OpenAI begins safety testing of new model o3 that uses ‘reasoning’ process to ensure reliability in complex programming, maths tasks

OpenAI said it has begun testing a new “reasoning” AI model family, o3 and o3-mini, amidst heated competition with major tech firms such as Google.

Google introduced the second generation of its own model, Gemini, earlier in December.

o3 is the successor to o1, which launched in preview in late September and in full on 5 December.

The company skipped the designation “o2” due to its similarity with the O2 telecommunications brand in the UK, The Information reported.

‘Reasoning’

o1 and o3 are both “reasoning” models, which are designed to spend time analysing their own responses before answering a query.

The process adds significant time and computing resources to each task, in turn making tasks expensive to carry out, but it also improves quality, according to advocates.

The reasoning models are particularly reliable for complex reasoning, science and programming tasks.

The models are the successors to GPT-4o, which was released in May with voice capabilities.

OpenAI chief executive Sam Altman said the company plans to launch o3-mini, which is fine-tuned for particular tasks, by the end of January and full o3 after that.

Safety research

The company said safety researchers can sign up for a preview of o3-mini beginning immediately.

Safety research is in part designed to ensure that a model has not become so intelligent that it begins intentionally deceiving humans.

The application for researchers closes on 10 January, the firm said.

OpenAI’s continuous release of new and more powerful models has helped it retain the interest of investors, which piled into a $6.6 billion (£5.25bn) funding round that concluded in early October, valuing the loss-making start-up at $150bn.