I Spent $40 And Got A Greeting

I used to spend money on AI APIs for testing. Now I spend money on AI APIs and immediately regret every life choice that led me to that moment. The prices have gotten out of hand and I need to talk about it before I have a breakdown in the middle of a terminal window.

The Current State Of Things

Let me lay out the pricing because I cannot be the only person who looked at these numbers and assumed there was a decimal point error somewhere.

Model	Input (per 1M)	Output (per 1M)
Claude Opus 4.6	$5.00	$25.00
Claude Sonnet 4.6	$3.00	$15.00
GPT-5 Pro	$15.00	$120.00
GPT-5.4 Pro	$30.00	$180.00

I stared at GPT-5.4 Pro pricing for a full minute. Then I closed the tab. Then I opened it again because I thought I misread it. I did not misread it. Output tokens cost $180 per million. That is not a typo. That is a cry for help.

My $40 Experiment

I put forty dollars on Anthropic last week. I wanted to test something small. A quick API call. Maybe a hundred tokens out. Nothing fancy. Here is what happened.

At this rate my forty dollars will last approximately one productive afternoon. Then I will be back to training my 100K parameter models that answer math questions with fish facts. At least those are free to run.

The Reasoning Tax

GPT-5.4 Pro does not need to be this expensive. The model is known for overthinking and overachieving everything. It spends tokens reasoning before it answers. That is great for accuracy. It is terrible for my bank account.

They charge per output token. Reasoning tokens count as output tokens. So the smarter the model tries to be, the more I pay. This feels like penalizing someone for reading the instructions before taking a test.

I doubt they changed the base architecture much. They probably just LoRA fine-tuned it to reason longer. If the model size is the same, the compute cost should be similar. Setting the price higher because it thinks longer feels like charging me extra for the waiter thinking about my order before bringing the food.

More output tokens does not equal more value. Sometimes it just equals more thinking that I did not ask for.

The Nine Hundred Dollar Question

Let us talk about real costs. A single complex query in a Claude Code-like harness could easily hit 5 million tokens. Do the math. 180 times 5. That is nine hundred dollars. For one query.

Yes. Nine hundred dollars. I could rent a small apartment for that. I could buy a lot of GPUs. I could feed a small village. Instead I am paying OpenAI so their model can think about whether it should write a Python script or a Bash script for twenty minutes.

This pricing structure assumes everyone is a corporation with an unlimited budget. It ignores the students. It ignores the hobbyists. It ignores people like me who just want to build things without needing a loan approval process.

A Quick Shoutout To Groq

Not sponsored. Just a person who likes fast things. Groq is genuinely great. They run open models at speeds that feel illegal. I can query Llama based models and get responses before my brain finishes formulating the next question.

Their free tier is actually usable. Their API is straightforward. And most importantly, they do not make me feel like I am burning money with every token. This is what AI infrastructure should feel like. Fast, accessible, and not designed to drain my bank account.

If you are tired of waiting for responses or watching your credits disappear, give Groq a look. Again, not paid to say this. Just a tired hobbyist who appreciates when things work well and cost little.

My Tiny Rebellion Continues

This is why I keep going small. This is why I train my own models. This is why I released TMLM-Haiku-1 to zero downloads and still feel proud about it. Because at least when I run my own model the only cost is electricity and my dwindling hope.

I am not saying large models are bad. Sometimes you need the big guns. But when the big guns cost more than my monthly grocery budget, I am going to stick with my tiny confused models that think France has a capital but will not tell me what it is.

Also I am not putting another forty dollars on any API until I absolutely have to. My wallet has spoken. It said no.

---