The reprice rule for the Sept 22 AI price cut

Your agent project is still sending the same context, and Anthropic released Claude Opus 5.5 on Tuesday, September 22. If you run an agent-style side project on evenings and weekends, the new list prices can feel like a free discount. The discount becomes real only after you reprice your own weekly token volume. A cheaper sticker price is not the same as a cheaper weekend for you.

Re-sent context often drives the surprising cost. Stanford Digital Economy Lab research found that re-sent context makes up roughly 62% of total agent inference costs. The 62% figure explains why a small input-price cut can look bigger than the bill change.

Most side projects stay side projects, and that is fine. They need a number close enough to stop a bad switch, not a finance exercise. A weekend reprice rule gives you that number. It is a quick check, not a budget rebuild or a spreadsheet project.

The cache-read line usually decides the bill

Cached context changes the bill. When an agent reads previously cached context, the next read can be much cheaper. A cached read is the model continuing from where it already is, instead of rebuilding the prompt from scratch. Check the cache-read line first.

Four steps turn a price cut into a decision

The reprice rule runs in a short sitting. You need your usage notes and the new list prices. If your project is small, the whole exercise can take less time than a model switch.

• Record your typical weekly input, output, and cache-read tokens. A small note with rows for input, output, and cache reads is enough; a normal week beats a perfect average.
• Multiply each number by the new list prices. For Claude Opus 5.5, the input rate is $4 per million tokens, and the output rate is $20 per million tokens, down from $5 and $25. On September 22, 2026, OpenAI set GPT-6 Sol at $2.00 per million input tokens and $10.00 per million output tokens, and GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens. Anthropic cache reads are 20 cents per million tokens after the Opus 5.5 release, compared with 50 cents for Opus 5.
• Compare the old and new totals. Put the bill before the cut and the bill after it side by side. If the new total is only a little lower, the cut is not doing much for your project. For a monthly view, scale the weekly total to the month you care about.
• Choose the next move. Pick keep, switch, or tighten context. If cache reads are the largest cost, tighten caching or context before switching models. If output tokens dominate, look at reasoning effort. Artificial Analysis's test battery measured GPT-6 Sol at $0.13 per task on low reasoning effort and $1.06 per task on max reasoning effort.

Keep the same week for both totals, so the comparison reflects usage rather than a busy or quiet period. Headline rates can mislead when the cache line is skipped. A cheaper input price can be swallowed by cache-read cost. Savings can hide or disappear in that line.

A cheaper model is not always the cheaper weekend

GPT-6 Sol's list price is 50% lower than GPT-5.6 Sol's in both input and output, while GPT-6 Luna's list price is 50% lower on input and 58.3% lower on output. That sounds like a clear win, but a model switch costs time: new prompts, new tests, and a weekend spent debugging instead of building.

Use the reprice rule before you spend that time. When the old and new totals are close, keep the model you already trust. A clearly cheaper new model with a small cache-read line is worth switching. Large cache reads point to fixing caching or context first. Your side project does not need the cheapest model. The rule turns a price cut into a decision you can make before the weekend is gone.