Credit: this is my take on I thought AI was gonna get cheaper?, an 18-minute video by Theo (t3.gg). The prices below are checked against the providers’ own pricing pages and Artificial Analysis. The opinions are mine.

The short version#

If you run anything on AI right now, your bill is probably going up, not down. That feels like a broken promise. In February 2025, Sam Altman wrote in Three Observations that “the cost to use a given level of AI falls about 10x every 12 months.” A few months later he said “intelligence too cheap to meter is well within grasp.”

Theo’s answer is that the promise held. We are just measuring the wrong thing.

The scary number#

Look only at the price tag and it does look backwards. GPT-5 lists at $1.25 per million input tokens and $10 per million output tokens on OpenAI’s pricing page. GPT-6 Astra, the new top model, lists at $10 in and $50 out. That is 8x more on input and 5x more on output.

Theo says his own usage went from roughly $800 to $1,000 across all of 2025 to days where he can burn more than $1,000 in tokens. Plenty of teams feel the same thing at smaller scale.

Why price per token is the wrong metric#

Theo breaks cost into three layers, and this is the part worth stealing:

  1. Price per token. The sticker price. It went up at the very top and down hard everywhere else. GLM-5.3 Flash lists at $0.15 in and $0.50 out, roughly 8x cheaper than GPT-5 on input and 20x cheaper on output, and Theo argues it beats GPT-5 at real agent work.
  2. Tokens per task. A cheap model that rambles can cost more than an expensive one that gets to the point. Theo’s example is Gemini Flash, which looks like a bargain per token and loses most of that edge once you count how many tokens it burns to finish the job.
  3. Cost at a fixed level of intelligence. Pick a capability level and track what it costs over time. This is where the 10x claim lives, and Theo says it has been beaten. His rough estimate is that GPT-5-level work is now about 20x cheaper.

Artificial Analysis now publishes a cost per task for every model it benchmarks, which makes this easy to check yourself. On its current numbers, GPT-6 Sol at max effort costs $1.06 per task, about half of what GPT-5.6 Sol cost for a similar score, according to its own write-up.

So why is the bill bigger?#

Because we use it for far more. A year ago the model finished one function. Now it takes a screenshot of a bug, finds the cause, and lands the fix, running for hours instead of seconds. Theo calls it a Jevons paradox: when steam power got cheaper, people burned more coal, not less.

The bill went up because the work went up. That is a sign the tools got good enough to trust with bigger jobs.

What I would do with this#

I build and host AI agents for small businesses, so I see this from the paying side. Three takeaways:

  • Stop defaulting to the top model. Most business tasks, like sorting email, drafting replies, and pulling data out of documents, sit well below frontier level. A Flash-class model does them at a fraction of the cost.
  • Budget per finished task, not per token. Track what a completed job costs end to end. That is the number that tells you whether the agent pays for itself.
  • Expect today’s frontier to be next year’s cheap tier. If a workflow is barely affordable on the best model now, design it anyway. The price for that level of capability is likely to fall fast.

Theo admits he once made a video saying AI had stopped improving and calls it the worst take of his career. Fair enough. The bet I would not make is that intelligence stays expensive.

Source video: Theo (t3.gg), I thought AI was gonna get cheaper?