The Numbers
The frontier token price index sits roughly 84% below its March 2023 base as of September 2026. Google's Gemini 3.1 Flash prices at $0.10 per million input tokens and $0.40 per million output tokens — a 99.7% reduction versus GPT-4's 2023 launch pricing. The pricing floor for mainstream APIs sits near $0.20 per million input tokens, with DeepSeek V4 Flash cheaper still at roughly $0.14/$0.40 per million.
What Was Previously Uneconomical
Features that call a model on every page load, every search query, or every support message had a hard unit-economics ceiling at 2023-2024 pricing — the per-request cost made them a bad bet at real traffic volume. At current pricing, that math shifted by roughly an order of magnitude, which is why AI features scoped down or shelved purely for cost reasons back then are worth re-scoping now.
Where the Money Actually Goes Now
With token cost dropping toward near-zero for many tasks, the real cost center in an AI feature shifts to the engineering around the model call: retrieval quality (see What Is RAG), prompt and context design, and review or guardrail infrastructure — not the API bill. Cheap tokens make sloppy engineering cheap to run, not correct.
The Practical Implication
The open question for a new AI feature is no longer 'can we afford to call this on every request' — it's 'is the output actually good enough to ship.' That's an engineering and product question, not a budget one, and it's where an integration engagement's time is actually worth spending.