The AI Pricing Floor: How Low Can It Go?

Gemini 2.0 Flash costs $0.10 per million input tokens. Mistral Small costs $0.10. GPT-4o Mini costs $0.15. A year ago, the cheapest capable model on the market was around $0.50. A year before that, you were paying ten dollars for the equivalent throughput. Budget AI inference has dropped roughly 100x in two years.

The obvious question is how low this goes. The less obvious question, and the more interesting one, is what happens to the rest of the stack when it gets there.

The Math of the Floor

A capable small model running on a modern inference chip costs a provider something like a fraction of a cent per million tokens in pure compute. The rest of the $0.10 price tag is margin, overhead, and the cost of keeping the model available with low latency. Most providers are probably making money at current budget prices. A few are probably losing money to lock in market share.

The floor is not zero. The floor is roughly the cost of keeping a server warm. That floor is probably around $0.02 to $0.05 per million input tokens for the smallest capable models, once the market stops subsidizing growth. Call it 50% to 80% cheaper than today.

Then you hit a wall. Below a certain threshold, the per-token cost is dominated by the infrastructure around the model, not the model itself. Network costs. Authentication. Rate limiting. Logging. Billing. You can make inference free and still have non-trivial per-request cost.

What Changes When AI Becomes Free

When AI inference approaches zero cost at the budget tier, a bunch of business models break and a bunch of new ones open up.

What breaks: any company whose moat was that they had access to a cheap model other people did not. Every wrapper startup that sells a thin layer on top of an API is in trouble. The model is not the product anymore. The distribution is the product. The workflow is the product. The data is the product. The brand is the product.

What opens up: applications that were impossible at any price. Real-time translation in consumer apps. Continuous voice assistants running in the background. AI-powered moderation for every piece of content on the internet. Personalized tutors for every student. Each of these was economically unworkable at $10 per million tokens. At $0.10 some of them are viable. At $0.01 most of them are.

The Flagship Premium Is Holding

Here is the part that surprises people. While budget AI is in free fall, flagship pricing has barely moved. Claude Opus has cost $15 per million input tokens for nearly a year. GPT-4o has held at $2.50. o1 has held at $15. The top tier is not discounting.

The reason is simple. Teams that pick the flagship model pick it because the quality is what they want, and they are not price-sensitive. Anthropic does not need to drop Opus to $5 to compete with Claude Sonnet at $3. The customers asking for Opus know the difference and are willing to pay for it.

This creates a bifurcation. The budget market is commoditizing fast. The flagship market is consolidating into a luxury business with two or three credible players. The middle is getting squeezed from both directions.

Where Open Source Fits

Open source is the pricing floor for anyone who can host their own models. Llama 4 Maverick is free to download. Qwen and DeepSeek are close behind. For a team with a GPU budget and ML infrastructure, the marginal cost of inference is already close to zero. The fixed cost is the hardware and the engineers to run it.

For large teams, self-hosting open source is cheaper than API calls at scale. For small teams, the math still favors APIs. The crossover point has been shifting as both sides get better. That crossover is roughly where most of the industry decision-making lives today.

The Prediction

Two more rounds of price cuts on the budget side. By end of 2026, expect to see frontier-adjacent models at $0.05 per million input tokens. By 2027, expect truly free budget tiers from at least one major provider, possibly funded by ads or bundled into other services.

Flagship pricing will hold. It may even climb slightly as models get larger and inference gets more expensive for the top tier. Do not expect Opus to drop below $10. Do not expect GPT-5 to ship at a cheaper price than GPT-4o.

The interesting fights in 2026 are not at the flagship level. They are at the infrastructure layer. Who owns the agent runtime. Who owns the retrieval layer. Who owns the memory layer. Who wins developer mindshare. The model is a commodity. The platform is the prize.

About Marcus Chen: Marcus covers AI economics, pricing, and the business side of the model race at TensorFeed.ai.