ChatGPT o3: Price Slashed by 80% – Has Quality Been Compromised?

Alongside the launch of its new flagship model, ChatGPT o3, OpenAI has significantly lowered the prices for its advanced reasoning model. The cost in the API has dropped by 80%, now standing at $8 per million outbound tokens and $2 per million inbound tokens. For ChatGPT Plus subscribers, which is the standard tier for most users, the limits for o3 have been doubled: users can now make up to 200 requests per week instead of 100.

OpenAI representatives attributed this dramatic price reduction to the optimization of the company’s server infrastructure. However, allegations surfaced on X that OpenAI may have actually made the model «dumber,» possibly through quantization or distillation.

**Quantization** involves compressing a neural network’s weights by replacing 32-bit numbers with 8- or 4-bit counterparts to reduce memory usage and increase processing speed with minimal quality loss. **Distillation**, on the other hand, is the process of training a smaller, more efficient model (the student) using the outputs of a larger model (the teacher) to maintain performance with fewer resources.

The first response to these accusations came from Aidan McLaughlin of OpenAI, but it’s important to note that he has a vested interest in the outcome of this discussion.

Subsequently, an independent evaluation emerged. Representatives from the non-profit organization ARC Prize reported that they re-tested the o3 model using the ARC-AGI benchmark, designed to assess the models’ abilities to reason, tackle fundamentally new problems, and adapt to unfamiliar situations. According to ARC Prize, the performance of the reduced-cost o3 in the benchmark remained unchanged.

On a personal note regarding claims of «dumbing down,» I find such comments often arise within one or two months of using nearly every model I’ve encountered. While there might occasionally be adjustments to settings or system prompts, most of the time, it’s more about psychology— the initial «wow effect» from a new model fades, we start to grasp its capabilities, become frustrated by its limitations, and then think it has «dumbed down.»

P.S. As usual, you can support me by subscribing to the channel “[escaped neural network](https://t.me/ai_exee),” where I discuss AI from a creative perspective.