Krux
March 4, 2026
OpenAI Cuts Hallucinations 27% in GPT-5.3 Update
Published: March 4, 2026 at 6:39 AM
Updated: March 4, 2026 at 6:39 AM
100-word summary
OpenAI's GPT-5.3 Instant, released today, makes fewer things up. High-stakes testing showed hallucinations dropped 27% when the model searches the web and 20% when it doesn't, compared to earlier versions. The update also stops littering answers with unnecessary warnings and caveats. It balances web results with its own reasoning instead of just parroting search results. The biggest shift: ChatGPT now gives you a straight answer instead of hedging with "I should mention" preambles. For anyone tired of fact-checking every AI response, this is the first meaningful trust upgrade in months.
What happened
OpenAI's GPT-5.3 Instant, released today, makes fewer things up. High-stakes testing showed hallucinations dropped 27% when the model searches the web and 20% when it doesn't, compared to earlier versions. The update also stops littering answers with unnecessary warnings and caveats. It balances web results with its own reasoning instead of just parroting search results. The biggest shift: ChatGPT now gives you a straight answer instead of hedging with "I should mention" preambles.
Why it matters
For anyone tired of fact-checking every AI response, this is the first meaningful trust upgrade in months.