GPT-6 Astra : ne sois plus jamais à court de token Codex ! [24 HACKS]

Summary

The video presents ways to make a GPT-6 Astra quota last longer in Codex. Its central advice is to match the model’s effort and tools to the task, reduce unnecessary context and output, and avoid spending quota on work that can be handled more cheaply.

Settings and workflow

1. Use the lowest suitable effort level. The presenter cites independent tests suggesting that higher effort can cost substantially more for only a small improvement in measured intelligence. Start low and increase it only if needed.
2. Avoid Max effort by default. The video claims that moving from Extra High to Max adds significant cost for a very small score increase.
3. Reserve Ultra for large tasks. In Codex, Ultra is described as providing maximum reasoning and delegation to sub-agents, which can add to the cost.
4. Consider a scheduled early-morning message. The presenter suggests sending a brief message at 6 a.m. to start a five-hour usage window earlier, so a later reset may arrive sooner. This depends on how the particular subscription’s limit window works.
5. Choose the model before starting a session. Switching models mid-session may mean the new model cannot use the existing conversation summary and must reread more context. Use a less expensive model when it is sufficient.
6. Turn off Fast mode unless speed is essential. The video says it uses quota more quickly for the same work.
7. Use Steer follow-up behavior. Sending corrections while the agent is working can redirect its current task rather than leaving the correction queued, potentially avoiding wasted work.
8. Ask for concise responses. The presenter recommends the “Caveman” skill and another brevity-focused plugin (called “I have ADHD” in the subtitles) to reduce unnecessary output.
9. Ask the agent to audit quota use. It may identify excessive effort settings, unused plugins, or secondary agents that repeatedly receive the full conversation.

Project setup and tools

1. Provide exact file paths when you know which files the agent needs, rather than having it search the project.
2. Create workspace maps or “file routers.” Small Markdown indexes can point to relevant files for different work areas. Reference these from AGENTS.md so the agent can locate information without loading or searching broadly.
3. Disable unused connectors and MCP servers. The video says enabled integrations can add context to new conversations even when they are not being used.
4. Clean up instructions, skills, plugins, and MCP servers monthly. Remove or archive duplicated, outdated, or overly long material. The presenter suggests scheduling the review near the end of a usage period.
5. Use an external image provider for image generation. The presenter says image generation within Codex can consume quota several times faster than a normal prompt. They describe using Fal.ai and an API key to generate media outside the Codex quota.
6. Minimize generated code. A tool name transcribed as “Ponytail” is presented as a way to encourage minimal implementations, which may reduce token use and make code easier to maintain.
7. Install RTK (Rust Token Killer). It filters command output before the agent reads it. The video cites claimed reductions of 60–90% for some common developer commands.
8. Use QMD for semantic file search. Unlike exact-word searching, it is described as helping find relevant files by meaning, especially in large workspaces.
9. Set an explicit token or quota budget in prompts. For example, ask the agent to stay below a specified percentage and seek confirmation if it might exceed the limit.

The subtitles cover tips through number 19 and refer to additional bonus tips in a free PDF guide, but do not describe those remaining tips. The presenter says the PDF includes resources and can be given to an agent to assess which suggestions fit a user’s setup.

Main speaker and sources

• Main speaker: The video’s presenter, identified by the channel metadata as Lou IA.
• Sources and tools mentioned: Artificial Analysis AI for model-effort comparisons; Codex settings and usage controls; GitHub-hosted skills and tools including RTK and QMD; Fal.ai for external media generation; and a free PDF guide from the presenter.