OpenAI Ships GPT-5.5 Instant as Default ChatGPT Model — 52.5% Fewer Hallucinations on High-Stakes Prompts

OpenAI promoted GPT-5.5 Instant to the default model across all ChatGPT tiers and the chat-latest API endpoint, replacing GPT-5.3 Instant. The release is explicitly not about benchmark competition — it targets the single most damaging problem in deployed LLMs: confident wrong answers in high-stakes domains.

OpenAI’s internal evaluations show GPT-5.5 Instant produces 52.5% fewer hallucinated claims than GPT-5.3 Instant on prompts in medicine, law, and finance. These are the domains where ChatGPT hallucinations carry genuine real-world risk — wrong drug interactions, incorrect legal citations, fabricated financial figures. A 52.5% reduction is not noise. It reflects deliberate training work on factual grounding rather than a general capability sweep.

What else changed

Beyond the hallucination reduction, the update delivers measurably tighter answers. GPT-5.5 Instant is less likely to pad responses with caveats and hedging when the underlying information is clear. Conversational naturalness improved, particularly in multi-turn sessions where earlier versions would occasionally lose conversational thread after context-heavy exchanges.

Personalization capabilities received a significant upgrade for Plus and Pro users: the model makes better use of context built from past interactions, uploaded documents, and connected Gmail accounts. Users who have built up a significant history in ChatGPT will notice the model feels more calibrated to their communication style and knowledge domain from the first message.

What hasn’t changed — pricing and the model tier structure

GPT-5.5 Instant maintains the same pricing position as its predecessor, competitive with similar frontier models at the everyday tier. OpenAI has not published updated per-token costs separately from the existing GPT-5 family pricing.

For developers, GPT-5.3 Instant remains accessible through explicit model configuration for a three-month transition window before retirement. The full GPT-5.5 model — with deeper reasoning and longer context processing — stays exclusive to Pro users and is not what was deployed today.

The practical takeaway

The Instant model line has always been OpenAI’s bet that “reliably useful” beats “occasionally brilliant” for the majority of use cases. A model that is 52.5% less likely to fabricate medical or legal information is more valuable to most organizations than a model that scores two points higher on MMLU. That positioning is increasingly defensible as enterprise deployments move from pilot to production and accuracy liability becomes a real procurement concern.

Personalization rollout to mobile and Free/Go/Business/Enterprise tiers follows in coming weeks.