Open-Source AI Is Rising… So Why Hasn't Anthropic Felt It Yet?

A Striking Paradox in the AI Market

In a paradox worth examining, the AI market is witnessing an unusual phenomenon: open-source models are dominating in usage volume, while advanced commercial models maintain a firm grip on actual spending. This is what Jesse Zhang, CEO of Decagon, revealed in a recent analysis published under the provocative title: "Everyone Is Wrong About Open-Source AI in the Enterprise."

A Lifecycle, Not Direct Competition

The core idea Zhang puts forward redraws the relationship between cutting-edge frontier models and open-source models. Rather than viewing them as competitors in a zero-sum race, he frames them as two consecutive phases in a single lifecycle:

• Exploration Phase: Expensive, advanced models are used to test new use cases and validate their feasibility.
• Production Phase: Once these use cases mature and stabilize, they gradually migrate to more efficient, lower-cost open-source models.
• Continuous Renewal: Meanwhile, new use cases constantly emerge, cycling back to frontier models for the initial testing phase.

The conclusion of this view: "Frontier labs will continue to own the discovery phase, while open-source AI will increasingly dominate the production phase."

What Does the Data Say?

Available figures support this analysis in a striking way. Vercel's AI Gateway dashboard shows DeepSeek jumping to the top in token volume processed, surpassing one-third of total platform traffic in just a single week. Yet the picture shifts dramatically when looking at actual spending: Anthropic still captures more than half of total spending on that same platform.

The OpenRouter platform tells a similar story. The DeepSeek V4 Flash model leads in usage by processing 5.3 trillion tokens per week, but Anthropic's Opus 4.8 model costs users approximately 23 times more per million tokens by comparison. The arithmetic means Anthropic earns a far greater share of revenue despite lower usage volume.

A Tiered Economy Entrenching Itself

A two-tier economic model is taking shape before us—one that may well become a near-permanent feature of the AI landscape:

• Upper Tier: Advanced frontier models that dominate new and complex use cases, justifying premium pricing through exceptional capability and early discovery.
• Lower Tier: Open-source models that absorb mature production workloads and deliver greater operational efficiency at lower cost.

There are two plausible explanations for why this balance persists: first, the total market for AI tasks is growing faster than the rate at which use cases migrate to cheaper models, ensuring frontier models always find fertile new ground. Second, many use cases remain too complex for substitution with lower-cost models to be practically viable.

A Lesson Still Being Written

Less than a year ago, analysts were questioning whether foundation model labs would be reduced to mere commodity suppliers, with upper-layer applications reaping the rewards of their efforts. Part of that scenario has indeed materialized: specialized AI products have shifted to lighter-weight models. But the more important part reveals a different truth—frontier models have succeeded in capturing the most valuable segment of the market: the premium token price. And that advantage shows no sign of being threatened in the near term.

The anticipated launch of Nvidia's Nemotron model adds a new variable to this equation, but it has yet to change the rules of the game. The landscape is evolving, and the competition is real—yet Anthropic and its peers are still winning the rounds that matter most.

✦ بقلم فريق دروب أيديا