Is Codex Fast Mode Worth It? 2.5x the Credits for 1.5x the Speed (I Tested It Live)

Codex Fast Mode trades credits for latency: roughly 1.5x faster inference for 2-2.5x the usage burn. The marketing question is whether it is fast. The real question is whether you can feel it, and whether the speed changes your output. I ran it live on a real build for 36 minutes to find out.

TL;DR: Fast Mode is real, and the trade is bad for most work. You get ~1.5x speed for ~2-2.5x credit burn, which is worse than linear. It costs no extra money, it just eats your existing weekly limit faster. In a live build test the speedup was invisible because the bottleneck was me, not the model. And speed did not fix quality: three faster iterations were still three broken iterations.

What Fast Mode Actually Is

Fast Mode is a serving tier, not a model. Same weights, same reasoning, lower latency — you are paying for priority on faster hardware.

The published numbers:

Do the division. At 1.5x speed for 2.5x burn, you are paying about 1.7x more credits per unit of work completed. That is the whole decision in one number.

How to Turn It On

Three ways, depending on how permanent you want it:

# per session, in the Codex CLI
/fast on
/fast status
/fast off

# per project, in config.toml
[features]
fast_mode = true

# at launch
codex --enable fast_mode

The /fast status check matters more than it looks. Fast Mode toggles silently, and the failure mode is leaving it on for a week of routine work and wondering why your limit evaporated.

The Live Test

I gave Codex one prompt with Fast Mode on: build a city game where you can borrow any resident’s life for 60 seconds. No cuts in the recording, so the real latency is on the clock.

What happened:

• 8 minutes to a playable proof of concept. Genuinely impressive.
• Try 1: working POC, broken game. The mechanic did not do what the prompt described.
• Try 2: fed it a reference repo. Still broken.
• Try 3: still broken.
• Opened a second terminal and let Claude take the same prompt with the same method. Different result in about seven minutes.

Two findings came out of that, and only one of them is about speed.

Finding 1: You cannot feel 1.5x

This is the part the multiplier hides. In an agentic coding loop, model latency is not the bottleneck — you are. The wall clock goes to reading the diff, running the thing, watching it fail, and deciding what to say next. Shaving 30% off the generation step disappears into that.

Fast Mode helps when you are genuinely blocked on tokens arriving: short interactive turns, quick edits, live debugging where you are staring at the cursor. It does nothing for long autonomous runs where you tab away anyway — and long autonomous runs are exactly where the burn multiplier hurts most.

The worst possible use of Fast Mode is a 40-minute unattended agent run. You pay 2.5x for speed you are not present to observe.

Finding 2: Speed is not quality

Three Fast Mode iterations produced three broken games. Fast Mode does not change reasoning, so it cannot change the outcome — it only changes how quickly you arrive at it. If the model is going to get your task wrong, Fast Mode gets it wrong sooner and charges you 2.5x for the privilege.

That inverts the intuition people bring to it. Fast Mode is not “more power for hard problems.” It is less economical on hard problems, because hard problems are exactly the ones that need many iterations.

Does It Cost Extra Money?

No, and this is the most common confusion in the threads about it.

Fast Mode has no separate charge. Codex and ChatGPT draw from the same usage pool your plan already includes. Turning Fast Mode on does not create a new line item — it drains the pool you already paid for, 2 to 2.5 times faster.

So the cost is real but indirect: you hit your weekly limit sooner and stop working. Several users on a $200/month plan have reported exactly that — usage evaporating faster than expected with Fast Mode left on. The bill never changes. The productive hours do.

When Fast Mode Is Actually Worth It

Narrow, but real:

1. Tight interactive iteration. Short prompts, you are watching the output, and the latency is the thing stopping you. This is the case OpenAI designed it for and it works.
2. A demo or a deadline. Wall clock matters more than credits for the next hour. Turn it on, then turn it off.
3. You are nowhere near your usage cap. If you consistently end the week at 30% of your limit, the multiplier costs you nothing real. Spend it.

When it is not worth it:

1. Long autonomous agent runs. You are not watching. You are paying 2.5x for latency you never perceive.
2. Hard problems needing many iterations. The multiplier compounds across every failed attempt.
3. Anywhere near your cap. Hitting the limit costs you more time than Fast Mode ever saved.

The Rule I Use Now

Fast Mode on for conversation. Fast Mode off for work.

If I am sitting there watching tokens land and my next move depends on this answer, on. If I am handing the agent a task and walking away, off. /fast status before any long run.

That single habit is worth more than the speedup ever was.

Watch the Full Test

The whole 36-minute run is unedited on YouTube, including the three failures and the Claude comparison at the end:

Codex Fast Mode: 2.5x the Credits, 1.5x the Speed. Worth It?

If your team is burning through agent credits faster than expected and you want a second pair of eyes on where it is going, that is exactly what we do — book a 1:1 or talk to us about a usage audit.