Skip to content

New Stuff Of OpenAI GPT 4 Turbo Longer Context More Control Many More OpenAI DevDay Part 2

Jump to bottom
FurkanGozukara edited this page Oct 22, 2025 · 1 revision

New Stuff Of OpenAI - GPT-4 Turbo - Longer Context - More Control - Many More - OpenAI DevDay Part 2

Full tutorial link > https://www.youtube.com/watch?v=4mQXxk8FeEE

New Stuff Of OpenAI - GPT-4 Turbo - Longer Context - More Control - Many More - OpenAI DevDay Part 2

image Hits Patreon BuyMeACoffee Furkan Gözükara Medium Codio Furkan Gözükara Medium

YouTube Channel Furkan Gözükara LinkedIn Udemy Twitter Follow Furkan Gözükara

This video is Part 2 of #OpenAI #DevDay #GPT

Full Event : https://www.youtube.com/watch?v=U9mJuUkhUzk

Credits : https://www.youtube.com/watch?v=U9mJuUkhUzk

00:00:00 🚀 OpenAI is launching GPT-4 Turbo, a new model that addresses user feedback and offers significant improvements.

00:01:20 🎮 GPT-4 Turbo provides more control to developers, including JSON mode, better function calling, and reproducible outputs.

00:02:30 🌍 OpenAI is introducing retrieval in the platform, enabling users to access external knowledge, and updating the knowledge cutoff to April 2023.

00:03:41 📷 GPT-4 Turbo supports image inputs via the API, generating captions, classifications, and analysis. It also includes a natural-sounding text-to-speech model.

00:04:51 🗣️ OpenAI is releasing Whisper V3, an improved speech recognition model, and expanding customization options for models.

00:06:29 💼 OpenAI is doubling the tokens per minute for GPT-4 customers and introducing Copyright Shield to protect against legal claims related to copyright infringement.

00:07:33 💰 GPT-4 Turbo offers a significantly lower price, with a 3x reduction in prompt token cost and a 2x reduction in completion token cost compared to GPT-4.

Video Transcription

  • 00:00:00 So now on to the new stuff and we have got a  lot. First we're going to talk about a bunch of  

  • 00:00:08 improvements we've made and then we'll talk about  where we're headed next. Over the last year we  

  • 00:00:14 spent a lot of time talking to developers around  the world. We've heard a lot of your feedback.  

  • 00:00:20 It's really informed what we have to show you  today. Today we are launching a new model GPT-4  

  • 00:00:28 Turbo. GPT-4 Turbo will address many of the  things that you all have asked for. So let's  

  • 00:00:39 go through what's new. We've got six major things  to talk about for this part. Number one, context  

  • 00:00:46 length. A lot of people have tasks that require a  much longer context length. GPT-4 supported up to  

  • 00:00:53 8k and in some cases up to 32k context length. But  we know that isn't enough for many of you and what  

  • 00:00:59 you've heard. So let's get started. GPT-4 Turbo  supports up to 128,000 tokens of context. That's  

  • 00:01:11 300 pages of a standard book, 16 times longer than  our 8k context. And in addition to longer context  

  • 00:01:18 length, you'll notice that the model is much  more accurate over a long context. Number two,  

  • 00:01:26 more control. We've heard loud and clear that  developers need more control over the model's  

  • 00:01:31 responses and outputs. So we've addressed that  in a number of ways. We have a new feature called  

  • 00:01:37 JSON mode which ensures that the model will  respond with valid JSON. This has been a huge  

  • 00:01:43 developer request. It'll make calling APIs much  easier. The model is also much better at function  

  • 00:01:50 calling. You can now call many functions at once.  And it'll do better at following instructions in  

  • 00:01:55 general. We're also introducing a new feature  called reproducible outputs. You can pass a  

  • 00:02:02 seed parameter and it'll make the model return  consistent outputs. This, of course, gives you a  

  • 00:02:06 higher degree of control over model behavior. This  rolls out in beta today. And in the coming weeks,  

  • 00:02:17 we'll roll out a feature to let you view log probs  in the API. All right. Number three, better world  

  • 00:02:26 knowledge. You want these models to be able to  access better knowledge about the world. So do  

  • 00:02:30 we. So we're launching retrieval in the platform.  You can bring knowledge from outside documents or  

  • 00:02:36 databases into whatever you're building. We're  also updating the knowledge cutoff. We are  

  • 00:02:42 just as annoyed as all of you, probably more,  that GPT-4's knowledge about the world ended  

  • 00:02:46 in 2021. We will try to never let it get that  out of date again. GPT-4 Turbo has knowledge  

  • 00:02:52 about the world up to April of 2023. And we will  continue to improve that over time. Number four,  

  • 00:03:01 new modalities. Surprising no one, DALI 3, GPT-4  Turbo with vision, and the new text-to-speech  

  • 00:03:10 model are all going into the API today. We have  a handful of customers that have just started  

  • 00:03:21 using DALI 3 to programmatically generate images  and designs. Today, Coke is launching a campaign  

  • 00:03:29 that lets its customers generate Diwali cards  using DALI 3. And of course, our safety systems  

  • 00:03:34 help developers protect their applications  against misuse. Those tools are available in  

  • 00:03:38 the API. GPT-4 Turbo can now accept images as  inputs via the API. It can generate captions,  

  • 00:03:46 classifications, and analysis. For example, Be My  Eyes uses this technology to help people who are  

  • 00:03:52 blind or have low vision with their daily tasks  like identifying products in front of them. And  

  • 00:04:00 with our new text-to-speech model, you'll be  able to generate incredibly natural sounding  

  • 00:04:05 audio from text in the API with six preset  voices to choose from. I'll play an example.  

  • 00:04:12 Did you know that Alexander Graham Bell, the  eminent inventor, was enchanted by the world  

  • 00:04:16 of sounds? His ingenious mind led to the creation  of the graphophone, which etches sounds onto wax,  

  • 00:04:22 making voices whisper through time. This is much  more natural than anything else we've heard out  

  • 00:04:28 there. Voice can make apps more natural to  interact with and more accessible. It also  

  • 00:04:34 unlocks a lot of use cases like language learning  and voice assistance. Speaking of new modalities,  

  • 00:04:41 we're also releasing the next version of  our open-source speech recognition model,  

  • 00:04:45 Whisper V3, today, and it'll be coming soon  to the API. It features improved performance  

  • 00:04:50 across many languages, and we think you're  really going to like it. Okay, number five,  

  • 00:04:56 customization. We've been doing really well for  GPT-3.5 since we launched it a few months ago.  

  • 00:05:03 Starting today, we're going to expand that to the  16K version of the model. Also starting today,  

  • 00:05:10 we're inviting active fine-tuning users to apply  for the GPT-4 fine-tuning experimental access  

  • 00:05:15 program. The fine-tuning API is great for adapting  our models to achieve better performance in a wide  

  • 00:05:22 variety of applications with a relatively small  amount of data. But you may want a model to  

  • 00:05:27 learn a completely new knowledge domain. Or  to use a lot of proprietary data. So today,  

  • 00:05:32 we're launching a new program called Custom  Models. With Custom Models, our researchers will  

  • 00:05:38 work closely with the company to help them make a  great custom model, especially for them and their  

  • 00:05:44 use case using our tools. This includes modifying  every step of the model training process,  

  • 00:05:50 doing additional domain-specific pre-training,  a custom RL post-training process tailored  

  • 00:05:55 for a specific domain, and whatever else. We  won't be able to do this with the many companies  

  • 00:05:59 to start. It'll take a lot of work, and in the  interest of expectations, at least initially,  

  • 00:06:04 it won't be cheap. But if you're excited to  push things as far as they can currently go,  

  • 00:06:08 please get in touch with us, and we think we  can do something pretty great. Okay. And then  

  • 00:06:13 number six, higher rate limits. We're doubling the  tokens per minute for all of our established GPT-4  

  • 00:06:19 customers so that it's easier to do more. And  you'll be able to request changes to further rate  

  • 00:06:24 limits and quotas directly in your API account  settings. In addition to the these rate limits,  

  • 00:06:30 it's important to do everything we can do to  make you successful building on our platform.  

  • 00:06:37 So we're introducing Copyright Shield. Copyright  Shield means that we will step in and defend  

  • 00:06:42 our customers and pay the costs incurred if you  face legal claims around copyright infringement,  

  • 00:06:48 and this applies both to ChatGPT Enterprise and  the API. And let me be clear, this is a good time  

  • 00:06:55 to remind people, we do not train on data from the  API or ChatGPT Enterprise ever. All right. There's  

  • 00:07:04 actually one more developer request that's been  even bigger than all of these. And so I'd like  

  • 00:07:09 to talk about that now. And that's pricing. GPT-4  Turbo is the industry leading model. It delivers  

  • 00:07:21 a lot of improvements that we just covered, and  it's a smarter model than GPT-4. We've heard from  

  • 00:07:28 developers that there are a lot of things that  they want to build, but GPT-4 just costs too  

  • 00:07:33 much. They've told us that if we could decrease  the cost by 20, 25 percent, that would be great,  

  • 00:07:39 a huge leap forward. I'm super excited to  announce that we worked really hard on this,  

  • 00:07:45 and GPT-4 Turbo, a better model, is considerably  cheaper than GPT-4 by a factor of 3x for prompt  

  • 00:07:53 tokens. And 2x for completion tokens starting  today. So the new pricing is 1 cent per 1,000  

  • 00:08:11 prompt tokens and 3 cents per 1,000 completion  tokens. For most customers, that will lead to  

  • 00:08:16 a blended rate more than 2.75 times cheaper  to use for GPT-4 Turbo than GPT-4. We worked  

  • 00:08:22 super hard to make this happen. We hope you're  as excited about it as we are. So, we've decided  

  • 00:08:32 to prioritize price first because we had to choose  one or the other, but we're going to work on speed  

  • 00:08:36 next. We know that speed is important, too. Soon,  you will notice GPT-4 Turbo becoming a lot faster.  

  • 00:08:44 We're also decreasing the cost of GPT-3.5 Turbo  16K. Also, input tokens are 3x less and output  

  • 00:08:52 tokens are 2x less, which means that GPT-3.5 16K  is now cheaper than the previous GPT-3.5 4K salah.  

  • 00:08:59 model. Running a fine-tuned GPT 3 .5 Turbo 16K  version is also cheaper than the old fine-tuned 4K

Clone this wiki locally