Cost anxiety
Last month, on a client project, I ran out of my monthly 10,000 Copilot Credits allotment (which costs approximately $100 USD). Limiting developers to 10,000 Copilot Credits per month is probably not enough for enterprise development, yet these limitations forced me to think about how best to use my tokens, specifically which models to use them with.
Token anxiety
Watching these credits being consumed was anxiety-provoking, especially as the context window filled up quickly while I desperately tried to get the AI to fix the creation of the Drupal model that had skidded off the rails. In just 2 hours, I consumed 2,000 credits, which forced me to research how best to get the most out of my credits/tokens. The most immediate mistake I made was spending 2 hours in one session, which created a huge context window that was being compacted as more and more tokens were consumed with each request and response. The biggest mistake I made was choosing the wrong model to begin with and, more specifically, using only one model and only one tier.
Choosing a model
The big three LLMs that I use are: Claude from Anthropic, GPT from OpenAI, and Gemini from Google.
Since Claude Code was the first coding harness to produce acceptable code, it is the default for many developers, but damn, Claude is expensive.
Gemini is said to be the cheapest model, but frankly, my one-off experiment didn't generate acceptable code. Simply put, I can live with an LLM getting a method's parameter order or data typing wrong, but when they start inventing nonexistent methods, I am out.
GPT has become my go-to model because the code is decent, readable, and refactorable.
Last week, OpenAI showed they were aware of the cost challenge when they adjusted their GPT 5.6 pricing, with the publication of advancing the price-performance frontier with GPT-5.6.
Choosing a tier
I want to be very clear that the goal of this post is NOT to make you start using GPT models, but to recognize that the pricing structure for GPT 5.6 Sol, Terra, and Luna defines which types of models we need, what the pricing should be, and how we should use them.
Starting today, GPT‑5.6 Luna, our fastest and most affordable model, will cost 80% less, while GPT‑5.6 Terra, our balanced model for everyday work, will cost 20% less. -- Advancing the price-performance frontier with GPT-5.6.
Below is the simplest summary of the GPT 5.6 pricing
- GPT-5.6 Sol - $5.00 / 1M tokens
Flagship model for ambitious agentic work - GPT-5.6 Terra - $2.00 / 1M tokens
Balanced model for efficient, high-volume work - GPT-5.6 Luna - $0.20 / 1M tokens
Fast, affordable model for everyday work
Sol is the most expensive, with Terra being 2.5x cheaper than Sol and Luna being 10x cheaper than Terra and 25x cheaper than Sol.
In simple terms, Luna is so damn cheap you have to try it, and I did.
Starting cheap
My quickest comparison between Terra and Luna was to have both models write a plan for a minor improvement. As an aside, word on the street is to always start in plan mode unless you are 100% confident the AI will complete a task correctly in a one-shot prompt.
Terra and Luna's plans were almost identical, though Luna mixed up the order of parameters in the SQL REPLACE function, which I also occasionally do. The fact that Luna was ⅒th the cost is huge, and I started to see a pattern and process for using all three models.
First off, I think that GPT 5.6 might be defining the ideal ratio for model pricing, which conceptually tells you to start cheap and scale up as needed. The cheap model is so inexpensive that, if it fails at a task, you can switch to the medium or highest-quality model without worrying about your initial expense.
A power-user trick and a huge potential cost-saver is this: if a cheap model fails you, your last prompt should ask it to summarize the initial prompt and the work you were trying to complete, then paste that response into a new session in the more expensive model. This pattern alone will probably lower your costs with the more advanced models. Still, over time, you will start to know when you need to start with a more expensive model.
General takeaway
The biggest takeaway I have is that, with the latest models, it's generally best to start with the cheapest one. I feel that the GPT team might have established that you need three tiers available, with costs going from a cheap tier, followed by a 10x-cost middle tier, and a 25x-cost high tier. Even with subscriptions, it might be worth getting comfortable with what the cheap models can and can't do, because word on the street is that the subscription approach to AI won't last forever.
Drupal takeaway
This post is pretty general, but I wanted to end with some Drupal-specific takeaways. As a community, we need to be aware of how the tangible cost of using AI affects who might be able to contribute. Things are moving so fast that I am sensing an open-weight, possibly locally running model will become our 'cheap' tier and daily workhorse in the next year or so.
I am unsure whether people are ready to hear this, but we might need to focus our AI best practices on a single cheap or free model to ensure it works best with Drupal to make AI contribution more accessible to all contributors.
AI-generated takeaway
I got the below response, with em dashes in all their glory, while asking Claude to break down GPT's pricing ratio and randomly asking it, "Metaphorically describe how the Luna to Tera to Sol ratio works," and I couldn't stop myself from sharing it with you
Think of it like climbing out of a small pond into an ocean.
Luna is a puddle — shallow, everywhere, costs you nothing to step in. You splash through it dozens of times a day without a second thought.
Terra is a lake — a real step up, deep enough to need a boat, but still something you visit often for serious but everyday journeys. Getting from the puddle to the lake is a huge leap (10x) — like stepping from your backyard into a whole different body of water.
Sol is the ocean — vast, powerful, built for the biggest voyages. Going from the lake to the ocean (2.5x) is a much gentler transition by comparison — more like sailing from a large lake into connected open water than jumping scales entirely.
So the shape of the journey is: a massive leap to get out of the shallow end, then a smoother glide into the deep end. The big psychological and financial "jump" happens early (Luna→Terra), while the jump from Terra→Sol feels more like a natural continuation than a shock — even though Sol is still the priciest and most powerful of the three.