Anthropic's discounts leave a 7.5x-to-37.5x gap, and the routing code decides the rest

Routing is the part that sticks. It is a code change rather than a procurement negotiation: a classifier decides which requests go to the cheap tier and which escalate, and once that logic exists the premium model stops being the default and has to earn each call. The write-up describes the net effect as less traffic and less revenue reaching premium products ^[7]. Nobody revisits it at renewal, because there is nothing to renew.

Discounting does not undo that logic. Apply the reported quarter off average prices paid to the reported spread and it narrows from 10x-to-50x to roughly 7.5x to 37.5x ^[17]. Against a sub-$1 competitor, the new mid-tier is still five times the price on input and twenty-five times on output ^[18]. A team that moved a workload for a thirtyfold saving does not move it back for a fifth off.

Two cautions on the evidence. All of it traces to a single trade write-up that credits abc7news ^[1], and that write-up calls Anthropic's top tier "Fable 5" ^[3] while presenting Claude Opus 5 as the mid-tier addition ^[8]; nothing in the material reconciles the two names, so treat the tier labels as reported rather than confirmed. The direction of the pricing is consistent throughout, which is the load-bearing part.

The rest of the report is harder for the premium case than the headline number suggests. It has Chinese models leading on intelligence-per-dollar ^[14], price sensitivity as the dominant force in enterprise AI procurement with sustained pressure on high-margin API usage ^[2], and Anthropic's share on specific AI platforms declining year over year even as absolute users grow ^[13]. Lindy, which moved its usage to DeepSeek outright on economics ^[6], is the visible version of a call most buyers make partially and silently, by threshold rather than by contract.

Segmenting the product line answers a different question than the one buyers are asking. It hands the routing layer a cheaper Anthropic option to select, which protects volume by cannibalising price, and it leaves the tier that has to carry margin serving only the requests the cheap model failed. That is the smallest slice of the traffic and the one where a rival's next release does the most damage. The $47bn run rate ^[11] shows the volume is real; nothing in this material shows the price holding.