I really cannot understand the policy OpenAI applies to model usage costs…
gpt-5.6 has come up with high prices, and then they were cut to affordable pricing, especially Luna, which, as of now, has 0.20 USD/1M tokens IN and 1.20 USD/1M tokens OUT.
That is very affordable, even more keeping in mind that gpt-5.6-luna is great at coding… I have recently added it to the available models in my vibe coding tool CertainThing, now equipped with 4 models (4o-mini, 5-nano, 5.5-nano, 5.6-luna) and while compiling the pricing table for the openai call cost calculation I have noticed that Luna now costs the same as gpt-5.4-nano (which is actually 1.25 USD/1M tokens OUT) and is even cheaper than gpt-4.1-mini which is 0.10 USD/1M tokens IN but 1.40 USD/1M tokens OUT.
And this is what i don’t understand. Frontier models are becoming sometimes cheaper than older (but still well performing) models.
Even more surprising is that the cheapest model as of now is gpt-5-nano, with an outstanding cost price of 0.05(!!!) USD/1M tokens IN and 0.40 USD/1M tokens OUT… and how about gpt-5-mini?! It has an output cost that is 1.5x that of Luna…
Technology is moving multiple directions as far as I can tell.
size
active parameters
distillation
quantization
context compression
hardware utilization
training quality
(let me know if I forgot any)
usually size and active parameters yield a better model, a more expensive model.
but all the other points reduce cost. There are a lot of optimizations and incremental improvements coming out of the various labs, and most are concerned with with pushing size down while retaining quality. Of course, these efforts aren’t always successful, but by my estimation that’s been the industry focus for a bit now.
So it’s not that surprising that a newer generation small model might outperform a last generation small model in terms of price, although not always necessarily in performance. But sometimes it does both. Lower cost for you, improved margins for OAI (maybe), everyone wins.
I understand your point but still i don’t understand… looks like OAI wants to discourage the use of older models that are still perfect in 90% of cases… i have my clients’ chatbots and code generation tools all based on 4.1-mini which is perfect for the scope… and now i should be forced to change it to luna because it is cheaper. For clients is a win… but.. i miss the logic. Honestly i as a producer would have kept oldest models always cheaper than new ones. But maybe i am missing something…
the thing is that models weights occupy space in the rigs. and they’re quite substantial.
keeping a terabyte size model on a box eats.. well, a terabyte of nand. and keeping a model ‘hot’ eats that in memory.
If they can sell you a smaller or faster model at a similar perceived performance point for slightly less money, it’s still a massive win for them.
I don’t know how big mini and luna are. Chatgpt seems to think luna is in the gemma 4 31BP ballpark, and mini drastically below that in terms of performance. Llama 70BP from 2 years ago underperforms gemma 4 at 31BP in my opinion. I don’t spend much time with OpenAI’s tiny models, so I can’t give a qualitative judgement on that, but it’s plausible.
While we don’t know what’s happening in the ivory tower here, the industry move is that you can indeed provide the same or better capability at less than half the cost (31BP vs 70BP). Keeping the bigger model around and loaded into your rigs really doesn’t make sense - it literally costs more than twice as much.
Plus - as I alluded to earlier - the newer generation models have a drastically lower context cost. You can now stuff many many more tokens into a gigabyte of ram than you could a year ago.
So the economics, the operational considerations basically dictate that old models must be retired to pry any sense of a margin into this business.
Hope that makes sense.
But I’m absolutely with you - I’m also shedding tears for some of the OGs.
I suspect if you want to guarantee long-term access to a model for business continuity reasons you won’t get around self-hosting.
ETA: to clarify - the jury’s still out on how much capability the newer, smaller models actually preserve. A lot of these money saving efforts have yielded mixed results, and if you say that 4.1 mini outperforms luna for your particular use-case… …that is (to me) a believable claim I wouldn’t discount.
