Rebuild a weekend AI stack after the GPT-5.6 Luna cut

The price cut changes the default, not the whole stack

You have a weekend project that works, but the model bill is a small thorn. The GPT-5.6 Luna pricing change gives you a reason to rebuild the stack without spending a Saturday.

GPT-5.6 Luna's price was cut by 80% on July 31, 2026. Input pricing fell from $1.00 to $0.20 per million tokens. The GPT-5.6 family has three tiers: Sol for sub-200ms latency, Terra for standard production use, and Luna for 1.5M-token agent capability.

GPT-5.6 Terra pricing fell by 20% at the same time as the GPT-5.6 Luna reduction. The developer price of the frontier GPT-5.6 Sol model fell by more than 20%.

On August 6, 2026, OpenAI said GPT-5.6 Luna would serve as the standard option for ChatGPT's Free and Go plans, and unlimited text chats would arrive the next week.

GPT-5.6 Luna remained twice as expensive as Gemini 2.5 Flash-Lite and DeepSeek V4 Flash after the cut.

The cut makes the long-context model easier to use by default, but it does not make every request cheap enough to ignore. A weekend project does not need a committee. It needs a default, a middle tier, and a narrow latency tier.

Reset the stack around the work you actually do. Use the cheap tier for work that can wait, the middle tier for work users touch, and the fast tier for the few moments where waiting is visible.

Reset the stack in the order you will actually use it

Start by listing the routes that call a model. Mark each one as long-context, normal, or latency-sensitive. The audit works when you can see the route, the model, the reason, and whether a human is waiting.

Most weekend projects have three patterns: an agent that reads a lot, a normal chat or form, and a fast interactive feature. If you cannot name them, the stack is doing more work than you thought, and the bill will keep hiding in the extras.

If a route is hard to classify, put it in the middle tier until you see the usage. Do not try to optimize every prompt. Try to stop paying for speed you do not use.

Move agent and batch jobs to Luna. Put overnight jobs, summarization, extraction, and multi-step agents there. You have moved the work when the long-context jobs no longer depend on the middle tier.

The common mistake is to move every request to Luna because it is cheaper. That works until a user-facing feature feels slow or the context window is wasted. A cheap default still needs a boundary.

Keep Terra for standard production. Use it for normal user-facing requests, support drafts, and anything that should feel reliable without a strict speed requirement. The main product path should stay on the middle tier.

Do not treat Terra as a fallback for everything. A fallback that catches every edge case becomes the real model, and the budget follows it. If a request belongs to Terra, say so in the audit and leave it there.

If Terra feels too slow, the fix is usually prompt trimming, caching, or a smaller response. Jumping to the fast tier because one request felt sluggish is how a weekend budget becomes a monthly argument.

Reserve Sol for sub-200ms latency. Use it only where a human is waiting and the answer must arrive before attention slips. The finish line is Sol appearing in a few routes, not across the app.

If a feature does not need that speed, moving it to Sol is paying for performance you will not use. Keep the latency tier narrow, and note the reason in your audit table. If you cannot name the waiting user, the route probably does not need Sol.

Latency is a user experience problem. If the answer is correct but arrives a beat late, the model may be fine and the interface may need a spinner, a queue, or a shorter answer.

Make the budget a habit, not a surprise

Keep a small sheet with monthly tokens, model, price, and total. After a price change, update the numbers and rerun the estimate. The budget works when you can name the most expensive route before the invoice does.

Review the sheet after a release, after a traffic spike, and after any price change. A near-zero budget means the bill should not need a meeting, even if it is not literally zero. If a line item jumps, you should be able to point to the route that caused it.

Most side projects stay side projects, and that is fine. A near-zero model stack lets the project live on weekends without becoming a second job or a company.

When the next price cut lands, do not rebuild from scratch. Keep the near-zero model stack rule:

• Default to Luna for agents and batch work.
• Use Terra for standard production.
• Reserve Sol for sub-200ms latency.
• Recompute the budget after every price change.