OpenAI closes gap to artificial general intelligence with GPT-5

OpenAI has recently unveiled the latest version of its large language model (LLM) in ChatGPT, now upgraded to GPT-5. This development marks a significant stride towards achieving artificial general intelligence (AGI). According to a recent blog post by the company, GPT-5 boasts improvements in accuracy, speed, reasoning, context recognition, structured thinking, and problem-solving capabilities.

OpenAI expressed its belief that the adoption of GPT-5 will pave the way for industry leaders to explore the full potential of AI powered by this advanced model, leading to enhanced decision-making, improved collaboration, and expedited outcomes for high-stakes projects within organizations.

From a technological standpoint, GPT-5 is built around a unified system that combines a smart and efficient model for answering most queries, a deeper reasoning model for tackling more complex problems, and a real-time router that swiftly determines the most suitable approach to take. The company highlighted that the router is continuously trained on real data signals, including user preferences, response rates, and accuracy metrics.

One notable feature of the router is its ability to allow the LLM to continue functioning using a smaller model once usage limits are reached.

Another significant update with GPT-5 is the introduction of a new safety training method known as “safe completions.” This training aims to teach the model to provide the most helpful responses while adhering to safety constraints. OpenAI mentioned that in certain scenarios, GPT-5 may offer partial answers or refuse to provide detailed information, but it is trained to explain its reasoning transparently and suggest safe alternatives.

Moreover, OpenAI claimed that GPT-5 excels in AI coding tasks compared to its predecessors. The model has demonstrated the ability to delve deep into complex codebases and offer insights on how different components function and interact. This capability has proven invaluable in accelerating daily coding tasks, particularly within intricate codebases like OpenAI’s reinforcement learning stack.

In a benchmark test known as SWE-Bench Verified, where models are tasked with generating code patches to resolve issues based on code repositories and problem descriptions, GPT-5 achieved an accuracy rate of 75% with approximately 10,000 tokens. This outperformed OpenAI’s o3 model, which scored 69% accuracy using 13,741 tokens.

We’re at a tipping point. GPT-5 promises even more realism, more precision, and more ease for the user. That’s great for innovation, but it’s also a gift to fraudsters

Grant Farhall, Chief Product Officer at Getty Images, shared his insights on the launch of GPT-5, emphasizing its potential to reshape the dynamics of content, creativity, and imagery within the industry.

As AI continues to evolve towards AGI, concerns about potential fraudulent use of such advanced systems are growing. Gary Hall, Chief Product Officer at Medius, cautioned about the risks associated with the increasing realism, precision, and user-friendliness of GPT-5, highlighting the challenges this poses for legacy finance systems.

Furthermore, GPT-5 is not limited to OpenAI’s platform but is also integrated across various Microsoft platforms, including Microsoft 365 Copilot, Copilot Studio, Microsoft Copilot, GitHub Copilot, Visual Studio Code, and Azure AI Foundry.