gpt-5-mini-2025-08-07 is being deprecated on December 11, and I’m struggling to find a replacement with comparable vision performance at a similar price.
I’ve tested several of the suggested alternatives on food image recognition, including gpt-5.6-terra, gpt-5.4-mini, gpt-5.6-luna, gpt-6-luna, and gpt-6-sol.
The main issue is identification accuracy. Several newer models consistently misidentify obvious foods that gpt-5-mini handles correctly. Examples include rice vs. hash browns, chicken vs. tuna or shrimp, and incorrect dish classification. For example, gpt-5.4-mini cannot identify Thai green curry, calling it “green soup”
gpt-6-sol performs well, but it is prohibitively expensive. gpt-5.6-terra, the recommended replacement, is also significantly more expensive without matching gpt-5-mini on image understanding in my tests.
For this use case, gpt-5-mini is still the best vision model out there. It still seems to occupy a useful middle ground between price and vision quality that the newer models don’t currently replace.
Is there a closer replacement planned before the shutdown date? I’d also strongly support keeping gpt-5-mini-2025-08-07 available longer while the newer vision models catch up.
fyiy; I ran over 8000 test analyses across 50+ dishes among all of the possible models, nothing came close to gpt-5-mini except for sol. which took significantly longer and cost a huge amount.
Happy to publish my findings if that helps. Please don’t sunset this model
The vision technology has changed after gpt-5.2, to “patches” instead of 512 pixel square tiles being billed.
There are a few tweaks you can do:
Understand the latest models’ “original”, and “high” input maximums for resolution, which by default is extremely high, and do your own resizing and downsizing. More tokens consumed by an image, a 12000 or 30000 token input image bill or more, does not necessarily lead to a more comprehensive “view” or a win due to the way self-attention works. Alternately, ensure upscale to a suitable size for vision.
Re-sample the constrained-size inputs with offsets. For example, shift the imagery 7 pixels diagonally for a duplicate copy of the image to be sent. This will produce a second embedding of the image with different semantics that can alter the understanding.
You will be able to find an input resolution for the type of imagery that yields the best performance when you take control of the imagery being sent, resizing and improving it on your end.
I took five images the cheaper models consistently misidentify and tried each at 512, 768, 1024, and 1536 px, plus a duplicated 1024 px image shifted slightly. Three runs each, same prompt, low reasoning.
gpt-6-luna barely improved with higher resolution, while gpt-5-mini got essentially all of them right at every size.
So in my testing:
Smaller images made performance worse.
Larger images barely helped and increased image tokens.
Sending a shifted duplicate did not help.
Token counts were identical across the models I tested for the same image.
That suggests this is not an image resolution or tiling issue. The same image data is reaching the models, but the models are interpreting it differently.
Also, these tests were run October 2–3, after the September 25 image encoding fix mentioned in the changelog.
I am hoping OpenAI will release some other option before Dec 11. I will be very sad to move to another LLM because my business logic is all built around openAI’s offerings
There’s not really a “well, we’ll need to pay a bit more” model after luna - there’s a big price jump.
I’ve also found gpt-5-mini to be a good performer, where slottiing in gpt-5.4-mini turns an app to hot garbage.
But “mini” wasn’t the smallest - that was “nano”. The smallest “gpt-5-nano” model was terrible - is it any surprise the smallest model remains bad?
So there might be a particular “magic” in that distinct size allocation and level of pre+(lack of) post training on gpt-5-mini and its type of vision.
How’s the reasoning effort at affecting the quality? I don’t think it should be set lower than default with newest models nor turned off with “none” as there is an expected shape the AI is trained on that perfoms better alongside the CoT itself.
I admit i am not an expert in models structures and functions, yet… my question is: is it possible to take one model’s typical feature (in this case gpt-5-mini vision functionality) and transfer it ‘as is’ to a new model?
If the answer is ‘yes’, why don’t they do it?
I don’t know how vision capability of newer models is build, but once it is granted that gpt-5-mini has the best vision capability and performance, why not using the same in new models?
Seeing features degradation in newer models is by far a constant at OpenAI, and this is disturbing. I myself have posted here more than once highlighting degraded features, this is another huge case.
I have always been a huge fan of OpenAI models, i have built more than 80 bots with different models and in different scenarios (from simple chats to intelligent webapps to vibe coding tools), from 3.5 to 5.4-nano and lately i have been experimenting with 5.6-luna in my vibe coding tools, but I admit that lately i have found myself suggesting another provider’s model to a client for peculiar tasks, because the new OpenAI models cannot replicate the performance of gpt-4.1-mini that was used in the preious version of the client’s tool.
On the Dec 11 date for gpt-5-mini-2025-08-07: the deprecations page (Deprecations | OpenAI API) lists gpt-5.6-terra as the recommended replacement. On the Standard pricing table (short context, per 1M tokens, input / output) gpt-5-mini is $0.25 / $2.00 and gpt-5.6-terra is $2.00 / $12.00, so 8x on input and 6x on output. The same table has gpt-5.6-luna at $0.20 / $1.20 and gpt-6-luna at $0.10 / $0.50, which are cheaper, though you found the newer cheap models misidentify foods.
@qhenkart, for your 8000+ run comparison: roughly what did it cost in time and spend, and what accuracy on your dish set would make you move off gpt-5-mini?