OpenAI changed the default model behind ChatGPT on 5 May 2026, replacing GPT-5.3 Instant with GPT-5.5 Instant across free and paid tiers. The company's framing was unusual for an AI lab: rather than emphasising new reasoning capability, the announcement focused on faster responses and fewer hallucinations in high-stakes domains — explicitly naming law, medicine, and finance. The change affects hundreds of millions of daily users.
Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, which is also named in this article. Readers should weigh that interest.
What changed for users
For the average ChatGPT user the change is invisible: same UI, same conversation flow, same product. What shifts is latency and reliability under load. OpenAI's own emphasis is on factuality: it says the new model is more dependable, with the largest gains in domains where accuracy matters most.
According to WhatLLM's May 2026 model survey, GPT-5.5 scored 60.24 on the Intelligence Index at high-effort reasoning — the highest score on the index, set when GPT-5.5 launched on 23 April. The Instant variant trades some of that reasoning for response speed, the trade-off OpenAI has settled on as the right default.
The "fewer hallucinations" claim
OpenAI's specific framing — fewer hallucinated claims on high-stakes prompts in areas such as medicine, law and finance — is targeted at the three verticals where its enterprise customers face the most pushback from professional regulators. The American Bar Association and the World Health Organisation have both published guidance since 2024 constraining how LLMs can be cited or used in regulated workflows. A measurable drop in hallucination rate inside those verticals is the kind of metric OpenAI's enterprise sales team can put in a procurement document.
The catch: OpenAI's figures — 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts, and 37.3% fewer inaccurate claims on conversations users had flagged for factual errors — come from its own internal evaluations, and as Air Street Press noted in its May 2026 state-of-AI summary, vendor-published behaviour claims await independent verification through structured evaluations like SimpleQA-Hard or HELM-Lite.
The wider May 2026 pattern
OpenAI is not alone. Google made Gemini 3.1 Flash Lite a new gateway default around 8 May, mirroring the same speed-over-scale philosophy. After April's frontier sprint — when GPT-5.5, DeepSeek V4, Kimi K2.6, and Claude Opus 4.7 all shipped in a single window — May has been a month of operational tuning rather than model races. The labs that didn't ship in April (Anthropic past Opus 4.7, Mistral, Qwen Max) are reportedly cooking new releases; the labs that did ship are optimising for cost, latency, and customer retention.
For developers building on top of ChatGPT's API, the practical implication is that the same prompt may now return slightly different output than it did before 5 May. Applications that rely on output stability should evaluate whether their pinned model version is GPT-5.5 Instant or an earlier Instant build. LLM-Stats maintains a live tracker of model swaps for production teams who need to know.
Correction, 18 October 2026: An earlier version said OpenAI had not published hallucination figures for the new default; it reported 52.5% fewer hallucinated claims on high-stakes prompts and 37.3% fewer inaccurate claims on flagged conversations, from internal evaluations. We have also removed an unsourced statement attributed to OpenAI about reasoning gains plateauing, and corrected when GPT-5.5 set its Intelligence Index score and the date of the professional guidance cited.
Sources and cross-checks: Primary: WhatLLM — New AI Models May 2026. Corroborated against: Air Street Press — State of AI: May 2026 and LLM-Stats AI Updates Today. 60.24 Intelligence Index score and default-swap date verified across all three sources 18 May 2026.
View author profile → · Editorial policy
