How to Save Your Codex Usage Limit: Fast Mode and Models

👁 117 reads

In Codex settings, under Usage and billing, there’s a sentence worth reading twice: “Shared limit for Codex, Work mode, workspace agents, and ChatGPT for Excel. Chat conversations don’t count toward it.” Below it is a weekly usage bar. I have a Pro plan, and on the first day the bar showed 100%. A few hours later, while looking through a configuration file, I found a setting that had quietly been using up that limit two and a half times faster for weeks.

That setting was Fast mode. Here’s what it costs, how to choose a model and reasoning level for each task, and which settings I’ve kept. All figures come from OpenAI documentation or from my tests on November 3, 2026.

Fast mode uses 2.5 times more of your limit

Fast mode makes the model respond more quickly. OpenAI says it works 1.5 times faster. The same document says that Fast mode uses 2.5 times more of your subscription limit than standard mode, and twice as many purchased credits. It works with GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Astra, GPT-6 Sol and GPT-6 Luna, where available.

There’s also Ultrafast, available only for GPT-6 Astra and GPT-6.1 Sol. It uses eight times more of your subscription limit than standard mode, and six times as many purchased credits. An hour of Ultrafast therefore uses as much of your limit as eight hours of standard mode.

Fast mode had been enabled by default in two places in my setup. In ~/.codex/config.toml, I found service_tier = "priority"; in the desktop app configuration, there was default-service-tier = "priority". In Codex, Priority is the same setting as Fast, under its older name. I removed both lines. When I really need speed, I enter /fast on in Codex, then /fast off when I’m done. I check the status with /fast status.

If you want Fast mode on all the time, the documentation gives these two lines:

service_tier = "fast"

[features]
fast_mode = true

Sol, Astra or Luna?

Codex offers several names in its model picker, but the descriptions in the app are brief. GPT-6.1 Sol is “the latest workhorse model for coding and everyday tasks.” GPT-6 Astra is “frontier intelligence for the most demanding tasks.” GPT-6 Luna is “a fast, affordable model for simpler tasks.” There’s also GPT-6.1 Terra, which I tested separately in GPT-6.1 Terra in Codex, along with older models from the 6 and 5.6 series.

Astra has stronger reasoning, but it’s from an earlier series than Sol 6.1. That’s why I haven’t made it my default. I use it for tasks where a mistake would cost me more than the extra time: planning a new app’s architecture, tackling a difficult bug that Sol has already misdiagnosed, or doing substantial research. For renaming files, writing a simple script or reading a log, Luna at a low reasoning level is enough.

Reasoning level is another control

Each model has a reasoning effort setting: low, medium, high, xhigh and max, and most also have ultra. Luna tops out at max. OpenAI’s Codex prompting guide recommends medium for everyday use, and high or xhigh for the hardest and longest tasks. A higher setting means more reasoning before the response, and therefore more usage.

I set high as my default. I mainly use Codex for end-to-end projects, where a cheap mistake can cost more than taking extra time to think. I switch to max or ultra manually for individual tasks. My configuration file looks like this:

model = "gpt-6.1-sol"
model_reasoning_effort = "high"
plan_mode_reasoning_effort = "xhigh"

The third line applies to planning mode. A plan is created once and shapes all the work that follows, so I think it’s worth giving the model more time at that stage. Execution then uses the regular high setting.

Even a short question uses tokens

While configuring Codex, I asked it quick questions from the terminal, such as “What’s your rule about dashes?” This is a translation of the original Polish test prompt; the test itself was conducted in Polish. The answer was a few lines long, but the counter showed between 19,000 and 24,500 tokens each time. Even in an empty conversation, Codex sends the model its instructions, my AGENTS.md file, the list of available skills and descriptions of its tools. That’s why ten one-sentence questions use more than one question with ten bullet points.

That led me to adopt three habits:

• I give Codex the whole task in one message, including context and the expected result, rather than working through it step by step.
• I keep AGENTS.md short. I moved the details into skills, which load only when needed.
• I don’t let long threads go on indefinitely. Starting a new conversation with a good AGENTS.md file is cheaper than sending the hundredth message in an old one.

Codex memory and your usage limit

I also enabled Codex memory (memories = true in the [features] section). Codex extracts notes from conversations and uses them later, much as Claude Code does. Creating those notes also uses your limit, but there’s a safeguard: with the default settings, Codex stops generating them when less than 25% of the limit remains. You can change the threshold with memories.min_rate_limit_remaining_percent.

When your weekly limit runs out, Codex can keep working with purchased credits. The same settings tab shows your balance and a toggle for automatic top-ups. OpenAI promises a discount of up to 40% when you buy credits. I haven’t enabled automatic top-ups. I’d rather have the limit running out serve as a signal than get an unexpected bill.

I covered limits on the cheaper plan in Codex limits on ChatGPT Plus. In my guide How to move Claude Code to Codex, I show how to bring your Claude Code rules and skills into Codex, so you don’t have to explain everything again in every message. But the biggest difference for me came from removing one line: service_tier = "priority". The same weekly limit now covers two and a half times as much work as it did with a setting I didn’t even remember enabling.

Sources

• OpenAI, Speed (Fast and Ultrafast): https://learn.chatgpt.com/docs/agent-configuration/speed
• OpenAI, Configuration Reference: https://learn.chatgpt.com/docs/config-file/config-reference
• OpenAI Cookbook, Codex Prompting Guide: https://developers.openai.com/cookbook/examples/gpt-5/codex_prompting_guide
• Codex Desktop settings, Usage and billing tab, as of November 3, 2026
• My Codex CLI 0.162.1 tests, November 3, 2026

// read moreClaude Fable 5.2 in Claude Code: Advisor, Fast Mode and AgentsNewsSync Claude Code and Codex with rulesync, Ruler and MoreAI ToolsCodex CLI on Windows: Step-by-Step Install GuideGuides