OpenAI has released GPT-5, and the results are striking. The model achieves state-of-the-art scores on every major benchmark, including MMLU, HumanEval, and the newly introduced Frontier Math dataset.

What’s New in GPT-5

The most significant improvement is in multi-step reasoning. GPT-5 can now decompose complex problems into sub-tasks, verify its own intermediate steps, and backtrack when it detects an error — a capability that was largely absent in GPT-4.

Multimodal understanding has also taken a leap forward. The model can analyze charts, diagrams, and photographs with a level of accuracy that rivals human experts in several domains.

Implications for Developers

The API is already available to developers, with a context window of 256,000 tokens. Early benchmarks from the developer community suggest that GPT-5 can handle entire codebases in a single prompt, dramatically simplifying tasks like refactoring and documentation generation.

The Competition Responds

Google DeepMind and Anthropic are expected to release competing models within weeks. The AI arms race shows no signs of slowing down.