Nov AI Coding Quota-Saving Tips

Don't let context grow too long: auto-compaction is triggered earlier at 400K, and starting a new session for each task reduces the historical tokens carried with every request.

Default to medium reasoning effort instead of always maxing it out. For the same task, switching models can change costs by more than 10x, so choose the model based on the task first.

Don't have the model reason repeatedly over steps that a program can handle: batch tool calls and cache steps that are already settled, which can noticeably reduce requests and tokens.

Slim down the material you feed the model first: convert documents to Markdown, and for desktop actions read the accessibility tree instead of screenshots, which saves a lot of tokens.

Simple agent tasks can switch to open-weight models, which are more cost- and latency-efficient; AMD's free interface can also be connected to Claude Code.

Starting December, ChatGPT and Codex share a quota pool, and only Instant chat is free. For Grok's new and old subscriptions, also work out the costs before switching.

Test cheap subscriptions before buying: OpenCode Go's share of reasoning usage is high, and the third-party model quota in Gemini Ultra is also less than the advertised multiple.

Message threading only makes requests smaller and responses faster; tokens are still billed as usual, so saving money still requires prompt caching.