In early 2023, calling GPT-4 wasn't cheap. Developers paid roughly $30 per million input tokens — and $60 per million output tokens. At that price, running an AI-powered app at any meaningful scale was a serious budget line item. Building autonomous agents that ran 24/7 was essentially economically prohibited for anyone without venture money.

Today, in early 2026, you can access intelligence at or above that quality level for under $0.30 per million tokens. Some providers go lower. Open-source models running on commodity hardware get you there for the cost of electricity.

That's a 100x cost reduction in roughly two years. For context: it took the semiconductor industry 18 years to achieve a 100x reduction in transistor costs under Moore's Law. AI inference just did it in 24 months.

This is the most important story in AI right now, and almost nobody is framing it correctly. The coverage focuses on capability benchmarks — which model scores best on MMLU, which one codes better. But the cost collapse is what actually changes the world. Capability without affordability is a demo. Capability at commodity prices is infrastructure.

The Numbers: What Actually Happened to Prices

Let's be precise about the timeline, because the drop wasn't gradual — it happened in discrete jumps driven by specific events.

March 2023 — GPT-4 launches: $30/1M input tokens, $60/1M output. The model was extraordinary but priced for enterprise buyers who could absorb the cost.

November 2023 — GPT-4 Turbo: OpenAI cuts to $10/1M input, $30/1M output. A 3x reduction in eight months. The market started to notice that prices could move.

May 2024 — GPT-4o: $5/1M input, $15/1M output. Another 2x cut. More importantly, GPT-4o mini launched at $0.15/1M input — demonstrating that near-GPT-4-quality intelligence could be delivered at dramatically lower cost.

Late 2024 — The Efficiency Revolution: Meta releases Llama 3.3 70B, a model that matches GPT-4-class performance and runs on a single high-end consumer GPU. Groq offers inference at speeds and prices that would have seemed impossible a year earlier. The open-source ecosystem begins running serious models on hardware anyone can buy.

January 2025 — DeepSeek R1: A Chinese lab publishes a model matching OpenAI o1's reasoning performance, trained for a fraction of the cost, with API pricing at $0.55/1M input tokens. The industry goes into shock — not because of the model itself, but because it proved that frontier-level intelligence didn't require frontier-level spend.

2025-2026 — The Commodity Era: Prices continue compressing. Inference providers compete aggressively. Open models proliferate. Claude Haiku, Gemini Flash, and a dozen open alternatives deliver strong performance at prices that make AI a rounding error in most application budgets.

What Drove the Collapse

The cost drop didn't happen by accident. Four distinct forces combined, each compounding the others.

1. Hardware Gets Better (and Cheaper)

NVIDIA's H100 GPU was the workhorse of 2023 AI inference — and it was expensive to buy, expensive to operate, and in short supply. By 2024-2025, H100s became more available, but more importantly, the broader GPU ecosystem caught up. AMD's MI300 series offered competitive inference performance. Custom silicon from Google (TPUs), Amazon (Trainium/Inferentia), and Meta began handling massive inference workloads at lower per-token cost. The hardware bottleneck that kept inference expensive started to ease.

2. Model Efficiency Improved Dramatically

The models themselves got dramatically more efficient. Techniques that were research curiosities in 2022 became production infrastructure by 2024:

  • Quantization — Reducing model weights from 32-bit or 16-bit floats to 8-bit or 4-bit integers, cutting memory requirements by 4-8x with minimal quality loss. A model that once required 8 A100s can now run on 2.
  • Speculative decoding — Using a small draft model to predict tokens, then verifying with the large model in parallel. Speeds up inference 2-3x without changing output quality.
  • Flash Attention and KV cache optimizations — Algorithmic improvements to the attention mechanism that dramatically reduce memory bandwidth requirements for long contexts.
  • Mixture of Experts (MoE) — Architectures where only a fraction of the model's parameters activate for any given token, delivering large-model quality at small-model inference cost.

The result: the same quality intelligence now requires significantly less compute to generate.

3. Competition Broke the Pricing Floor

In 2023, OpenAI had enough of a capability lead that pricing power came with it. By 2024, Anthropic, Google, Meta (via open weights), Mistral, and a wave of startups had closed most of the gap. When multiple providers offer roughly equivalent intelligence, price becomes the competitive lever. OpenAI's price cuts weren't charity — they were competitive responses to a market that was diversifying fast.

DeepSeek R1 was the shock that crystallized this dynamic. It showed that a smaller team with a fraction of the training compute could match frontier performance. If DeepSeek can do it, the assumption that only a handful of massively-capitalized labs can build leading models gets shaky. That assumption was the implicit foundation of pricing power.

4. Scale Creates Its Own Efficiency

Inference at massive scale enables optimizations that don't work at small scale. Batching requests together, continuous batching, request routing to underutilized hardware, geographic load balancing — all of these become more effective as volume grows. As the major providers handled more requests, their cost-per-token fell even holding hardware and model efficiency constant. This is a compounding effect: lower prices bring more users, more users enable better optimization, better optimization enables lower prices.

What's Now Economically Viable That Wasn't Before

The technology didn't change. The economics did. And when economics change, the set of viable applications changes with them.

AI in Every Application

At $30/1M tokens, adding AI to a consumer app required a conscious decision about whether the feature was worth the cost. At $0.30/1M, it's a rounding error in hosting costs. This changes how developers think about AI integration — from "can we afford this feature?" to "why wouldn't we include this?" The question shifts from budget to value.

The practical consequence: AI capabilities are proliferating into applications where they previously couldn't pencil out. Customer support, content moderation, personalization, search enhancement, data extraction — every category of software application is gaining AI capabilities because the marginal cost of adding them dropped 100x.

Autonomous Agents Running 24/7

This is the one that changes the nature of AI itself. An agent that monitors a system, makes decisions, and takes actions needs to run continuously. At 2023 prices, a reasonably active agent processing a few thousand tokens per minute would cost thousands of dollars per month per agent. That's not a product — that's an experiment.

At current prices, the same agent costs under $50/month. That's not just cheaper — it's a different product category. You can now run agents that continuously monitor production systems, research markets, draft communications, coordinate workflows, and take actions — at costs comparable to a SaaS subscription. The economics of autonomous AI agents just became viable for most businesses.

Real-Time Voice and Multimodal Processing

Voice interfaces require low latency and continuous processing. Multimodal inputs — images, documents, screenshots — are token-expensive by nature. Both were cost-prohibitive at 2023 prices for most applications. At current prices, real-time AI voice assistants and rich multimodal interfaces are becoming table stakes rather than premium features.

AI for Emerging Markets and Smaller Businesses

The 2023 pricing tiers effectively put frontier AI out of reach for businesses in lower-income markets and for small operations without significant capital. Cost compression democratizes access. A small business in Southeast Asia or Eastern Europe can now access the same AI capabilities as a Fortune 500 company. The geographic and economic distribution of who can build with AI is expanding rapidly.

Where This Trajectory Leads

If prices have fallen 100x in two years, what does 2027 look like? There are a few things we can say with reasonable confidence.

The Jevons Paradox Will Play Out

The Jevons Paradox, from 19th century economics, observes that when a resource gets dramatically cheaper, total consumption of that resource increases rather than decreases. Cheap electricity didn't reduce energy use — it massively expanded it. Cheap cloud computing didn't reduce server costs — it enabled applications that wouldn't have existed at 1990s mainframe prices.

AI inference will follow the same pattern. As tokens get cheaper, developers use more of them. Longer contexts, more iterations, richer prompts, more aggressive agent loops, more embedded AI calls. Total inference spend will likely increase even as per-token prices continue falling. This is why the major labs are investing so aggressively in capacity — they're building for a world where inference demand is orders of magnitude larger than today.

The Moat Shifts

When AI intelligence is a commodity, the competitive advantage moves elsewhere. A 10% better model matters a lot when you have a monopoly on good models. It matters much less when three providers offer good-enough models at similar prices.

The moats that matter in a commodity intelligence world are: proprietary data (training signal that others can't replicate), distribution (being embedded in workflows users already have), product (the interface and experience layer), and trust (reliability, security, compliance). Pure model quality as a differentiator has a diminishing half-life.

Intelligence Becomes Infrastructure

The end state of this trajectory is AI intelligence as infrastructure — something like electricity or bandwidth. You don't think about where your electricity comes from or optimize for the most efficient electrons. You just use it. AI inference is moving toward that model: abundant, cheap, fungible, and assumed to be present in any serious application.

When that happens — when developers stop thinking about AI integration as a special decision and start treating it as table stakes — the application landscape transforms completely. Every piece of software will be AI-augmented the way every piece of software today is cloud-connected. The question isn't whether, it's how fast.

The Takeaway: Price Is the Story

The AI discourse obsesses over capability — which model is smartest, which benchmark it leads. That conversation matters, but it's incomplete. A 2x better model at the same price has a certain impact. A model at 1/100th the price — even if it's only equally good — restructures entire industries.

We are living through the fastest cost collapse in the history of intelligence-as-a-service. The $30/million-token world of 2023 and the $0.30/million-token world of 2026 are not the same world with a lower price tag. They're categorically different environments with different viable business models, different application possibilities, and different competitive dynamics.

The hardware is improving. The models are getting more efficient. The competition is intensifying. And the Jevons flywheel is just starting to spin.

Whatever your mental model of where AI is headed — make sure it accounts for the price. That number, more than any benchmark score, is what determines what AI becomes.