OpenAI auto-switches to GPT-5.5 Thinking

ChatGPT’s default mode just got a little less literal. What used to look like one fast model is now more of a traffic cop — OpenAI says “Instant” can quietly decide when a prompt is simple enough for the lighter model and when it should hand the job to GPT‑5.5 Thinking instead. (help.openai.com) That matters because most people do not want to babysit a model picker. They want speed when a task is easy and deeper reasoning when a task turns messy. OpenAI is basically trying to hide that tradeoff behind one button. ### What actually changed in ChatGPT? The headline change is in the default experience. GPT‑5.5 Instant is now the standard model for logged-in ChatGPT users, and when you choose Instant, ChatGPT can automatically switch from the fast path to GPT‑5.5 Thinking for more complex requests. (openai.com) OpenAI also says the interface can show when GPT‑5.5 Thinking or GPT‑5.5 Pro starts reasoning, sometimes with a short preamble before the answer. ### Why hide the switch behind “Instant”? Because the old model-picker problem is real — most users do not know when to trade latency for reasoning depth. OpenAI’s answer is a single auto-switching system: easy prompts stay fast, harder prompts get more deliberate reasoning. (help.openai.com) The catch is that the thinking trace may not always appear if the routed reasoning is brief, so the system is not turning every hard answer into a big visible chain-of-thought moment. ### Who gets manual control? Paid users do. OpenAI says Plus, Pro, and Business users have access to the model picker, where they can manually choose GPT‑5.5 Instant or GPT‑5.5 Thinking. GPT‑5.5 Pro sits one tier higher and is limited to Pro, Business, Enterprise, and Edu plans. (help.openai.com) OpenAI also lets users turn automatic switching on or off in the picker’s Configure settings, which is the part developers and power users will care about most. ### Is this just branding, or is the model better? There is a real model update underneath it. OpenAI says GPT‑5.5 Instant improves factuality, STEM responses, image understanding, and when to use web search. In OpenAI’s internal evaluations, GPT‑5.5 Instant produced 52.5% fewer hallucinated claims than GPT‑5.3 Instant on high-stakes prompts, and 37.3% fewer inaccurate claims on especially difficult conversations flagged for factual errors. (help.openai.com) Those are big numbers for a default model people use all day. ### What’s happening on the cyber side? OpenAI paired the ChatGPT change with a security-access expansion. Trusted Access for Cyber — TAC — now includes GPT‑5.5 for verified defenders, and OpenAI is rolling out GPT‑5.5‑Cyber in limited preview for teams securing critical infrastructure. (help.openai.com) The idea is simple: lower refusals for legitimate defensive work like malware analysis, reverse engineering, detection engineering, and patch validation, while still blocking offensive abuse like credential theft or malware deployment. ### Why make cyber access harder to get? Because OpenAI thinks cyber capability is one of the clearest dual-use risks in frontier models. The company says TAC is identity- and trust-based, not open access, and users with the highest cyber-capable permissions will need phishing-resistant protections. (openai.com) More specifically, members using the most permissive cyber models must enable Advanced Account Security starting June 1, 2026. ### So what does this mean in practice? For regular ChatGPT users, the product is moving toward “one mode that figures it out for you.” For power users, OpenAI is still keeping manual knobs. And for security teams, OpenAI is drawing a sharper line between general access and high-capability access under tighter identity checks. (openai.com) That split — convenience for everyone, controlled capability for sensitive work — looks like the real strategy here. ### Bottom line? OpenAI is trying to make stronger reasoning feel invisible in ChatGPT, while making stronger cyber capability feel gated. Same company, same model family — but two very different distribution philosophies. (help.openai.com) (openai.com)