Introducing GPT-5.4 mini and nano — our most capable small models yet
It seems just about every bit of documentation is wrong.
- The models comparison page has rounded values for token pricing that are incorrect:
gpt-5.4-mini vs gpt-5-mini
Bad rounding: $0.075 → 0.08 $0.025 → 0.03
- The “patches” multiplier in documentation is all screwed up:
What it should look like, given the currently realized costs:
Model Multiplier gpt-5.4-mini-2026-03-171.2xgpt-5.4-nano-2026-03-171.2xgpt-5.4-2026-03-051.2xgpt-5.3-codex1.2xgpt-5.2-2025-12-111.2xgpt-5-mini-2025-08-071.2xgpt-5-nano-2025-08-071.5xo4-mini-2025-04-161.72xgpt-4.1-mini-2025-04-141.62xgpt-4.1-nano-2025-04-142.46xgpt-5.3-codex1.2xcodex-mini-latest1.72x
The maximum vision input size of the new GPT-5.4-mini and nano models is wrong. You can send 1600x1600 and get billed for 50x50 patches = 2500 (haven’t tested “original”).
Chat Completions with gpt-5.4-mini and nano is resizing wrong or billing wrong (cheaper). Here is sending that 1600x1600px for 2500 patches/tokens:
| model | vision | mult | ChatC | Ccalculated | Responses | Rcalculated |
|---|---|---|---|---|---|---|
| gpt-5.4 | patch | 1.2 | 2813 | 2338 | 3008 | 2500 |
| gpt-5.4-mini | patch | 1.2 | 2813 | 2338 | 3008 | 2500 |
| gpt-5.4-nano | patch | 1.2 | 2813 | 2338 | 3008 | 2500 |
| gpt-5-mini | patch | 1.2 | 1834 | 1522 | 1833 | 1521 |
| gpt-5-nano | patch | 1.5 | 2290 | 1522 | 2289 | 1521 |
The amount of input “usage” received back per endpoint is the columns “ChatC” and “Responses”. Image consumption then by input difference due to inclusion of the image, and then reversing the apparent multiplier.
If it were downsized, like the original mini and nano:
1248 × 1248 px (39 × 39 patches) = 1521 tokens.
This is not the only model disparity in billing between API endpoints. Here’s sending the image to others to where it should be downsized. Price should be the same between Chat Completions and Responses, but is not.
| model | vision | vision_mult | chat input | calculated | responses input | calculated |
|---|---|---|---|---|---|---|
| gpt-5.3-chat-latest | patch | 1.2 | 1667 | 1383 | 1833 | 1521 |
| gpt-5.3-codex | patch | 1.2 | - | - | 1833 | 1521 |
| gpt-5.2-chat-latest | patch | 1.2 | 1667 | 1383 | 1833 | 1521 |
And I have even another way of sending the same image to Chat Completions - the same vision, cheaper price for me still on the “chat” model:
| model | vision | vision_mult | chat input | calculated | r | rerror |
|---|---|---|---|---|---|---|
| gpt-5.4 | patch | 1.2 | 2813 | 2338 | E400 | E400 |
| gpt-5.3-chat-latest | patch | 1.2 | 1548 | 1284 | E400 | E400 |
Vision price inflation
Ultimately, when I integrate what the API is currently costing into my own calculator, - the same image, downsized manually to the old “high” resolution (which doesn’t happen as documented), vision still costs 3x on mini and 3.2x on nano new models.
Updated models:
hotnova.com
Discussion in the ATmosphere