OpenAI's most powerful AI: New GPT-5.4 model unveiled with 'thinking' feature

In addition to the standard version, users have access to GPT‑5.4 Thinking (for reasoning) and GPT‑5.4 Pro (for high performance).

The developer version can process up to one million units of information (tokens) simultaneously — the largest data capacity of any OpenAI product.

The model has also become more efficient: it performs the same tasks while using significantly fewer resources than its previous version.

Moreover, GPT-5.4 achieved record results in tests measuring computer skills, general knowledge, and professional task performance.

Features of the new GPT model

According to Brendan Foody, CEO of Mercor, the new model leads in professional knowledge tests in law and finance. GPT-5.4 has also become significantly more accurate: it makes 33% fewer errors in specific facts and 18% fewer general mistakes compared to the previous version.

OpenAI has updated the technical side as well, adding a Tool Search feature. Now, the AI doesn’t load instructions for all tools at once; it searches for the needed ones only when required. This makes the system faster and more cost-efficient.

Special attention has been given to safety: developers introduced a "chain-of-thought" monitoring system. This allows users to see the AI’s internal reasoning when performing complex tasks. Such oversight ensures that the neural network does not conceal its actual logic from developers.

Deception is less likely in the GPT-5.4 Thinking version, indicating that the model cannot hide its reasoning, and chain-of-thought monitoring "remains an effective safety tool," OpenAI notes.

Research has shown that AI-based chatbots lose reliability and make more mistakes during extended conversations.

For instance, when executing a single command, models demonstrate a 90% success rate, but in multi-step dialogues, this drops to 65%. While overall task suitability decreases by only 15%, their unreliability increases by 112%.