What's new in OpenAI - Announcements at Nov'2023 DevDay

GPT-4 Turbo, APIs getting cheaper and they are getting more capable

It's been 6 hours since the conference concluded and there's a lot of exciting news for us, developers!

Here're the highlights:

• GPT-4 Turbo API
  • Huge 128K context - the current GPT4 API has 8k and 32k flavours, the largest context available in commercial LLM so far is 100k in Anthropic's Claude 1/2
  • !!! 4 to 6 times cheaper than GPT-4 32K
    • $0.01 per 1k input tokens vs $0.06 in GPT-4 32K
    • $0.03 vs $0.12 for 1k output tokens
  • Knowledge cut-off is now April 2023
• GPT-4 Turbo with vision can now accept images along the text
  • Available via gpt-4-vision-preview
• Opening GPT-4 for fine-tuning
• GPT-3.5 Turbo also gets updates in new API versions
• New APIs versions (gpt-3.5-turbo-1006 and gpt-4-1106-preview)
  • Improved instruction following
    • 38% improvement on format following tasks such as generating JSON, XML and YAML for GPT-3.5 Turbo
  • Function calling with improved accuracy and multiple functions called in one request
  • JSON mode via response_format param
  • Reproducible outputs via seed param
  • logprobs param in Completion API (log probabilities for the most likely output tokens)
• Other services and features
  • Assistants API - build and integrate AI agents in apps. One of the features is persistent and infinitely long threads (remember those context size limits and the necessity to send entire conversations via Chat APIs)
  • Better quality Text-to-speech
  • dall-e-3 image generator
  • Model customisations for enterprise customers
  • Whisper v3 and Consistency Decoder

Prices are going down

Coming soon

The new changes are expected to be rolled out in the coming weeks and become available to the paying customers. Also looking forward to the new features becoming available in Azure!

Check out OpenAI's blog for more detail!