This lesson's interactive features are locked, please to keep using them
Open-Weight Models
Have another look at this "cheaper models" table using the developers' standard API prices:
| Developer | Model ID | Input Cost (1M Tokens) | Output Cost (1M Tokens) |
|---|---|---|---|
| OpenAI | gpt-5.6-luna |
$0.20 | $1.20 |
| Anthropic | claude-haiku-4-5 |
$1.00 | $5.00 |
gemini-3.8-flash |
$0.75 | $3.75 | |
| Moonshot AI | kimi-k2.6 |
$0.95 | $4.00 |
Google's introductory pricing lasts through December 31, 2026, then rises to $1.50 input and $7.50 output per million tokens.
The kimi-k2.6 model is open-weight, as is qwen3-coder. That means the numerical parameters, the "weights," are available to download and run.
It's kind of like open-source software... but not quite. The training data and training process can still be private. So, while you can run and fine-tune the model, reproducing it from scratch isn't really possible without the code and data used to train it.
What Are Weights?
This isn't a course on LLM architecture, but at a high level the process looks like this:
- Gather a massive dataset – Reddit posts, GitHub repos, tweets, Wikipedia articles, etc.
- Run that text through a training process that adjusts the model's internal parameters, its "weights," to reduce prediction error.
- The end result is a giant file full of numbers that encode the patterns the model learned during training.
Then, to use the model, you feed it input text and it generates output based on those learned patterns.
When people talk about "AI costs," they usually mean one of two things:
- Training costs: the cost of turning the training data into model weights. This is a massive upfront expense for labs like Anthropic and xAI – sometimes hundreds of millions of dollars for a frontier model.
- Inference costs: the cost of using the trained model to generate predictions. This is what you're paying for when you use OpenCode Go or any other provider to code with AI.
For an AI lab to turn a profit, they'd need to charge users enough to cover the costs of running their models (inference) and to recoup development/training costs. So far this has proven quite difficult!