Chinese AI models have narrowed the performance gap with leading American systems while offering dramatically lower costs. A new wave of releases from firms like DeepSeek, Moonshot AI and Zhipu has forced executives at U.S. companies to reconsider their spending on frontier models from Anthropic and OpenAI.
One study highlighted in The Information found that certain Anthropic models could prove cheaper to run than some Chinese alternatives under specific workloads. Yet fresh benchmarks tell a different story. DeepSeek’s V4-Flash version costs just three cents per test on average. Anthropic’s Claude Fable 5 runs at $3.15 for the same evaluation. That’s more than 100 times the expense.
Artificial Analysis conducted the tests. The research firm tracks real-world task costs that factor in not only token pricing but also how many steps a model needs to finish a job. DeepSeek charges 14 cents per million input tokens and 28 cents per million output tokens. Those figures make it one of the cheapest well-known models globally.
But. Performance still matters. V4-Flash scored 50 out of 100 on Artificial Analysis’s Intelligence Index. The index combines results from nine benchmarks covering coding, reasoning and workplace tasks. Claude Opus 5, Claude Fable 5 and OpenAI’s GPT-5.6 Sol each scored at least nine points higher. Moonshot’s Kimi K3 reached 57.
So companies face a choice. Pay premium rates for top scores. Or accept good-enough results at a fraction of the price. Many have started to shift.
OpenRouter, a platform that routes queries across dozens of models, saw Chinese models account for more than 30% of tokens processed in some weeks this summer. The share hit 46% at peaks. It averaged just 11% before. Justin Summerville, who works on data and analytics at the company, told CNBC that open-source Chinese models run 60% to 90% cheaper than leading Anthropic and OpenAI offerings.
Businesses have taken notice. Lindy, an AI automation startup, saved millions after switching workloads to DeepSeek. Vercel reported GLM 5.2 from Zhipu as its fastest-adopted model ever. Token volume jumped 27 times. Customer count grew 80 times.
Zhipu’s GLM 5.2 landed within a percentage point of Anthropic’s Opus 4.8 on a prominent agentic benchmark. It did so at roughly one-fifth the cost. Some researchers say the model matches top U.S. labs on certain cybersecurity tests. The New York Times described how Silicon Valley engineers flocked to the release. Rehaan Ahmad, co-founder of alphaXiv, used it for over a week and concluded the gap with American systems had grown “very slim.”
Brookings Institution scholar Ryan Chan put the lag at six to nine months. “They operate close to the top American frontier models,” he said in the CNBC report. “The new open source models are performing well and prove capable for all but the most complex LLM tasks.”
Costs for U.S. models have climbed. Token prices rose at several labs. Enterprises suddenly confronted unexpectedly high bills. That pressure created an opening. Chinese developers, facing export controls on advanced chips, focused on efficiency. They distilled knowledge from existing models. They optimized inference. Training runs stayed lean.
DeepSeek claimed its earlier V3 model trained for about $5.6 million using 2.8 million H800 GPU hours. U.S. labs often spend far more. Their closed models also bundle enterprise features, safety layers and uptime guarantees. Those extras justify higher prices for some buyers. Yet many developers simply want capable outputs without the markup.
Alibaba’s Qwen 3.8 Max reached 58 points on the Intelligence Index in recent tests reported by The Elec. The model costs $1.13 per task. Claude Opus 5 in medium mode scored 59 points at 72 cents. Kimi K3 from Moonshot hit 60 points at 84 cents. The price-performance curve has flattened.
Earlier this month Reuters detailed how DeepSeek aims to regain momentum after rivals like Moonshot and Alibaba grabbed attention. The company prepares a more powerful V4-Pro. Alibaba unveiled its largest model yet on the same day as some of these reports circulated. Competition inside China has grown fierce. ByteDance, MiniMax and others push similar low-cost options.
U.S. adoption carries risks. Concerns over data privacy, government ties in Beijing and potential intellectual property issues linger. Some companies avoid Chinese models for sensitive work. Others run them in isolated environments or for non-critical tasks. The calculus differs by industry.
Still the momentum builds. Six of the 10 most popular models on one major leaderboard came from China this summer. Open-weight releases let anyone host, optimize and compete on price. That dynamic exerts downward pressure U.S. closed models rarely face.
Anthropic has responded by highlighting strengths in reasoning, safety and enterprise readiness. Its models often require fewer follow-up prompts. They produce cleaner code in some developer tests. Those qualities command premiums when accuracy and reliability outweigh raw expense.
Yet the market has fragmented. Startups mix models. They route simple queries to cheap Chinese systems. They reserve frontier calls for complex problems. This hybrid approach cuts bills without sacrificing quality where it counts.
Researchers continue to test. Artificial Analysis updated its rankings in early August. Kimi K3 leads certain Chinese leaderboards with an 80.3 score on BenchLM as of mid-month. Qwen variants follow closely. The gap on raw capability has shrunk faster than many predicted even a year ago.
Policy makers watch closely. Export controls sought to slow Chinese progress. Instead they pushed local labs toward efficiency and open innovation. The result? Models that deliver solid performance at prices American developers find hard to ignore.
Enterprises outside the U.S. show even less hesitation. Developers in Europe, Asia and Latin America cite data sovereignty worries with American providers. They prefer open models they can audit and run locally. Chinese offerings fit that preference.
The original study covered by The Information suggested scenarios where Anthropic’s systems could undercut Chinese rivals on total cost of ownership. Factors like output quality, context handling and reduced need for human review played roles. Those nuances matter. A cheap model that generates reams of low-value text can end up costing more in engineering time.
Real-world deployments will decide winners. Early data from platforms like OpenRouter points to sustained growth for the low-cost tier. Whether that growth erodes margins at Anthropic, OpenAI and Google remains an open question. Supply constraints at U.S. labs have kept prices elevated. More compute capacity could change the equation.
For now the trend is clear. Chinese AI has moved from curiosity to viable alternative. Companies that ignore the shift risk paying too much for marginal gains. Those that experiment stand to save significantly while maintaining competitive output. The balance between cost and capability has tilted. Smart operators are adjusting fast.