Budget your weekend agent to $4/$20, not the promo date
Write the monthly agent bill to a sustainable rate, then set a cap that survives the promo.
Your weekend agent can feel cheap while a promo is running, then become a surprise on the next billing cycle. Calculate the monthly bill at the rate you can keep paying, then set a cap that survives the promo.
The sustainable rate is the planning rate
OpenAI announced a developer price cut for its frontier GPT-5.6 Sol model on August 21, 2026, with a reduction of more than 20 percent lasting for the next three months. The discount reaches OpenAI's API and eligible credit plans for ChatGPT Work and Codex, while Pro, Plus, and Business subscription prices remain unchanged.
If your agent runs through Codex, check whether your plan is eligible before you treat the lower number as permanent. For the worksheet, GPT-5.6 Sol is the model to plan around, because it is the one your agent will use on ordinary calls.
Put the planning rate on the worksheet first. Put the promotional rate beside it, with the end date if your plan shows one. That keeps the planning number from being overwritten by a discount that expires while your agent is still running.
Keep the worksheet small enough to update after each weekend. A few lines are enough: input estimate, output estimate, rate, monthly total, and cap. If the total moves, change the agent before you change the budget.
Token math belongs before the promo expires
- Estimate your monthly input and output tokens. Count prompts, retrieved context, tool results, conversation history, and retries. Done looks like a number for what the agent sends and a number for what it returns.
- Multiply by the sustainable rate. Use $4 per million input tokens and $20 per million output tokens for standard short-context calls to GPT-5.6 Sol. You are done when the monthly number can stand without a discount.
- Compare the result against the earlier $5 and $30 levels and Anthropic's Claude Fable 5 rate of $10 per million input tokens and $50 per million output tokens. The step is complete when you can name the gap before the promo ends.
- Set a monthly cap at the rate you can sustain. Put the cap where you track other weekend project costs, and make it visible in your billing dashboard.
Treating the promotional rate as the planning rate is the mistake. Discounts can make a heavy agent look affordable. Pay the bill that remains after it ends.
When the sustainable number is too high, do not wait for the next promo. Reduce context length, remove unnecessary tool calls, or move routine steps to a cheaper model. A smaller agent is easier to keep alive than a bigger one that needs a discount to justify itself.
Do not let the agent grow until the bill is stable. A new tool call, a longer memory, or another user can all raise the monthly total. Test the change in a low-traffic weekend before you make it the default.
The comparison sets your limit
Anthropic's Claude Opus 5 adds a middle rung at $5 per million input tokens and $25 per million output tokens. That cost is useful as a limit test: if your agent cannot survive near it, it needs smaller prompts, fewer turns, or a cheaper model.
OpenAI also cut the price of the mid-tier GPT-5.6 Terra model by 20% and the lower-cost Luna model by 80%. Cheaper models fit simple tasks, because they keep the frontier model for the steps that need it.
Finish with the agent monthly cost you can pay without stress. Keep the cap when the number fits your weekend project cost. Shrink the agent before you scale it if it does not.
Once the cap is set, make it the limit. A weekend agent does not need to be impressive every night; it needs to be cheap enough to keep running while you improve it. Most side projects stay side projects, and that is fine.