OpenAI's newest default model for ChatGPT hallucinates 52.5% less than its predecessor. That number—confirmed across internal evaluations covering medicine, law, and finance—represents the most significant factuality improvement in the company's recent history, and it arrives at a moment when competitors have been aggressively marketing themselves as more trustworthy alternatives.
GPT-5.5 Instant replaces GPT-5.3 Instant as ChatGPT's baseline model, meaning every user interaction now runs through a system that OpenAI claims produces substantially fewer fabricated claims in high-stakes domains. The company reported 37.3% fewer inaccurate statements on conversations users had previously flagged for factual errors, suggesting the model learned from real-world error patterns rather than synthetic benchmarks alone.
The timing matters. For roughly eighteen months, Perplexity and Anthropic's Claude positioned themselves as safer choices for research-intensive workflows, exploiting OpenAI's well-documented tendency to generate confident falsehoods. Medical professionals, lawyers, and financial analysts had begun maintaining elaborate prompting strategies—citation checks, cross-referencing, explicit uncertainty requests—simply to work around models that could not stop making things up. GPT-5.5 Instant aims to eliminate the need for those workarounds.
OpenAI published a system card alongside the launch, documenting safety evaluations and known limitations. The company acknowledged these results come from internal testing, a detail worth noting: external researchers have not yet replicated the 52.5% figure. Historical context suggests some skepticism is warranted. OpenAI has revised hallucination metrics before, and the definition of what constitutes a "hallucinated claim" can shift the numbers meaningfully.
What remains unchanged is latency. GPT-5.5 Instant maintains the sub-second response times that made the Instant line popular, meaning this accuracy improvement arrives without the speed penalty that typically accompanies more careful reasoning. For developers building user-facing applications, the calculus has shifted: a model that answers questions correctly on the first pass eliminates the retry loops and verification steps that slow down production systems.
The release also includes enhanced personalization controls, though OpenAI provided few details on how these differ from existing memory features. The factuality story dominates the announcement, which is deliberate. After months of competitor ads highlighting AI's trustworthiness problem, OpenAI is reclaiming the narrative with the one metric that cuts through: a concrete percentage reduction in made-up information.
GPT-5.5 Instant is available now as ChatGPT's default model at no additional cost to existing subscribers. OpenAI has not announced plans to extend the underlying architecture to its API, leaving developers to decide whether the improvements justify migrating from GPT-4o or other models currently in production.