Why and how is it even logical that o1 is 10x+ more expensive than o3, when o3 is supposed to be better?

Community
, , , ,
1

Hi everyone,
I’m genuinely confused by OpenAI’s API pricing here and I’m hoping someone (ideally OpenAI) can explain the rationale.

From the pricing tables:

  • o1: $15 / 1M input, $60 / 1M output

  • o3: $2 / 1M input, $8 / 1M output

So o1 is ~7.5x more expensive than o3 on both input and output — and in practice it feels like 10x+ depending on usage patterns.

But at the same time, o3 is positioned as the newer / stronger reasoning model, so:

Why would the older model cost dramatically more than the newer model that’s supposed to perform better?
How is that pricing logical?

2

You’re assuming something that has never really been true for OpenAI’s API pricing:

“Newer + stronger model ⇒ must be more expensive.”

Historically that’s just wrong.

Take gpt-4-32k as a concrete example. Its official pricing was:

  • $60 / 1M input tokens
  • $120 / 1M output tokens

That’s dramatically more expensive than both o1 ($15 / $60) and o3 ($2 / $8). And yet gpt-4-32k is strictly weaker than the current frontier models.

So if you apply your logic consistently, the old, weaker gpt-4-32k being 3–10x the price of today’s stronger models would make even less sense. But that’s exactly how pricing has looked in the past: newer models tend to become both better and cheaper over time.

The reason is simple:

  • Pricing is about product tier and compute cost, not “publishing order.”
  • o1 is a specialized, slower, more compute-heavy deliberate reasoning product.
  • o3 is optimized to be the default workhorse: fast, cheap, good-enough reasoning for most use cases.

If you don’t need the deliberate behavior of o1, just use o3 and enjoy the fact that in 2025 you’re paying a fraction of what people paid for gpt-4-32k, for far better performance. The “why isn’t the newer model more expensive?” complaint ignores the entire history of API pricing.

3

If you really think “newer + stronger ⇒ must be more expensive” then you’re going to have a hard time explaining Anthropic’s pricing too.

Compare just these two:

  • Claude 3 Opus (older flagship “big brain” model):

    • $15 / 1M input tokens
    • $75 / 1M output tokens

  • Claude 4.5 Haiku (newer “small” model with near-frontier code + reasoning):

    • $1 / 1M input tokens
    • $5 / 1M output tokens

So Anthropic’s newer, extremely capable Haiku 4.5 is literally ~15x cheaper than their older Opus 3, and it’s good enough to replace Opus in a ton of real workloads.

Exactly the same pattern you see with OpenAI:

  • high-end “deliberate / frontier” models are priced as a premium tier,
  • smaller, highly optimized newer models are positioned as fast/cheap defaults.

In other words, across vendors the actual pattern is:
“model role + compute profile ⇒ price,”
not
“release date or version number ⇒ price.”

If you want to argue the pricing is bad, fine — but the “why is a newer model cheaper than an older one” line just ignores how literally everyone in the industry prices their model families.

1 Like