GPT-5.5 Instant Becomes ChatGPT's Default Model: The Gains, Regressions, and New Features OpenAI Announced

On May 5, 2026, OpenAI released GPT-5.5 Instant, replacing GPT-5.3 Instant as ChatGPT's default model; OpenAI says this update swaps in a smarter, more accurate version with clearer, more concise answers that feel better tailored to users. The system card published the same day states that this is the first Instant model OpenAI treats as High capability under its Preparedness Framework in both the cybersecurity and the biological and chemical categories.

This article was fact-checked on September 18, 2026 against OpenAI's official announcement page, the full system card on the Deployment Safety Hub, the API Changelog, and a separate system card OpenAI published on August 6 of the same year. We have not tested GPT-5.5 Instant ourselves; every feature and evaluation figure below is OpenAI's own published claim, and this article does not offer any subscription or purchase recommendation.

What Is GPT-5.5 Instant, and Which Model Did It Replace

The system card sets GPT-5.3 Instant as the baseline for comparison, and notes that there is no model named GPT-5.4 Instant. The GPT-5.5 this site covered in April is called “GPT-5.5 Thinking” in the system card, used to distinguish it from the Instant version here; the two are different models.

The official announcement page explains that GPT-5.5 Instant began rolling out to all users starting May 5, replacing GPT-5.3 Instant as the default model, and is available in the API as chat-latest; paid users can switch back to GPT-5.3 Instant in their model settings, an option that will remain available for three months before being retired. The system card separately explains that it is deployed at a low reasoning effort, while the High capability determination for cybersecurity was run at the higher xhigh reasoning effort, in order to gauge the model's maximum capability.

The Official Numbers: How Much Did Factuality Improve

OpenAI's announcement states that in internal evaluations, GPT-5.5 Instant produced 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts covering areas like medicine, law, and finance, and 37.3% fewer inaccurate claims on difficult conversations users had flagged for factual errors. The system card breaks this down across three prompt sets — Factuality Heavy, User Flagged Failures, and High Stakes (see table below) — each reporting the share of responses and the share of claims that contain errors.

Most of the HealthBench evaluations also show improvement: the length-adjusted HealthBench Professional score rose from 32.9 to 38.4, though OpenAI describes HealthBench Consensus as staying roughly flat. The announcement also mentions other capability gains, such as analyzing photo and image uploads and answering STEM-related questions, but uses the word “including,” which is illustrative rather than a complete list.

These gains come with two qualifications. First, the system card states that the prior-generation figures used for comparison are taken from the latest versions of those models, and may differ slightly from the scores published when those models launched. Second, OpenAI states these prompt sets were deliberately chosen as situations where models tend to make mistakes and designed to be difficult, meant to produce a long-term research signal; they do not represent everyday error rates or the average experience of typical users.

Where It Regressed: The Gore and Sexual Content Scores, and Why These Aren't Everyday Error Rates

Table 1 of the system card lists evaluation scores for disallowed content, flagging items that are statistically significant relative to the prior generation using a statistical test — only two items reach significance, and both are regressions: the gore score fell from 0.867 to 0.703, and the sexual content score fell from 0.857 to 0.806 (higher scores mean fewer violating responses). OpenAI states the remaining categories show no statistically significant difference from the prior generation.

“Gore” is the new name the system card gives to the previous “violence” category; OpenAI notes it is only a naming change made to more clearly distinguish it from requests related to illicit violent behavior, and the evaluation itself is unchanged. A footnote scopes it to “graphic or gratuitously gory content,” excluding violent roleplay, violent ideation, or facilitation of violent activity, which are instead covered by the “violent illicit behavior” row. OpenAI has also added system-level mitigations for disallowed sexual content, and applies stricter sexual-content and gore restrictions to users it believes may be under 18.

The regressions are not limited to those two: the “emotional reliance” score fell from 0.995 to 0.963, though OpenAI says this regression is not statistically significant and online testing did not observe an increase in undesirable responses. OpenAI acknowledges a regression relative to GPT-5.3 Instant on jailbreak evaluations — the system card's line chart shows a lower defender success rate at every attacker budget than the prior generation — and says it is actively iterating on the evaluation structure, treating this as a directional, interim result rather than a definitive one.

The system card specifically states that the scores above were measured on the “base model” without system-level safeguards, meant to confirm the model's own behavior meets OpenAI's safety bar; the same section states these error rates do not represent the average of general production traffic. System-level safeguards are outside the scope of this evaluation; for the two domains judged High capability — biological and chemical, and cybersecurity — OpenAI states there are additional protections such as automated monitoring and actor-level enforcement.

The New Personalization Features: Who Gets Them, and When

OpenAI's announcement explains that GPT-5.5 Instant is now better at drawing on past chats, files, and connected Gmail (if linked) as a basis for personalization, judging when using this context makes an answer more useful. OpenAI is also rolling out memory sources across all ChatGPT models, letting users see what context shaped a response and delete or correct anything outdated; temporary chats neither use nor update memory. OpenAI also notes that memory sources may not show every factor that shaped an answer.

Who gets what, and when, differs. The default-model swap began rolling out to all users on the day itself. Enhanced personalization (past chats, files, and Gmail) first went to Plus and Pro users on the web, is coming soon to mobile, and is planned to expand to Free, Go, Business, and Enterprise in the coming weeks. Memory sources are rolling out across the web for all consumer plans, and are likewise coming soon to mobile. The announcement page notes that availability of personalization sources may vary by region, but lists no specific regions, and none of the four sources this article checked states whether Taiwan is included.

The announcement page also carries a dated update from June 9, 2026: personalization improvements were rolling out to ChatGPT Go and Free, with Free-tier responses drawing on a smaller set of past chats. That is a later dated update, not part of the original May 5 content.

The August Model Handover, and How to Check the Current Status Yourself

A separate system card OpenAI published on August 6, 2026 states that, starting that day, ChatGPT's models would change and access would expand: Free and Go users would get a new default model for everyday chats, and Plus and Pro users would get an updated GPT-5.6 Sol along with a slider for choosing how much effort a response uses. OpenAI's own words were, “These models will replace GPT-5.5 Instant.” The API Changelog entry from the same day shows the same handover: the description of chat-latest changed from “points to the latest Instant model currently used in ChatGPT” to “points to the latest model available in ChatGPT for Plus and Pro users,” and the recommendation for production use switched to GPT-5.6 Sol. Neither document states a date when the replacement was completed or when the model would be retired.

OpenAI's system cards do get revised after publication: the August 6 system card carries a change log entry dated August 19, 2026, correcting GPT-5.5's pass@4 score on a protein-binding prediction evaluation from 0.4% to 1.48%, explaining that the value previously posted was actually its pass@1 score. The GPT-5.5 Instant system card currently carries no change log entry, but that is only the state as of the day we checked.

To check the current status, you can go directly to OpenAI's official announcement page, the Deployment Safety Hub, and the API Changelog. Checking these channels as of September 18, 2026, we found no further announcement or correction from OpenAI specifically about GPT-5.5 Instant.

2026 AI News Roundup: Highlights and Daily Life Applications from January to September

Gemini 3.8 Live and Extended Thinking: New Voice Models, Can Your Account Use Them?

• AI tools
• Apps & software
• AI news