Codex Fast and the Price of Intelligence on a Clock

OpenAI recently added a switch to Codex called Fast. On GPT-5.4, Fast mode increases speed by 1.5x while consuming credits at 2x the rate, and OpenAI describes the feature as “Increase speed without sacrificing intelligence”. Most reactions stop at the surface arithmetic: pay double, wait less. That arithmetic leaves out the unusual part of the product. We almost never get a clean market price for the same reasoning delivered sooner. Historically, intelligence came bundled with time, access, reputation, and human availability. Codex Fast pulls time out of that bundle and sells it directly.

My first instinct is still economic. If a task takes 10 minutes at normal speed, Fast cuts it to about 6.7 minutes. The user gains about 3.3 minutes and pays as if they ran the task twice. That is a weak trade. But simple cost-per-minute logic still feels too shallow here. Fast belongs in a stranger category: what OpenAI is selling is accelerated cognition.

This Substack is reader-supported. To receive new posts and support my work, consider becoming a free or paid subscriber.

That category has older analogs, even if the product feels new. Psychometric research on timed testing has long treated performance as a combination of ability and speed, with a speed-ability trade-off shaping the measured result. Give people more time and many more will eventually solve the same pattern. Tighten the clock and the ranking changes. A timer does not merely change convenience. A timer changes what gets measured. That point helps explain why timed IQ-style tests have always been a little slippery. They capture raw reasoning, but they also capture how quickly someone can turn reasoning into an answer.

Markets have traditionally handled the same problem from the business side. Expertise and urgency were sold separately. The good lawyer had one rate. The same lawyer on overnight turnaround had another. The consultant with a normal cadence was one product. The consultant who moved your work to the top of the pile was a different product. Queueing theory has modeled purchased priority for decades, and more recent work on priority pricing with delay-sensitive customers makes the same point in cleaner form. Some buyers care much more about delay than others, so a paid priority lane emerges naturally.

Finance gives the harsher analogy. In high-frequency trading, latency can be valued directly. The faster signal is not necessarily wiser. The value comes from actionability. The window is open, then it closes. Earlier intelligence can command a premium when the surrounding opportunity decays quickly enough. That idea is particularly useful when thinking about faster intelligence: an answer arriving three minutes earlier is sometimes just nicer but in other cases, those three minutes change what can be done next.

AI adds one more twist, and I think this is the key one. Human experts sell lumpy blocks of time: an hour, a day, a retainer, a project. AI sells at much lower granularity — AI sells turns and that changes the economics. The gain from an earlier answer may sit less in the answer itself and more in the extra move the answer unlocks. One more compile-run-debug loop before the meeting. One more pass before context slips. One more chance to catch the issue before a deploy window closes. Horvitz’s work on continual computation and expected value of computation gives a strong conceptual frame for this. Computation has value, but delay also has a cost. Classical anytime algorithms usually ask when extra thinking time is worth spending to improve a result. Fast’s mode flips the problem around. The quality is held roughly constant and the question becomes how much value comes from compressing the same deliberation into less wall-clock time.

Once I look at Fast through that lens, the price of accelerated intelligence breaks into four separate premia.

First, there is queue value: how painful is the wait itself?

Second, there is decay value: how quickly does the answer lose usefulness as time passes?

Third, there is option value: does the earlier answer buy another iteration, another test, another draft, another decision before the window closes?

Fourth, there is blocking value: what scarce asset is sitting idle while Codex is thinking? Sometimes that asset is my own attention. Sometimes it is another engineer. Sometimes it is a live incident, a customer conversation, or a narrow deploy window.

Plain cost-benefit analysis catches only the first layer. Most of the actual value sits downstream.

That is why the phrase “mispriced intelligence” is close, but not quite precise enough. The thing being priced here is intelligence under time pressure. That is a more faithful description of the commodity. Psychometrics says the clock changes the task. Queueing theory says delay-sensitive buyers will pay for priority. Metareasoning says computation time and decision quality belong in the same objective function. Put those three traditions together and a cleaner picture appears. Fast is a rush fee on a general-purpose reasoning engine, plus whatever option value the saved minutes create inside the user’s workflow.

Now go back to OpenAI’s numbers. Fast mode gives 1.5x speed at 2x credit consumption. Throughput rises by 50 percent. Human waiting time on a fixed task falls by about one third. Those are the same facts, framed in two very different ways. “50 percent faster” flatters the product. “About 3.3 minutes saved on a ten-minute loop” tells the human story. Every time someone leaves Fast on, they are making a concrete claim: these saved minutes are worth another full standard run.

That claim can make perfect sense near a real threshold. Live debugging is one. A blocked teammate is another. A short work session before a meeting, a production issue, a fragile stretch of concentration, or any workflow where one extra loop changes the odds, those are all strong candidates. Outside those conditions, the premium gets soft quickly. Repo reading, broad exploration, refactors, overnight agents, background work, and long-running tasks usually have a long enough utility half-life that the earlier answer does not change much.

So my conclusion is simple: Fast belongs in the toolbox, not in the default config. Use it when the saved minutes protect something scarcer than credits, or when unused weekly credits push the true marginal cost close to zero. Leave it off the rest of the time. OpenAI has found a clean way to price urgency. The product is clever. The default economics still look weak.

This Substack is reader-supported. To receive new posts and support my work, consider becoming a free or paid subscriber.