Cost-per-task pricing emerges as Fable 5 lags at 11% of spend

OpenAI and Anthropic are pushing enterprises to evaluate AI models by cost-per-task rather than per-token, a defensive pivot as Ramp data shows Anthropic's most expensive model Fable 5 captured just 11% of Claude software spending in July.

"Low-cost models may have cheaper tokens, but getting good results may require more attempts, more time or more human review," Sarah Friar, OpenAI's chief financial officer, wrote in a July blog post. Anthropic's financial services lead Jonathan Pelosi echoed the view, calling token cost "only useful as a proxy metric" for what enterprises actually pay per completed task.

Ramp's July data shows Fable 5 accounted for 6% of tokens purchased from Anthropic but 11.4% of spend — a gap that reflects premium pricing rather than broad adoption. The model costs $10 per million input tokens and $50 per million output tokens, exactly double Claude Opus 4.8's $5 and $25 rates. By comparison, OpenAI's flagship GPT-5.6 Sol captured 25% of tokens and 23% of spend within OpenAI's customer base. Anthropic's customer penetration is growing — 43.5% of US companies now pay the company, up 1.1 points from June — but that growth hasn't translated into top-tier model adoption.

The pricing narrative shift comes as Anthropic prepares for a fall IPO with investors floating a $2 trillion valuation at roughly 30x projected 2026 revenues of $100-120 billion. If enterprises continue to cap spending on premium models, those revenue projections face headwinds. Meanwhile, DeepSeek's April open-source flagship model prices input and output tokens at a fraction of Anthropic's rates, and Meta and SpaceX are pushing even lower-cost alternatives.

Token Price vs. Task Cost: The New Pricing Battleground

The metric shift is a direct response to competitive pressure from low-cost token providers. DeepSeek's open-source flagship, released in April, prices input and output tokens at a fraction of Anthropic's most advanced product. D.A. Davidson technology research head Gil Luria framed the US labs' argument bluntly: "American labs want to convey that using open-source tokens is like using cheap toilet paper — it does clean up, but you need more of it, and there are other unpleasant side effects."

Third-party benchmark platforms are gaining traction as enterprises seek independent validation. Vals AI co-founder and CEO Rayan Krishnan noted that when labs self-report cost-per-task figures, "they often use internal benchmarks, so true apples-to-apples comparisons aren't possible." By per-task cost, Anthropic and OpenAI models remain the most expensive on the market, with Claude Fable 5 at the top — though both companies offer cheaper alternatives like OpenAI's GPT-5.6 Luna.

Enterprise Budgets Cap Premium Model Adoption

The Ramp data suggests a structural ceiling on what companies will pay for AI. In July, median AI spend per employee was $7,400 for the top 1% of spenders, $650 for the top 10%, and just $11.95 for the median company. While leading enterprises invest heavily, that money isn't flowing disproportionately toward the most expensive models. Companies are increasingly using cheaper models for routine tasks and reserving top-tier models for high-stakes work like large-scale code migrations.

The shift to usage-based billing is accelerating across the industry. Anthropic now bills business customers based on actual AI usage, with varying credit allotments per subscription tier before pay-as-you-go kicks in. OpenAI has introduced usage limits for its coding agent, and Microsoft's GitHub has rolled out usage-based pricing after users burned through monthly quotas. Walmart has capped staffers' use of an in-house AI agent, and Uber limits each employee's monthly spending on AI coding tools to $1,500 per tool.

The stakes are significant for Anthropic's IPO. The company's Claude Code and Claude Cowork products have become the fastest-growing enterprise products in tech history, generating over $50 billion in revenue as of June. But if enterprises continue to treat premium models as a niche tool rather than a default, the $100-120 billion revenue projections for 2026 that underpin the $2 trillion valuation talk may prove optimistic. Anthropic shares, if they debut at 30x revenues, would need sustained enterprise willingness to pay premium prices for top-tier models — a willingness that Ramp's data suggests has limits.

This article is for informational purposes only and does not constitute investment advice.