During an uncharacteristically long video stream yesterday, OpenAI announced GPT-5, alongside a series of interface and usability improvements to its chatbot. Here’s everything new with ChatGPT.
One model to rule them all
After many years of confusion with similarly-named models imbued with overlapping abilities, OpenAI finally streamlined the user experience and trimmed down its model offerings to:
GPT-5
GPT-5 Thinking
GPT-5 Pro (limited to the US$200/mo plan)
OpenAI says that ChatGPT will intelligently decide how much time it should spend calculating the answer to the user’s prompt, but the user can manually select GPT-5 Thinking if they want to ensure a longer reasoning process.
And for the first time, free users will have access to the same model as Plus users, but with lower usage limits. Once those limits are reached, they’ll be switched to a ‘mini’ version of the model. The same applies to Plus users, though their limits will be more generous. In both cases, however, the exact thresholds remain unclear.
OpenAI also decided to deprecate all its other models. As a result, past conversations with older models “will automatically switch it to the closest GPT-5 equivalent. Chats with 4o, 4.1, 4.5, 4.1-mini, o4-mini, or o4-mini-high will open in GPT-5, chats with o3 will open in GPT-5-Thinking, and chats with o3-Pro will open in GPT-5-Pro (available only on Pro and Team).”
As for the Voice Mode, OpenAI says ChatGPT will keep on being powered by GPT-4o for now.
And speaking of the Voice Mode, it now works with custom GPTs, and lets paid users give specific instructions regarding answer length, speed, tone, and more.
Performance
OpenAI showed (with a couple of messy charts) that while the new models do perform better than their predecessors in multiple benchmarks, including SWE-bench Verified for software engineering, Aider Polyglot for multi-language code editing, and MMMU for visual problem solving, the performance jumps in general aren’t as massive as from past GPT version upgrades.
Still, one promising improvement is in the hallucination rate, including for health-related conversations, which OpenAI said are dramatically lower than on o3 and GPT-4o.
Personalization
Claude users probably know the Styles setting, which lets them choose between Normal, Concise, Explanatory, and Formal modes, or define custom behaviors for the model to follow.”
Now, ChatGPT offers basically the same thing, with four personalities other than the Default (which is user-customizable). Here’s OpenAI’s official description of each new personality:
Cynic – Sarcastic and dry, delivers blunt help with wit. Often teases, but provides direct, practical answers when it matters.
Robot – Precise, efficient, and emotionless, delivering direct answers without extra words.
Listener – Warm and laid-back, reflecting your thoughts back with calm clarity and light wit.
Nerd – Playful and curious, explaining concepts clearly while celebrating knowledge and discovery.
Alongside the new personalities, OpenAI will also let users apply accent colors to the interface.
Google connectivity
OpenAI says that starting next week, Pro users will be able to connect their Google Calendar, Google Contacts and even their Gmail account, in order to get more personalized help directly on ChatGPT.
During the event, one demo included an OpenAI engineer using ChatGPT to help organize their schedule, with the model voluntarily helping them squeeze in a run between appointments.
OpenAI says the Google integration will roll out to other tiers in the future, but didn’t give a specific timeline.
Safety and health
OpenAI spent a good portion of the event showcasing ChatGPT’s new “safe completion” behavior. Rather than refusing to comply with a prompt, it will explain why it can’t proceed with the conversation.
If that sounds like a great way to help users looking to jailbreak the model and find ways around its safety limitations, I agree. It will be interesting to see the feedback of the trust and safety community on this one.
The company also spent time addressing how people use ChatGPT for medical advice. In addition to highlighting its lower hallucination rate in health-related conversations, they featured a conversation between CEO Sam Altman and an oncology patient and her husband, who spoke about how GPT-5 has helped them better understand her condition and treatment.
On that subject, while it is obviously not advisable to blindly trust what ChatGPT says about a serious health condition, the company highlighted the chatbot’s ability to help explain and present information in a responsible way, which is definitely an improvement and can be a net positive, since people will turn to the chatbot for such advice either way.
Development
This was easily the segment that OpenAI dedicated the most time to.
The company showcased tools and use cases that went from quickly creating a language teaching game, to an interactive demo that helped understand how planes stay airborne.
OpenAI also announced that developers will be able to customize how verbose GPT-5 should be, which is great to keep better control of API and token usage, and even invited Cursor CEO Michael Truell to prop up the new model’s coding capabilities and integration with what is currently one of the more famous AI-powered IDEs.
Developers will be able to access GPT-5, and two other variants (mini and nano) via the API, with a 256,000-token context window, up from 200,000 tokens from the o3 model. These are the API prices:
GPT-5: $1.25/million input tokens, $10/million output tokens;
GPT-5 mini: $0.25/million input tokens, $2/million output tokens;
GPT-5 nano: $0.05/million input tokens, $0.40/million output tokens.
What did you think of GPT-5? Let us know in the comments.
Limited Mac deals on Amazon