in-house AI processor Zhenwu M890. This hardware diversification mitigates vulnerability to US export controls and chip embargoes. For product architecture, the takeaway is: to reduce costs and minimize supply chain risks, one must engage with local inference options and open model weights.


Source: www.it-daily.net · Published July 20, 2026
Lumi AI News — AI-assisted curation pursuant to Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.

Share on: