Anthropic’s cheaper Opus 5 overtook its flagship Fable 5 in corporate spending within a month of Opus 5’s July 24 launch, according to Ramp, the corporate card and spend-management company. Ramp did not disclose Opus 5’s exact share of that spending. The shift shows buyers are no longer picking one model and sticking with it.
Ramp’s data, reported by Implicator.ai on August 23, found Fable supplied just 6% of the Anthropic tokens it could identify in July but drove 11.4% of attributed spending, a gap explained by Fable’s higher per-token price. By August 23, Fable’s share of identified Anthropic spending had settled near 11%. Ramp priced Fable at roughly $10 per million tokens in its comparison, about double what it assigned to GPT-5.6 Sol, and found Fable generated only about 75% as much attributed spending as Sol despite being Anthropic’s strongest generally available model at the time.
Vercel’s numbers complicate that picture. Its AI Gateway index, covering July 2026 production traffic, put Fable at 13.2% of all gateway spending, trailing only Opus 4.8 (the index predates Opus 5’s full-month availability). Nine in ten teams using Fable that month were new to it. Ramp leans toward technology companies and can’t explain why any single buyer chose a model; Vercel values traffic at list price rather than negotiated rates and doesn’t cover the full market. Neither source settles Fable’s actual standing, but both agree the mix moves fast: on Vercel’s gateway, changing models takes one line of code, and three in four teams running over 10 million tokens in both June and July had swapped out over 10% of the systems they route work to.
This is the real story, and it is easy to flatten: a cheaper model taking more of the spending pie is not the same as winning on cost per successful task. Opus 5 launched at $5 per million input tokens and $25 per million output tokens, half Fable’s rates. But a lower sticker price doesn’t guarantee a lower bill. A cheaper model can require more attempts, longer prompts, or heavier human review before a task actually clears, and each of those steps adds real cost that a per-token price tag hides. Anthropic’s own evaluations show Opus matching or beating Fable on bounded coding and office tasks once you tally what each finished job actually required, not just what one token cost, a narrower and more defensible claim than “Opus is winning.” Fable is still built for autonomous work that has to hold together across many linked steps over multiple days, a job Anthropic product lead Dianne Penn described as choosing “Opus 5 for value and Fable 5 for days-long, very autonomous projects.”
The pattern here matters more than either product’s individual share. Low switching costs mean businesses now treat models as a routed portfolio rather than a single vendor choice, sending routine requests to the cheapest system that clears the bar and reserving the expensive one for jobs that fail elsewhere or demand sustained autonomy. That changes what a frontier lab is actually selling: not one flagship product but a menu, priced and monitored task by task. Anthropic’s own growth backs up the portfolio reading rather than a Fable decline. Its annualized revenue rate reached $65 billion in July, up from $47 billion in May, and Ramp separately found 43.5% of U.S. businesses paying for Anthropic products in its broader July sample, ahead of the 39.7% paying for OpenAI.
Operators evaluating Anthropic contracts should ask for cost-per-completed-task figures, not just per-token rates, before assuming a cheaper model is actually cheaper to run at scale.
Reporting by Marcus Schuler for Implicator.ai, published August 23, 2026.
