Anthropic’s best AI model struggles to attract users as cheaper tools thrive

Why is the strongest model the least used?

Simon Willison recently shared data from the Financial Times and the Ramp AI index that reveals a counterintuitive trend: despite Anthropic's highly anticipated flagship models Fable 5 and Opus 5, their actual enterprise spending share in July was only 8.0% and 3.5% respectively. Meanwhile, the previous generation Opus 4.8 still dominates with 28%, while older Sonnet and legacy models make up the rest. This signals a clear shift: the AI market's hype phase is over, and we have entered the ROI phase.

Why are the best models struggling to gain traction?

You might assume enterprises would pay a 3x premium for a 1% accuracy boost, but most business use cases simply don't need peak performance. For code generation, document summarization, and customer support, older models like Opus 4.8 or Sonnet 4.6 are already overqualified. When marginal performance gains no longer translate to marginal business value, CTOs shift their focus from chasing benchmarks to optimizing cost-efficiency. Fable 5 and Opus 5 are underperforming not because they lack capability, but because their pricing misaligns with real-world demand. High per-token costs make them economically unviable for routine tasks, while complex workflows now rely on hybrid routing strategies that reserve flagship models for critical decision points only. This architectural分流 naturally dilutes flagship billing visibility.

The revenue surge behind the scenes

Anthropic's annualized revenue hit $65 billion in July, with Q3 profitability expected. What does this tell us? The monetization logic of AI has fundamentally changed. Instead of relying on a few marquee enterprise deals, growth is now driven by 6,000 customers spending over $100,000 annually, plus millions of smaller developers. OpenAI mirrors this trend, with GPT 5.6 pushing its annualized revenue past $40 billion. AI is no longer a luxury product; it's becoming utility infrastructure. Legacy models may have lower unit prices, but their stability, affordability, and mature ecosystems make them the workhorses of enterprise cost optimization.

What this means for developers

For technical decision-makers, this data offers three actionable insights. First, stop chasing the newest and strongest. Ask whether your use case truly requires Opus 5-level reasoning. For RAG pipelines or standard text processing, Sonnet or Haiku will likely deliver better ROI. Second, model routing is becoming standard practice. Future AI architectures won't rely on single models. Smart systems will use a dispatcher layer: cheap models for simple tasks, flagships for complex reasoning, and fallback mechanisms for low-confidence outputs. Third, measure cost per completed task, not just price per thousand tokens. API pricing is just the surface. Real ROI comes from calculating which model minimizes total cost, latency, and human intervention across a full business workflow.

A counterintuitive perspective: the competition is shifting from capability to ecosystem

While everyone obsesses over benchmark scores, billing data tells a different story. The real battleground is no longer lab performance, but enterprise spending patterns. As foundation models converge in capability, competition will pivot to toolchain maturity, developer experience, compliance support, and pricing transparency. Anthropic and OpenAI are both printing money, but through different playbooks. Anthropic thrives on enterprise lock-in and legacy model tail revenue, while OpenAI rides the GPT 5.6 adoption wave. But regardless of strategy, the ultimate winners will be companies that make AI as reliable, affordable, and accessible as electricity. So next time you hear about a model breaking world records, ask yourself: does its billing data reflect real adoption? In the AI era, affordability and usability matter more than raw speed.