Anthropic's newest model gets 8% of its own spend

Anthropic's model spend runs 28% to Opus 4.8, an older release, versus 8% to Fable 5, the newest and most capable model it ships. The 20-point gap upends a common assumption. Agent teams price for the frontier model, but production traffic runs on the release two versions back.

Anthropic's annualized revenue reached $65 billion in July, up from $47 billion in May, per Simon Willison's analysis of an FT report. The company told investors it expects a profitable third quarter, and six thousand customers now spend $100,000 or more a year with it. OpenAI's annualized revenue is over $40 billion, up 35% quarter to date after GPT-5.6 launched in July.

That breakdown comes from the Ramp AI index, which reads model adoption off the card and billing data of 70,000 companies. Opus 5 draws just 3.5% of Anthropic spend in the same data.

Ramp counts corporate card spend, which misses the direct enterprise contracts where those 6,000 six-figure accounts likely sit. The FT's revenue figures come from unnamed sources, not a filing. Both gaps probably understate Fable's real share, but an 8-to-28 split survives either correction.

Anyone running an agent framework with a model fallback chain should check which model it calls under load, then price that against the flagship rate. Most fallback chains get configured once and never audited, so real usage may sit closer to Ramp's numbers than assumed.