A few weeks ago, OpenAI released GPT-5.6. However, they only released them to “trusted partners”, because of bullshit fear-mongering that Anthropic has decided to stir up with Mythos. A few days ago, they finally released GPT-5.6 to the public, and I gotta say, its a good model. Problem is, it eats up subscription limits like I eat food. I can’t tell if this is due to them lowering subscription limits, cache write costs being implemented, or if the model overall just costs more.
I’m grateful for their choosing to get rid of the 5 hour limits, but I’ve used Sol for a small Unity project and I went down from 99% → 92% with around 15 minutes of work. Interrupted. With GPT 5.5, I wouldn’t even get anywhere close to the 5 hour limits. I once sent GPT 5.5 on a cryptography challenge, it worked for over 3 hours and my 5 hour limit went down by 10% in total. For roughly 20$ of inference. That is insane, and Sol drives it down.
GPT-5.6-Terra also exists. It’s actually a good model. Cheap, fast, and actually useable for work. I’m probably going to switch over to using Terra because of the limits that have been imposed, as I cannot afford a 100$ a month plan or paying API pricing.
Now OpenAI has caught up in the developer side, with a model that will not stop working until it has completed the task, but I think that OpenAI should slow down and with GPT 5.7, instead of making the model smarter, they should fix where GPT-5.6 is still lacking. GPT-5.6 isn’t exactly great at frontend. If you push it for a specific design style, it will follow it well. It just wont do good frontend by itself. OpenAI models have also not been very conversational since the GPT-5 days. GPT-4o was the most conversational GPT model that OpenAI has ever released. They really need to make GPT-5.7 conversational, as it’ll improve overall user experience.
One final thing that I really doubt that OpenAI will crack with GPT-5.7, or any AI company for that matter, critical thinking. LLMs exhibit no critical thinking. They never come up with good ideas. You cannot prompt your way around this—It’s a fundamental limitation of our brains being much larger than LLMs. Our brains contain roughly 80B neurons. Some might say, but LLMs have over 1 trillion parameters! Parameters are NOT the same as neurons. Parameters describe the connections between neurons. A simple neural network that has 3 input, 2x10 hidden layers, and 3 output neurons has 30 + 10 parameters for the weights + biases of the input layer, 100 + 10 parameters for weights + biases of the hidden layer, and 30 + 3 weights and biases for the output layer. Those 26 neurons produced 183 parameters. Now 80 billion neurons? Depending on the configuration, that could be over 100T parameters! And, we run on roughly 20W. Compared to the 120,000W that a single GB300 NVL 72 rack could use.
OpenAI, you make great AI models. Address the small things with GPT-5.7, and then after make GPT-6 top every benchmark. Also you should probably release some more OPEN models. You could make some wildly efficient openweight models (both in tokens and actual model size) with what you are currently doing.