Inference-First GTM: Re-Framing Compute as Customer Acquisition Cost (CAC)
Update (Oct 7–8, 2026) — the model-layer price war that opened the week of Sep 22 continued: Anthropic's Claude Haiku 5.5 launched at roughly one-tenth the per-token price of its predecessor for most requests, and Mistral put its new trillion-parameter flagship on sale at half price. The cost of an AI answer keeps collapsing — and with it, the effective CAC of inference-subsidized GTM motions.
Per PYMNTS (Oct 8): Anthropic's Claude Haiku 5.5 (launched Wednesday Oct 7) "costs one-tenth the price of its predecessor, Claude Haiku 4.5, for most requests and about 75% less on average." The concrete token math: $0.10 per million input tokens and $0.50 per million output tokens, versus $1 and $5 respectively for Haiku 4.5. The model targets exactly the high-volume, GTM-adjacent workloads — "customer support, summaries and data lookups" — where inference cost is the binding constraint on usage-based pricing. Anthropic also claims Haiku 5.5 "outscored [GPT-6] Luna on every benchmark both models were tested on," including 72.4% vs 48.9% on a computer-use test; a customer, financial-research firm Rogo, offered the task-level proof: the model "goes into the 10-K and pulls the segment revenue line the deck needs."
This follows September's Opus 5.5 debut at 40% below Opus 5. The pattern across labs: price cuts are no longer annual events but launch features — each new model generation ships with a step-function price reduction as the headline. Mistral's half-price launch period for Mistral Large 4 (Oct 6) shows the same tactic at the open-weights layer.
For the inference-as-CAC thesis: every dollar of model cost now buys ~10x the output volume it did a generation ago on small models. Startups running agent-heavy, usage-based products can subsidize customer acquisition with inference and still hold margins — the way cloud vendors once subsidized pilots with compute credits1. The buyer's question shifts from "can we afford the AI" to PYMNTS's own framing of the big picture: "which model to trust with their customers, and where it should run." See Open Weights as a Competitive Wedge Against Incumbents for the open-weights flank of the same price collapse.
-
An instance of Collapsing inference prices turn compute giveaways into the customer-acquisition budget. — Step-function model price cuts make subsidized tokens an affordable customer-acquisition budget, the precise mechanism the theme states. ↩︎