A deleted GitHub post gives an early look at OpenAI’s next major model, GPT-5

• gpt-5: Built for logic and multi-step tasks
• gpt-5-mini: A lightweight option for cost-sensitive scenarios
• gpt-5-nano: Tuned for speed and low-latency applications
• gpt-5-chat: Designed for advanced, multimodal, and context-aware conversations in enterprise settings

GitHub says GPT-5 supports more autonomous task execution, working with fewer and shorter prompts. The model is designed to provide clearer explanations and be more context-aware, especially in enterprise and software development environments.

Incremental, not revolutionary

A recent report from The Information lowers expectations for GPT-5. Internal testing suggests the model brings improvements in math, coding, and following instructions, but the performance jump is smaller than what users saw going from GPT-3 to GPT-4.

OpenAI's original candidate for GPT-5, a large language model codenamed "Orion," didn't meet expectations and was instead released as GPT-4.5. That version only offered marginal improvements, ran slower and cost more than GPT-4, and quickly faded from view.

OpenAI has also tested "reasoning models" like o1 and o3, which performed well in specialized domains but lost key abilities when adapted for chat. For example, o3-pro excelled in expert benchmarks but struggled with basic conversation, sometimes using excessive compute just to generate simple greetings.

With GPT-5, OpenAI is aiming for a better balance between advanced reasoning and reliable, everyday communication. The new model reportedly includes mechanisms to dynamically allocate compute resources based on task complexity, possibly avoiding the kind of overthinking that plagued earlier models.