Ramp's July card data puts Opus 4.8 at 3.5 times Claude Fable's spend share

Somebody set a default model string in a config file, ran an eval against it, and moved on. Multiply that by a few thousand teams and you get a spend-share table. It records what passed a check the last time anyone looked, not what would win a fresh bake-off, which is how the frontier and the deployment drift apart.

Drew Breunig named the moment the drift started. Before Fable, he wrote, a better model would arrive at the same price or cheaper and paper over most harness and context problems, so tuning either felt like wasted effort ^[9]. Then Fable landed, still incredible and expensive, Opus was good enough for most of the code his team needed (as were 5.6, K3 and GLM), and they started asking what work went where ^[9].

Group the Ramp lines by generation and that question has a visible answer. The 4-series entries, four Opus versions plus two Sonnets and a Haiku, come to 47.9 percent; the three 5-series entries add up to 15.1, a ratio a little over three to one ^[3]. All ten lines together reach 63.0 percent, so more than a third of whatever Ramp is measuring sits outside the named models, which is reason to read any single figure as direction rather than market share ^[1].

Two dates argue against the harshest reading. July was Fable's first uninterrupted month on sale, after 19 days of the previous one spent suspended ^[7]. Opus 5 booked 3.5 percent off eight days of availability ^[6]. And Ramp is reading card bills at 70,000 companies ^[5], a channel that misses committed enterprise contracts, which is where a top tier would most plausibly sell.

Capability is not the thing in doubt here. Simon Willison handed Fable 5 a sandbox evaluation inside Claude Code for web, the container turned out to have no /dev/kvm and no nested virtualisation, and the model went and ran the real test battery on GitHub Actions runners through a temporary workflow it deleted before the final commit ^[11]. That is what the premium tier is sold on, and it coexists with a single-digit slice of Anthropic's own spend ^[6].

The revenue line moved the other way, up 38 percent in two months on the FT's sourcing ^[4], with OpenAI's figure in the same story past $40bn and up 35 percent in the quarter to date after 5.6 shipped in July ^[4]. The six-figure account list Anthropic showed investors accounts for at least $600m annualised at its floor, under 1 percent of the $65bn ^[5], so either those accounts spend well above the floor or most of the money is arriving from smaller buyers, and smaller buyers are the ones Ramp can see routing down.

For the team that has to pick a default on Monday, the two axes are whether your harness detects the cheaper model's failures on its own, and whether a failure costs more than the price gap. Detectable and cheap, route down and let the eval do the arguing. Detectable and expensive, route down with a retry on the top tier. Undetectable and cheap, route down and accept noise you will never measure. Undetectable and expensive is the only quadrant where the frontier price defends itself, because price only helps you when you can tell what you bought. Ramp's split is a record of other companies' cost tolerance in one month, and it does not transplant into your routing table.