Can DeepSeek Compete With OpenAI, Google and Anthropic? Low Prices Help, but Gaps and Chip Limits Remain

NEW YORK — Nearly two years after Chinese startup DeepSeek rattled global markets with a low-cost AI chatbot, the company has established itself as a serious challenger to American AI leaders on price and in some technical areas. But published benchmarks, hardware constraints and reliability problems suggest it still trails the top U.S. labs in several key respects.

The question of whether DeepSeek can truly compete with OpenAI, Google and Anthropic has taken on new urgency as the Hangzhou-based company prepares a major shift to Chinese-made chips and as its service suffered another outage this week.

Strong showing in coding

DeepSeek's latest flagship, V4-Pro, was released as a preview on April 24, 2026, and reached general availability on Aug. 13. It is a 1.6 trillion parameter model released under the MIT open-source license, allowing developers to download and run it themselves.

On several coding benchmarks published by DeepSeek, the model outperformed leading U.S. systems available at the time. V4-Pro scored 3,206 on the Codeforces competitive programming rating, ahead of OpenAI's GPT-5.4 at 3,168 and Google's Gemini 3.1 at 3,052.

On LiveCodeBench, another coding test, V4-Pro scored 93.5%, ahead of Gemini at 91.7% and Anthropic's Claude Opus 4.6 at 88.8%.

On SWE-bench Verified, a widely watched test of real-world software engineering, DeepSeek's model tied Gemini 3.1 Pro at 80.6%, though OpenAI's GPT-5.5 led at 88.7%.

Where it trails

The same comparisons show DeepSeek behind rivals in other areas. On Humanity's Last Exam, a difficult test of expert-level reasoning, V4-Pro scored 37.7%, below Claude at 40.0%, GPT-5.4 at 39.8% and Gemini 3.1 Pro at 44.4%.

On the GPQA Diamond science benchmark, DeepSeek reported 90.1% against Gemini's 94.3%. GPT-5.5 led on Terminal-Bench 2.0, which measures autonomous command-line tasks, at 82.7% compared with V4-Pro's 67.9%. Claude led on long-context retrieval, scoring 92.9 on one test against V4-Pro's 83.5.

Analysts have also cautioned that several of DeepSeek's published comparisons were made against older versions of rival models, such as Claude Opus 4.6 and GPT-5.4, rather than newer releases.

Independent testing has produced mixed results. In one independent coding benchmark that evaluated 24 models on the same software project, Anthropic's Claude Opus 4.7 scored 97 out of 100, while DeepSeek V4-Pro reached 89 only when run through a third-party tool.

Among smaller, faster models, an October ranking by an AI benchmark tracker placed DeepSeek V4.1 Flash at 39.5 on a widely cited intelligence index, behind Anthropic's Claude Haiku 5.5 at 43.4 and Google's Gemini 3.8 Flash at 40.9.

U.S. labs have continued to release new models through 2026, meaning the leaders are moving targets.

Price is the biggest weapon

Where DeepSeek clearly competes is cost. V4-Pro has been priced at about $3.48 per million tokens, compared with about $25 for Claude's top model, according to Digital Trends. Analysts estimated that DeepSeek charges roughly 11 times less than Gemini on input pricing while matching it on SWE-bench Verified.

Combined with open weights that companies can run on their own hardware, that pricing has made DeepSeek attractive for businesses seeking capable AI at lower cost and with more control over their data.

Some analysts describe this as a structural challenge to the business models of closed AI companies that charge premium prices for access to their most advanced systems.

The chip problem

DeepSeek's biggest long-term obstacle may be access to computing power. U.S. export controls have restricted Chinese companies from buying Nvidia's most advanced AI chips, which American labs use to train their largest models.

DeepSeek CEO Liang Wenfeng told investors he aims to rely on Huawei chips to train future models, with a batch of Huawei training chips expected to begin arriving as early as the fourth quarter of 2026, according to reports. The company plans to deploy at least 160,000 Huawei Ascend accelerators at a new data center in Inner Mongolia.

Liang estimated that training a model comparable to OpenAI's largest would require roughly 50,000 Nvidia GB300 chips or about 200,000 Huawei Ascend 950 chips, according to a report by The Information. That gap highlights the efficiency challenge DeepSeek faces.

The company optimized V4 for Huawei processors, but supply constraints have limited availability. Nvidia's share of China's data center AI chip market fell to 34% in the first half of 2026, while Huawei's rose to 28%, according to research firm IDC. Nvidia CEO Jensen Huang has criticized the export rules, saying they forced a U.S. company to "cede China's US$50 billion AI chip market to Huawei."

Reliability concerns

DeepSeek's consumer service has also faced repeated disruptions. The company recorded 12 incidents in the past 30 days, according to an AI service status tracker, and users reported widespread problems with its app and website on Wednesday evening, when the company acknowledged a partial outage affecting its chat service and API.

U.S. competitors also experience outages, but frequent downtime can weigh on enterprise customers that depend on stable service.

Trust and regulation

DeepSeek also faces scrutiny outside China. Several governments have restricted its use on official devices over data security concerns, and OpenAI raised concerns about the company in a 2025 letter to U.S. policymakers. Those concerns could limit its adoption among Western businesses and government agencies, regardless of performance.

Growing ambitions

DeepSeek is preparing for a possible initial public offering as soon as 2027, according to reports by Bloomberg and the Financial Times, which could provide capital to expand its computing resources.

The verdict

DeepSeek has shown it can match or beat U.S. rivals on specific coding benchmarks and dramatically undercut them on price, making it a genuine competitor for cost-conscious developers and businesses. But it still trails the top American models on several reasoning and agentic tests, faces significant hardware constraints and must overcome trust and reliability concerns in Western markets.

Whether it can close the remaining gaps may depend less on research talent than on whether Chinese chips can deliver the computing power needed to keep pace.