DeepSeek bursts into token volume; Anthropic dominates spending

What happened? DeepSeek V4 revolutionizes the token market

According to Vercel's AI Gateway Production Index (May 2026), the launch of DeepSeek V4 has caused a radical shift in token volume distribution. In April, DeepSeek accounted for less than 1% of tokens and less than 0.2% of spending. In May, its volume share jumped to 17%, making it the third-largest provider by volume, ahead of OpenAI. Almost all volume comes from two models: deepseek/deepseek-v4-flash and deepseek/deepseek-v4-pro, both launched in May.

DeepSeek V4 Flash pricing is $0.14 input / $0.28 output per million tokens, 20 to 50 times cheaper than comparable Anthropic models, and 8 to 12 times cheaper than other mid-range models like Qwen 3.6 Plus or Kimi K2.6. This price differential has driven massive adoption.

Why it matters? Low-cost quality for production

Historically, low-cost models failed to capture significant volume share because they didn't meet the quality threshold needed for production workloads. DeepSeek V4 is the first to achieve this. According to Vercel, teams that tested V4 against their existing evaluations found the output good enough to ship to production, not just cheap enough to test.

This milestone opens the door to a new category of 'value-tier' models that compete on quality, not just price. However, total AI spending continues to rise. The average cost per token increased in May, despite DeepSeek's pull effect, because work demanding frontier models grew faster than work that doesn't.

Consequences: market segmentation and pressure on leaders

The AI production market is polarizing. On one hand, low-cost models (DeepSeek) absorb large token volumes at low prices; on the other, frontier models (Anthropic) concentrate spending. Anthropic increased its token share from 26% to 32% and its spending share from 61% to 65%. OpenAI maintained its token share near 13%, but its spending share rose from 12% to 13%, indicating customers pay more per token on OpenAI.

The AI coding agent use case illustrates this divide: high-value tasks go to frontier models, while low-risk tasks are routed to cheap models. Intelligent routing has become a key practice for budget optimization.

The slow adoption of Google's Gemini 3.5 Flash, launched in May at a higher price than its predecessor, shows that teams don't automatically migrate to new models if the price isn't competitive or the quality benefit isn't clear.

What readers should know?

• DeepSeek V4 has validated the high-quality low-cost business model. We're likely to see more competitors in this segment.
• AI spending continues to grow. Although DeepSeek reduces costs, increased use of frontier models drives total spending upward.
• Model routing is essential. Companies must implement routing strategies to assign tasks to the most cost-effective model without sacrificing quality.
• Anthropic maintains its spending lead. With 65% of total spending and 70-80% in critical use cases, it remains the dominant premium provider.
• Quality remains the decisive factor. DeepSeek's success is not just due to price, but because it crossed the quality bar for production.

"Cost savings alone wouldn't have moved DeepSeek's volume so much in one month; teams found the output good enough to ship to production." — Vercel AI Gateway Production Index, June 2026.

Historical context and comparisons

The phenomenon recalls the SSD vs. HDD disruption: initially more expensive per GB, but speed advantages transformed the market. Here, DeepSeek offers drastically lower prices with sufficient quality, which could force leaders to adjust pricing or differentiate more clearly on quality. The reaction from Anthropic and OpenAI will be key in the coming months.