OpenAI announced yesterday that they will be announcing new products and features every day for the upcoming 12 days. Today was the first day of these announcements and here's what Sam(-ta Claus) and his band of elves have in the bag for users.
ChatGPT now has a new subscription tier – ChatGPT Pro. It's a new plan that costs about $200 per month. That's a steep difference from the existing $20 per month individual subscription for ChatGPT Plus, which will still be available.
With ChatGPT Pro, users will get unlimited access to OpenAI o1 (and o1-mini), 4o, as well as Advanced Voice Mode. ChatGPT Plus users get access to these models with certain limits:
- o1: 50 messages per week
- o1 mini: 50 messages per day
- GPT-4o: 80 messages every 3 hours
- Advanced Voice: Fixed daily limit (which can change depending on usage; user speculation puts it close to around 30 minutes)
- GPT-4o mini: No limit
With the ChatGPT Pro subscription, none of these limits are in place for o1, o1-mini, and GPT-4o.
ChatGPT Pro users will also get access to the 128K context window, which was previously only available with the Enterprise plan or through API. Plus and Teams users get a 32K context window whereas Free users get a 8K context window.
Sam Altman also said that a few new things will also be added to this subscription as they release that stuff during the next 12 days.
Also, note that o1 doesn't carry the words "preview" with it anymore. It's because the reasoning model is now out of preview. The model, in full version, is now smarter and faster than o1-preview.
In my testing of the o1-preview, I found that it sometimes has a tendency to think unnecessarily long even for the simplest problems. That was also the experience for most users. Altman says that by incorporating this feedback, they have improved thinking time for o1 for simple problems. The model is also smarter and makes major mistakes about 34% less often than o1-preview. As demonstrated, o1 performs considerably better than o1 preview on Competition Math, Coding, and PhD level science questions.
The company also says that o1 would also be better for everyday use cases now and not just math, science, or coding problems.
o1 is also now multimodal and can take image input. However, they're still working on adding support for file uploads, web browsing, and data analysis to o1.
For ChatGPT Pro users, the company is also introducing the o1 Pro mode. In the Pro mode, o1 uses more compute power to find solutions for the hardest problems. It does show improvements in accuracy than o1 in competition Math and PhD-level science questions.
Notably, Noam Brown, a researcher at OpenAI, previously said that they aim for o1 to think for hours, days, and even weeks instead of merely seconds. While the o1 pro mode doesn't seem to be there yet, it can think considerably longer than the o1. The interface will display a progress bar and you'll even get an in-app notification once the model is done thinking in case you want to switch to another tab.
o1 in pro mode is also more reliable than o1 and significantly more so than o1-preview as demonstrated in the 4/4 reliability test. A 4/4 test only considers a model as having passed a problem when it gets the answer right all 4 times.
Currently, there's no news about the usage limit for o1 Pro mode. ChatGPT Pro and the o1 full version are available starting today.
That was about it from today's announcements from OpenAI. Let's see what the next 11 days have in store for us. There are speculations of OpenAI releasing their video model Sora and possibly even GPT-5.
Member discussion