Key takeaways:

• Ramp data: Fable 5 is only 6% of Anthropic tokens and 11.4% of dollars among US business customers.
• Claude Opus 5, launched in late July at roughly half Fable 5’s price, overtook it in Ramp-tracked spending within weeks.
• OpenAI’s Sol draws 25% of OpenAI token volume among Ramp customers — more than double Fable 5’s share.
• Ramp’s read: businesses have found a willingness-to-pay ceiling for frontier AI models.
• Operator posture: test a mid-tier model on your actual work before committing to flagship pricing.

Among businesses in Ramp’s US corporate-card customer base, Anthropic’s most capable model is not where the money is going. Ramp’s August 2026 AI Index shows Fable 5 accounted for just 6% of Anthropic tokens and 11.4% of Anthropic dollars spent by Ramp customers in July. Ramp explicitly notes its sample skews toward US businesses and tech-adjacent companies — actual Fable 5 adoption across all industries is likely even lower.

The Financial Times reported using the same Ramp data that Claude Opus 5 — launched in late July at roughly half Fable 5’s price per The Decoder — overtook Fable 5 in Ramp-tracked spending within weeks of launch.

Why Are Ramp’s Business Customers Routing Around Fable 5?

The price gap is the clearest explanation. Secondary reporting places Fable 5 at approximately $10 per million input tokens and $50 per million output tokens; Opus 5 is priced at roughly half those rates. OpenAI’s Sol — broadly comparable in market positioning — draws 25% of OpenAI token volume among Ramp customers at a lower price than Fable 5.

Ramp’s own interpretation: “more performance is not worth the price tag.” The index notes that encouraging businesses to adopt the latest frontier models requires capability gains that consistently outpace cheaper alternatives — a harder bar as capable mid-tier and open-weight options expand. Open-source and model-serving platforms now account for 6.1% of AI-spending businesses in Ramp’s sample, up from negligible a year ago.

Median AI spend among Ramp customers was $11.95 per employee in July; the top 1% paid $7,400. Most businesses in Ramp’s sample operate with modest AI budgets, and Fable 5’s pricing sits above where most can justify it for general workflows.

What Does the Ramp Data Tell Operators About Model Selection?

Interpret the signal carefully. Ramp’s sample is US businesses on a corporate card platform, likely skewing toward companies with formalized procurement. The spending pattern may not reflect every sector or company size. What it does suggest: paying a significant premium for a frontier model is not an automatic decision for most organizations represented here.

The practical implication is not that mid-tier models will work for your specific tasks — it’s worth finding out whether they do. Run 50–100 real work items through a mid-tier model and score them with the rubric your team already uses. Set your own pass/fail threshold based on the actual cost of errors in that workflow: what a mistake costs you in rework, compliance risk, or customer impact. Only then does the rate-card comparison become meaningful. Also ask whether your AI platform allows per-workflow model routing and spending caps — that control is worth having before any contract expansion.

See related context: AI token cost planning and OpenAI’s Sol pricing.

Watch next: Track Ramp’s monthly AI Index for model-mix movement — whether agentic and reasoning-heavy workloads drive a Fable 5 share recovery in Q4 would materially update the procurement calculus.

FAQ

Does low Fable 5 adoption in Ramp data mean Anthropic is losing enterprise market share overall?

Not overall. Ramp’s index shows Anthropic leads US business AI adoption at 43.5% among its customers, ahead of OpenAI at 39.7%. The concern is narrower: whether the highest-priced model tier can sustain its premium as capable cheaper alternatives expand. Overall Anthropic spend within Ramp’s sample is growing; the model-mix is shifting down-tier.

How should we decide which model tier our workflows actually need?

Test with real work items, not synthetic benchmarks. Score outputs using the rubric you already apply — accuracy, format compliance, downstream error rate, or whatever reflects actual consequences in your operation. Set your pass/fail bar based on what errors cost your business. Only once you have those numbers does the comparison between mid-tier and flagship pricing become a decision rather than a guess.

What should we watch before expanding an AI contract that includes flagship-tier models?

Watch Ramp’s AI Index for model-mix shifts as agentic workloads mature. Watch whether Anthropic adjusts Fable 5 pricing or introduces new tier options. And track open-weight model benchmarks — if open-source alternatives keep narrowing the capability gap, short contracts with per-workflow model flexibility become better procurement hygiene.