GPT-3 launched at $60 per million tokens. By 2026, models beating its benchmark scores cost $0.06 per million tokens or less.1 A thousand-fold collapse in under five years. No commodity in computing history has deflated this fast.
Blended AI token prices fell 67% year over year, from $18.40 to $6.07 per million tokens between Q1 2025 and Q1 2026.2 And yet 73% of enterprises still blew through their AI budget last year.3 Uber’s CTO told the company in April that it had burned through its entire annual AI coding budget in four months.
Prices crashed. Bills went up. Seems paradoxical. Six forces are driving this.
Force one: China broke the price floor.
DeepSeek’s V4-Flash runs at roughly $0.14 input and $0.28 output per million tokens.4 Its nearest Western equivalent, GPT-4.1, runs at $2 and $8. That’s 14x cheaper on input, 29x cheaper on output, for competitive benchmark performance. DeepSeek’s architecture activates only about 37 billion of its 671 billion parameters per forward pass, which is the efficiency trick behind the price.5 The share of tokens routed to Chinese models by U.S. companies on the OpenRouter platform has climbed as high as 46% this year.6
Force two: nobody knows how to budget for a token.
Ask a finance leader how many tokens it takes to process one invoice. There’s no standard answer. It depends on document length, how much context the agent re-reads on every retry, how many tool calls it fires, whether the model reasons out loud or just answers. A cloud VM has a knowable, stable cost per hour. A token has no such anchor. That’s why enterprises keep building budgets off last year’s chatbot usage patterns and getting blindsided by agent consumption they never modeled.
Force three: agentic AI is not always cheaper than offshore.
The instinct in every boardroom right now is that agents beat an offshore FTE on cost. Not true. A generic AI agent runs $6 to $8 fully loaded per hour today, against roughly $10-15 for offshore human resource. But this comparison only holds for repetitive, protocol-following work with clear completion criteria and a human catching the exceptions.
It gets worse fast once you move up the complexity curve. A single-model call is one thing; multi-agent orchestration on complex processes is another. Coordinating multiple specialized agents on one task, planners, workers, verifiers, each handing off context to the next, can multiply token consumption 10 to 50x versus a naive single-pass estimate, because every retry or hand-off resends the accumulated conversation history. Outside the simple, repeatable lane, don’t assume AI beats humans on cost.
Force four: TCO, not token price, is the real number.
The model and compute costs is roughly 30% of total AI project spend. The rest of the real cost of enterprise AI is orchestration, RAG pipelines, monitoring, governance, human-in-the-loop review, integration (and I am not including change management). Anyone comparing providers on cost-per-million-tokens alone is comparing the cheapest part of the bill and ignoring the rest of it.
Force five: tokens are not outcomes, so why is everyone pricing like they are?
A token is not a unit of value delivered. It’s an obscure, unstandardized proxy for compute consumed, and it varies by tokenizer, by model, by vendor. Nobody can tell you what a token “did.” But both buyers and sellers want AI to deliver value with business outcome-tied pricing structures. But what’s being billed today, even under an “outcome-based” label, is still metered in tokens. Until tokens are translated and benchmarked to “task” or “outcome,” buyers are trading one opaque unit for another.
Force six: what happens to pricing when Anthropic and OpenAI go public?
Both companies have filed confidentially. Once quarterly earnings calls enter the picture, the pressure runs in two directions at once. Investors will want either margin expansion, which points toward price increases or tighter free-tier limits, or continued land-grab pricing to defend share against DeepSeek and the open-weight pack, which points the other way. So pricing gets more volatile, not more stable, once Wall Street is in the room.
The Big Picture. The winners won’t have the cheapest tokens. They’ll have the discipline not to waste them.
When the marginal cost of intelligence falls a thousand-fold but the true TCO doesn’t move nearly as fast, the constraint on enterprise AI stops being model access and starts being orchestration discipline. Coinbase is one of the more concrete examples of what that discipline looks like in practice: nearly halving its AI bill by routing every call to the smallest model that could do the job, while usage kept growing the whole time.
It was never going to be about who has the cheapest tokens. It’s about who has the operating discipline to not waste the cheap ones, the honesty to model TCO instead of sticker price, and the patience to wait for outcome-based pricing to actually mean what it claims. Enterprises that treat token budgets like a fixed software license will keep getting Uber’d.
Sources
1. a16z, LLMflation index (2024); EpochAI figures via Forbes (July 28, 2026).
4. Published API pricing compiled in layer3labs.io, AI Token Cost Chart (2026).
5. DeepSeek’s model documentation (DeepSeek-V3, Hugging Face).
6. CNBC, “Chinese AI models are gaining ground with U.S. companies” (July 7, 2026)
