Keep Argon for Hard Parts: A Dec 1 Cost Plan
Treat the Dec 1 price change as a weekend checkpoint: keep the paid model for hard tasks, route routine calls elsewhere, and batch overnight runs.
Dec 1 is a scope checkpoint
Your small project is already calling Gemini 4 Argon, and beginning 1 December 2026 the listed API price doubles to $4 per million input tokens and $20 per million output tokens. The introductory rate runs through 30 November 2026, at $2 per million input tokens and $10 per million output tokens. Google made Gemini 4 Argon available to developers and businesses on 13 October 2026. That gives you a short window to decide which calls deserve the paid model.
Most side projects stay side projects, and that is fine. The model is a tool, not a business plan. Ask whether the paid model is doing work that only it can do well. The price change is a filter: move routine calls, protect hard calls, and schedule overnight jobs. The token costs are the part you can control. The sample workload in the checklist shows the cost at a fixed volume.
Repeated prompts can also change the math. Google lists implicit cached input for standard Gemini 4 Argon at $0.10 per million tokens until 30 November 2026 and $0.20 afterward. If your app sends the same system prompt, the same few examples, or the same tool description in the same order, that repeated part may already be cheaper than it looks.
Split the calls into three lanes
Start with the calls you already know. The routine lane is the first place to look. The hard lane is the one to protect. Hard calls are the ones where a wrong answer costs you time: a long document, a tricky refactor, a customer-facing draft, or a tool call that needs careful formatting. Routine calls are the kind you would not mind running on a cheaper model: short summaries, label checks, simple rewrites, or status updates. Overnight calls are the ones that do not need to answer before you go to bed: bulk cleanup, report generation, data reshaping, or long test suites.
The audit does not need to be perfect. It needs to be honest. Pick a recent day, read the prompts, and mark each call with a lane. If a call is doing several jobs, split it. If a call is mostly boilerplate, move it. If a call is the reason the project works, keep it on the stronger model.
Routing makes the plan real. Give each lane its own prompt, its own log, and its own failure mode. The hard lane can be slower and more expensive. Keep the routine lane fast and boring. Schedule the overnight lane, retry it, and check it in the morning.
Run the swap as a triage sheet
Use this three-lane swap checklist before the deadline, not after. Each item should take a minute. Make the swap visible before the invoice does. If an item feels slow, shrink the sample and run it again.
- Label a recent batch of calls as hard, routine, or overnight. Why: the lanes make the swap visible.
- Check that routine calls can run on a free-tier model, because Argon is excluded from the Gemini API free tier.
- Keep hard calls on the paid model when the task needs long context, nuanced judgment, or careful formatting.
- Batch long overnight runs into scheduled jobs instead of scattered daytime calls.
- Estimate cost with the standard-mode sample workload: 10 million non-cached input tokens plus 1 million output tokens costs $30 during the introductory period and $60 after the price change.
- Set routing so hard, routine, and overnight lanes have separate prompts and logs.
Set the routing before the weekend, then check the logs to confirm the lanes are holding. If a routine call still lands on the paid model, move it. If a hard call feels thin, keep it on the stronger model.