OpenAI AI Updates: April 26, 2026

1. GPT-5.5 Tops Artificial Analysis Intelligence Index With an 86% Hallucination Rate

OpenAI. Independent post-launch evaluation pegs GPT-5.5 at 60 points on the Artificial Analysis Intelligence Index — the new top score — and 57% accuracy on AA Omniscience, but with an 86% hallucination rate when the model is wrong, versus 36% for Claude Opus 4.7. The pricing math is also more nuanced than the raw $5/$30 input/output per million tokens suggests: token-list prices are double GPT-5.4, but reported 40% lower token usage on equivalent tasks brings the effective cost increase to roughly 20%. The takeaway for buyers: GPT-5.5 is the most accurate frontier model on offer when it is right, and the most confidently wrong when it is not — pick a deployment pattern accordingly. Source

2. Sam Altman Apologizes to Tumbler Ridge Over Mass-Shooting Disclosure Failure

OpenAI. Sam Altman issued a public apology in a letter to residents of Tumbler Ridge, BC, over OpenAI’s failure to alert authorities about an 18-year-old user whose ChatGPT account was banned in June 2025 after he described detailed gun-violence scenarios — months before he carried out a shooting that killed eight people. OpenAI says it decided internally not to escalate to law enforcement and only contacted Canadian authorities after the attack. The company says it is rolling out new safety protocols and direct law-enforcement contact channels for future high-risk signal cases, an episode likely to feed into ongoing debates about how providers should handle pre-incident warning signs from chatbot users. Source