At its first ever DevDay event, artificial intelligence research company OpenAI unveiled a slate of major upgrades and new capabilities for its AI platform. The announcements centered around more advanced language models, new multimodal features, and tools to customize conversational AI experiences.
New GPT-4 Turbo Language Model
The headliner was GPT-4 Turbo, OpenAI’s latest natural language model that improves upon the previous GPT-4 in several key ways:
- Offers a 128,000 token context window, 16x larger than GPT-4’s 8,000 token limit. This enables much longer conversations and workflows.
- Contains knowledge of world events up to April 2023. Previous models cutoff in 2021.
- Priced at just $0.01 per 1,000 prompt tokens and $0.03 per 1,000 completion tokens, making it 3x and 2x cheaper than GPT-4 respectively.
- Enhanced function calling abilities, allowing developers to invoke multiple functions in a single request and always returning valid JSON responses.
The Turbo model is available in preview now via the API, with plans to release a production version later this year.
Cheaper, More Capable GPT-3.5 Turbo
OpenAI also announced upgrades to its mid-sized GPT-3.5 language model:
- The 16,000 token version is now cheaper than the previous model, priced at just $0.001 per 1,000 input tokens and $0.002 per 1,000 output tokens.
- Fine-tuning this 16K model is also now up to 75% cheaper.
- It offers the same improved function calling abilities as GPT-4 Turbo.
New Assistants API
For developers building conversational apps, OpenAI launched a new Assistants API that makes it easy to create customized AI agents. The API includes:
- Persistent threads to maintain long conversations without context limits.
- Built-in tools like code execution, retrieval and enhanced function calling.
- A platform playground to prototype assistants without writing code.
Multimodal Capabilities Added
OpenAI brought new multimodal features to its models:
- GPT-4 Turbo can now process visual inputs to generate captions, classify images and more.
- The DALL-E 3 image generation model can also be used directly in apps via the Image API.
- A new text-to-speech model offers natural voice generation with 6 voice options.
Customizable GPTs in ChatGPT
Finally, users can now create customized “GPTs” within ChatGPT that combine instructions, data and capabilities tailored to specific use cases. GPTs will also be shareable via a forthcoming GPT Store.
The series of upgrades aims to provide developers more advanced AI building blocks for creating innovative applications across industries. OpenAI is looking to continuously iterate and expand capabilities over time.