In January 2023, accessing GPT-4-level intelligence cost $0.06 per 1,000 tokens. Today, you can get the same capability for $0.0015 — a 97.5% price drop in under 30 months.
This is not a pricing strategy. It is commoditization in real time — and it is the most underappreciated structural shift in the AI market right now. Every startup whose pitch deck contains the phrase "powered by the latest AI models" needs to rethink what they are actually selling.
The Price Collapse in Numbers
The scale of the drop is hard to internalize until you see it laid out:
State of the art at launch
83% cheaper than original GPT-4
Faster, cheaper, same capability tier
99.75% cheaper than GPT-4 at launch
Google racing to zero
Open source matches 2023 GPT-4 benchmarks
The open-source trajectory is even more aggressive. By the time most enterprise AI vendors finish their 2026 sales cycles, the models they're selling access to will be available for free.
What Gets Destroyed
Not everything in the AI stack survives commoditization. The casualties are predictable:
Pure-play API wrappers
If your core product is a cleaner interface on top of an OpenAI or Anthropic API with no proprietary data layer, you are one pricing update away from irrelevance. Dozens of $50M-valued AI writing tools discovered this in 2024.
Model-selection as a feature
Startups that differentiated on access to multiple models — 'we let you pick GPT-4, Claude, or Gemini' — found that as all three became commodities, the selection itself had no value.
Prompt engineering businesses
Charging for expertly crafted prompts had a 12-month business window. Once models improved enough to interpret natural language instructions accurately, the premium on prompt crafting collapsed.
Generic AI copilots
Horizontal AI assistants without vertical depth or proprietary context compete purely on model quality. When model quality equalizes, they compete on price — and lose.
What Actually Survives
The same dynamics that commoditized SaaS infrastructure 15 years ago now apply to AI models. The winners in that era were not the infrastructure providers — they were companies that built on top of cheap infrastructure to own workflows. The same playbook applies here.
Proprietary data flywheels
Unique data that improves with every user interaction cannot be replicated by switching models. This is the only durable moat in a commoditized model world.
Workflow lock-in
Companies integrated 18 months deep into an operational workflow have switching costs that no price decrease overcomes. The integration is the product.
Vertical-specific fine-tuning
A model trained on 10 years of radiology reports or insurance claim denials outperforms generic models in its domain regardless of the base model's benchmarks.
Regulated industry distribution
Healthcare, financial services, and defense require vendor relationships, compliance certifications, and procurement relationships that take years to build — not months.
The Infrastructure Layer Is the Real Play
If models are free, the next bottleneck is inference infrastructure — and that market is enormous. Groq, Cerebras, and a dozen well-funded competitors are racing to make inference not just cheap but instantaneous. The global AI inference market is projected to exceed $200B by 2028, growing 40% annually.
More importantly: when model costs trend toward zero, volume explodes. The economics shift from cost-per-query to cost-per-millisecond and cost-per-concurrent-session. That is an entirely different optimization problem — and the companies solving it at scale will extract significant value even in a world where the models themselves are commoditized.
As an investor, I am looking at two tiers: companies that own irreplaceable data in high-value verticals, and infrastructure plays capturing the volume explosion that cheap models enable. Everything in the middle — generic AI tools with no vertical depth and no infrastructure advantage — is a compression trade.
The model is not the product anymore.
The data, the workflow, and the distribution are the product. Build those — or you're selling a commodity.
Track AI market dynamics at Value Add VC's AI Landscape Dashboard. Originally published in the Trace Cohen newsletter.