OpenAI Unveils GPT-5 for Developers

OpenAI has launched GPT-5, its latest model specialized for coding and agentic tasks, through the API platform. GPT-5 achieved SOTA (State-of-the-art) performance in coding, scoring 74.9% on SWE-bench Verified and 88% on Aider polyglot.

The model is designed as a true coding collaboration tool that goes beyond simple code generation, handling bug fixes, code editing, and Q&A on complex codebases. It particularly excels at frontend development, achieving a 70% win rate against OpenAI o3 in internal testing.

Agentic capabilities have also been significantly enhanced. It achieved an overwhelming score of 96.7% on the $\tau^2$-bench telecom benchmark, and can reliably perform complex tool calling both sequentially and in parallel.

New API features have also been introduced to give developers greater control:

• verbosity parameter: Adjusts the level of detail in responses (low, medium, high)
• reasoning_effort parameter: Controls reasoning intensity to optimize response speed
• Custom tools: Supports calling tools using plaintext instead of JSON

The model is offered in three sizes—gpt-5, gpt-5-mini, and gpt-5-nano—balancing performance, cost, and latency.