Theo-brainUnofficial. Not affiliated with Theo or t3.gg.
Sign inYou
18:18
0:00

I thought AI was gonna get cheaper?

Sunday, Oct 11 · 18 min · about 6 min to read

Theo-brain watched all 18 min

Theo argues that AI did get cheaper, by more than the 10x Sam Altman promised, once you measure the cost of a fixed level of intelligence instead of the price per token. Bills are rising because models now do far bigger, longer jobs and we use them much more, not because intelligence got pricier.

Briefing is enoughThe notes carry the argument and the key numbers, so you can skip the video unless you want to hear Theo walk through it himself.

Tap any time to jump the video there.

Key takeaways

Price per token is a poor way to judge cost. Theo uses three measures: the price per token, how many tokens a model burns to finish a task, and what it costs to reach a given level of intelligence. The last two matter most.

come from. The second one is how many tokens it takes to complete a given piece of work. And the third is the types of work the models can complete.

Theo says DeepSeek V4 Flash matches GPT-5's level for far less. Even if it were 3x less token-efficient, he says it still nets out around 20x cheaper. That beats the goal of 10x cheaper intelligence at a given level.

changes here, but even if you assumed that Deep Seek was 3x less efficient, the difference still nets out to around a 20x improvement. So if the goal was to get 10 times cheaper intelligence at a given level, we blew that out the [ __ ] park. We're at 20 times cheaper.

Bills are going up because usage went up, not token prices. Models used to autocomplete one function. Now they take a bug screenshot and land the fix end to end, running longer with less hand-holding.

bills are getting higher. I'm not trying to gaslight you into thinking otherwise, but it's not because the token cost went up, it's because the usage went up. We are using these models so much more.

A cheap per-token model can still be expensive if it is token-hungry. Gemini 3.8 Flash is over 10x cheaper per token than top models, but its wasted tokens shrink the real cost gap to under 3x.

that it's doing. So even though it is comically cheaper per token, the token waste that 3 8 Flash has results in this gap going from over 10x to under 3x. And

Once a level of intelligence exists, it gets cheap fast. A level that cost $15 per task with o1-preview was matched by Grok-3 mini at roughly a hundredth of the price about four months later.

And that was September of 2024, and by February of 2025, Grok-3 mini had four points on top of it at under a hundredth the price. Sorry, roughly a hundredth the price. Pretty crazy that the price has dropped by almost 99% in just 4

One skill here is new to you.

Match the model to the job and the cost per task

What it is
Pick models by what a whole task costs, not by price per token. For jobs that read a lot and write little, input price dominates. For agent coding work, caching (reusing already-read text cheaply), reliability and token efficiency matter more.
Why Theo cares
The last few benchmark points are expensive. A score of 45 to 46 cost 82 cents per task, while the 50s cost $3.26, about three times more for a few points.
How to try it
  1. For bulk reading or sorting jobs, ask your agent to use the model with the cheapest input tokens.
  2. For coding-agent work, compare models on total cost per finished task, not the price list.
  3. Decide what level of quality the job really needs before paying for the top model.
what. At that point, optimizing for the cheapest model for the input token read is probably the right call. If you're doing agentic dev work, things like caching, reliability, and token efficiency suddenly matter a hell of a lot more. We're also not looking at this
massively. Like if you're happy with a 45 to a 46 as the score here, your cost will be 82 cents per task. Where if you want to bump up into the 50s, suddenly it's $3.26. So for just a few more points, you're spending three times more money. I still

GLM 5.3 Flash New

What it is
A cheap model Theo calls one of his favorite releases ever: 15 cents per million input tokens and 50 cents per million output, often half off.
Why Theo cares
At a 20th to a 30th of GPT-5's per-token price, he says it runs circles around GPT-5 for a lot of real-world agent work.
How to try it
  1. Try GLM 5.3 Flash for routine agent tasks you now send to pricier models.
  2. Compare the results and the bill on the same task against your usual model.
which I absolutely adore, it's one of my favorite model releases ever. This model is 15 cents per million input tokens and 50 cents per million out. It's also on
a 30th the cost of GPT-5. For a 20th to a 30th the price per token, JLM 5.3 Flash will run circles around GPT-5 for a lot of real-world agentic stuff.

Hot takes

Small models with good tools beat last year's smartest models. Theo would rather use GLM 5.3 Flash for everything than be stuck with GPT-5 for anything.

His reasoning: It handles vision, tool calls and hours-long agent work much better. Its only gap is less built-in knowledge, and with a search tool it would still beat GPT-5 because it searches and uses tools well.

3 Flash is a way, way better model than GPT 5. I would rather use GLM 5 3 Flash for everything than be stuck using GPT 5 for anything because 5 3 Flash has a lotbrain. Small models with good tools massively outperform last year's smartest models even with tools because we've gotten so much better at that.

Altman's "intelligence too cheap to measure" isn't happening as implied, but his promise of 10x cheaper intelligence has been blown out of the water.

His reasoning: The tasks people used AI for last year really are that much cheaper now. The difference is that we now use it for bigger, tougher things.

promise of intelligence 10 times cheaper has been blown out of the water. We absolutely achieved that. The tasks that we used AI for last year are actually that much cheaper. The difference is we can do bigger and tougher things with

How Theo thought about this

He flipped the question. Instead of asking how much intelligence a dollar buys, he fixed a level, GPT-5's capability, and checked what it costs today. He then assumed the cheap model was 3x less efficient, to keep his estimate conservative.

think about it though. I want to flip it. Instead of thinking about this as how much intelligence can you get per dollar, let's take a level of intelligence and see how much it costs over time. If we took at the capability

He owns a past mistake. Two years ago he made a video saying AI wouldn't keep improving, because benchmark scores had flattened. He now says the real problem was that we needed new benchmarks, and that we are still looking for ways to measure these levels of intelligence.

that we were looking at. That's because we needed new benchmarks. I was so wrong. I've never in my life had a take that was that bad. And IBut yeah, the point I'm trying to make is that we are still trying to find ways to measure these levels of intelligence.

Sign in to mark what you've watched and keep notes.

Have a question about this one? Ask about this video →

Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.