Switching APIs from 4.1 mini to 5.1 mini - a lot more tokens generated

API
, ,
1

We use 4.1 mini for our APIs and have many custom instructions to avoid answer specific questions, etc. When experimented with 5.1 mini the same instructions have generated a lot more tokens. We really cannot drop the instructions, they are important. Does anyone has advise on how to have comparable number of tokens used with 5.1 and preserve lengthy instructions?

2

You can set a max_completion_tokens to force it, though you might want to experiment with your prompt to make the model more succinct. Literally ask it to summarise, be succinct etc.

1 Like
3
  1. There is no 5.1 mini model. You’ve got gpt-5-mini-2025-08-07, and a codex-mini model that is not appropriate for general use.
  2. GPT-5 AI models are reasoning models. They produce internal language that you pay for but don’t see, thinking about the task.
  3. The control for this amount of thinking is the API parameter reasoning.effort, where your lower-expense option is to send “low”, with the side-effect of lower-quality results on a model reliant on this reasoning.

GPT-5 family summary:

  • comparable number of tokens: no
  • comparable cost and latency: no
4

Yes, but on Completions, reasoning is default OFF on gpt-5.1 , so you have to be explicitly setting it (which wasn’t obvious from the OP)

5

“mini” does not appear in the name “gpt-5.1”.

Any concern seems to be about the cost, and not the output length, as reducing the length of input instructions was a budgeting option that was considered and ruled out.

Thus, using the full-size model also must be ruled out, “5.1” merely a mis-step in communication that does not meet the goal, as its output is already 6.25x the cost.

gpt-5-mini at “minimal” reasoning is a dot that appears almost identical in performance to 4.1 in 2d benchmark plots.

6

and 5.1 doesn’t appear in 5-mini either :man_shrugging:

7

And “system” as a role of a message doesn’t appear in any context that you can send to a gpt-5 model, if instruction-following was the goal. It gets its own text from OpenAI, often counter to your application.

Chet the cowboy, on whether it’s easier to just brand all the cattle he don’t own:

gpt-4.1-mini

Haha, you got that right! Out here, sometimes it feels like the whole county’s cattle end up wanderin’ through my pastures. Keeps me on my toes, that’s for sure. But I reckon that’s part of the fun — ropin’ and ridin’ through the open land, makin’ sure everyone’s stock is where it oughta be. Helps me know every critter by name — or at least by their brand! What about you, ever spent much time out on a ranch?

gpt-5-mini, minimal - you get disclaimers

Ha! I get that a lot. Folks around here joke I oughta start a “lost and found” for cattle. Truth is, I only brand what I own — it’s about respect and keeping the peace out on the range. Branding’s a legal thing too; you don’t want to be the one folks come looking at courthouse time.

1 Like