The Collapse of AI API Pricing

The Collapse of AI API Pricing

Frontier-model API pricing has fallen roughly 85-95% per unit of equivalent capability since early 2024. A task that cost $15 per million output tokens on the best model available then now costs $2-4 on a model that scores higher on every public benchmark. This is not a discount cycle — it is the commoditization of raw inference, and it relocates the margin.

Root cause

• Supply caught up. GPT-class capability is served by at least six independent providers plus a dozen strong open-weight models. Switching is an afternoon of work.
• Serving got cheaper. Speculative decoding, KV-cache reuse, MoE routing and FP8 serving cut cost-per-token 3-5x at constant quality.
• Prompt caching changed the unit economics. Cached input tokens bill at 10-25% of standard rate, so the effective blended price fell faster than the headline price for agent and RAG workloads.

Capability-adjusted $/Mtok (best model of the quarter, indexed to 100)
100 | #
 60 |     #
 35 |          #
 18 |               #
  9 |                    #    #
    +----------------------------
    24Q1  24Q3  25Q1  25Q3  26Q2

Decision matrix

Actor	Exposure	Response
Model labs	Raw-token revenue per user falls faster than usage grows	Move up-stack: hosted agents, evals, fine-tuning, priority capacity
Inference resellers	Margin toward zero on undifferentiated serving	Compete on latency SLA, region, compliance — not price
Application companies	COGS per request drops — a one-time tailwind, not a moat	Bank the margin; assume competitors get the same input cost
Local-hardware buyers	Cloud break-even moves out 6-18 months	Re-run the TCO before buying a rig sized for last year's prices

Playbook

1. Re-price your product against current, not historical, token cost.
2. Instrument blended effective price — measure cache hit rate, route accordingly. 70% cache hits beats a 20% list discount.
3. Keep a warm second provider. Capability parity makes switching a lever you control.
4. Defer local-inference capex unless a residency, latency or privacy constraint forces it.