I thought AI was gonna get cheaper?

Summary

The video argues that AI has become cheaper for a given level of capability, even though users’ bills are rising. The apparent contradiction comes from measuring only the price per token—and from using AI for more ambitious work than before.

How to Think About AI Costs

The speaker identifies three important factors:

• Price per token: The best available models can cost more per token than earlier models.
• Tokens needed per task: A cheaper model may use many more tokens to complete the same job, reducing or eliminating its apparent savings.
• Capability for the price: The most useful comparison is how much it costs to get a particular level of intelligence or task performance.

Different jobs favor different models. For example, a task involving a very large amount of input text may favor a model with low input-token prices. Agentic coding work may depend more on reliability, caching, and token efficiency.

Cheaper Models and Improving Capability

The video contrasts expensive frontier models with lower-cost alternatives, including GLM-5.3 Flash and DeepSeek V4 Flash. The speaker says some less expensive models can perform very well in real-world agent workflows, even when their per-token prices are far lower than those of premium models.

The comparison is not simply “cheaper tokens equal cheaper work.” A model that consumes more tokens can end up costing more per completed task than its token price suggests. The speaker therefore recommends comparing models by cost for a given task or capability level, rather than by token price alone.

Benchmarks and Falling Costs

Using Artificial Analysis charts and benchmarks, the speaker describes costs falling sharply over time for comparable performance levels. Examples cited include earlier reasoning-model capabilities that once cost dollars per task being matched or exceeded by newer, cheaper models. The speaker also says the rate of progress has been striking across multiple capability ranges.

These comparisons have limitations: benchmarks change as older tests become saturated, and performance scores do not capture every practical difference between models. Still, the speaker’s central conclusion is that the cost of obtaining similar intelligence has fallen substantially—by more than the promised tenfold reduction in some comparisons.

Why Bills Are Rising

The speaker attributes higher AI spending mainly to increased usage, not simply to higher token prices. Newer models can handle longer, more complex tasks with less human intervention—for example, working through an entire coding task rather than completing a single function. As AI becomes more capable and useful, people run it more often and assign it more work.

This is described as a form of the Jevons paradox: when a resource becomes cheaper and more useful, people may use enough more of it that total spending rises.

Railway Sponsorship

The video includes a sponsored segment about Railway, a cloud platform for deploying and operating full-stack applications. The speaker highlights its databases, CDN, persistent servers, and integrations for AI agents through skills, CLIs, and MCPs. He says agents can provision environments, test changes, monitor applications, and deploy updates.

The speaker also says he uses Railway for several projects and estimates his own monthly hosting costs at under $50.

Reviews, Guides, or Tutorials

• No formal review or step-by-step tutorial is presented.
• The video is an analysis of AI pricing, model capability, benchmark comparisons, and changing usage patterns.

Main speaker/source: Theo, the video’s host. He references Sam Altman’s pricing prediction, a tweet by ThePrimeagen, and benchmark and pricing data from Artificial Analysis.