This Is an Architecture Update, Not Just a Pricing Update
OpenAI has reduced the price of GPT-5.6 Luna by 80%.
| Before | Now | |
|---|---|---|
| Input | $1.00 per million tokens | $0.20 per million tokens |
| Output | $6.00 per million tokens | $1.20 per million tokens |
That sounds like a pricing update.
For software developers, I think it is also an architecture update.
It is tempting to use the most powerful model for every task. Architecture decisions, bug fixes, documentation, classification, test generation, and background automation all get sent through the same expensive model.
We would never design the rest of our software like that.
We do not run every workload on the largest possible server. We choose the right infrastructure for the job and scale the expensive parts only when necessary.
AI workflows should work the same way.
Route Repetitive Work to the Right Model
A cheaper model such as Luna can handle the repetitive, high-volume work:
- Extracting structured information
- Summarizing documents and logs
- Generating routine tests
- Classifying issues
- Running background agent tasks
- Preparing context for a stronger model
The more capable models can then focus on ambiguous bugs, architecture decisions, and tasks where deeper reasoning actually matters.
This does not mean automatically choosing the cheapest model.
Token Price Is Not Task Price
Token price is not the same as task price.
When a cheaper model needs additional retries, produces unnecessarily long responses, or generates code that requires more repair work, the apparent savings can quickly disappear.
The metric that matters is not cost per token.
It is cost per successful outcome.
Cheaper Usage Makes More Workflows Practical
The price reduction is also reflected in usage limits for paid Codex and ChatGPT Work subscriptions. That makes it easier to experiment with longer-running agents and workflows that previously felt too expensive to run continuously.
That changes the kinds of automation developers can try. A workflow does not need to be perfectly optimized from day one if it is affordable enough to measure, improve, and run again.
Treat Intelligence Like Scalable Infrastructure
The interesting shift is not that AI is becoming cheaper.
It is that developers can start treating intelligence as another scalable infrastructure component: route tasks, measure results, and use expensive capacity only where it creates additional value.
Are you already routing tasks between different models, or are you still using one model for everything?