OpenAI marked Day 9 of its "12 Days of Shipmas" event with a substantial package of updates aimed at developers and businesses building AI applications. The announcements include the general availability of the o1 model, new developer tools, and significant price reductions for audio processing. Let's break down the most significant changes that matter for businesses and developers.
1. o1 Model Available via API
The star of today's announcement is the general availability of o1 in OpenAI's API for tier 5 developers. The numbers tell an impressive story - o1 achieved a 79.2% success rate on AIME 2024 mathematical problems, nearly doubling its predecessor's performance. For developers working on coding applications, o1 hit a 76.6% success rate on LiveCodeBench, up from 52.3%. Perhaps most importantly for businesses watching their bottom line, the model uses 60% fewer thinking tokens than its preview version, translating to faster responses and lower costs.
2. More Developer Controls
OpenAI has introduced a new "reasoning_effort" parameter that lets developers fine-tune how much time the model spends processing before responding. The update also brings function calling capabilities for connecting with external APIs and data sources, plus structured output features that ensure responses follow specific JSON schemas. A new "developer messages" feature gives more precise control over the model's behavior and output style. You can try many of these features in the updated developer playground.
3. Major Price Cuts for Audio Processing
OpenAI has slashed GPT-4o audio token prices by 60%. The new rates are $40 per million input tokens and $80 per million output tokens. They've also launched GPT-4o mini for audio processing at just one-tenth of the previous rates, opening up new possibilities for businesses exploring voice-enabled AI applications.
4. WebRTC Support in Realtime API
Voice application developers can now take advantage of WebRTC support in the Realtime API. This simplifies the creation of real-time voice applications across different platforms, making it easier to build robust voice assistants and interactive customer support systems. The implementation requires minimal code, reducing the technical barrier for teams looking to add voice capabilities.
5. Preference Fine-Tuning
The new Preference Fine-Tuning feature uses Direct Preference Optimization to help customize models based on specific preferences. Early adopter Rogo AI reported impressive results, boosting their financial analysis model's accuracy from 75% to over 80%. This feature is particularly valuable for applications where tone and style matter as much as accuracy.
6. New SDKs for GO and Java
Enterprise development teams now have more options with new beta SDKs for Go and Java, complementing the existing Python, Node.js, and .NET libraries. This expansion makes it easier for organizations to integrate OpenAI's technology into their existing systems, regardless of their preferred programming language.
7. Vision Capabilities in o1
The o1 model now includes vision capabilities, allowing it to process and reason about images. This addition opens up new possibilities for applications in scientific research, manufacturing, and coding where visual inputs are crucial.
8. API Access to o1 pro is Coming
OpenAI confirmed that the team is working on accessing o1 pro via the APIs. In ChatGPT, Pro mode uses the o1 pro model and currently costs $200/month for unlimited use. It is unclear how this will translate to API pricing.
9. Developer Quality-of-Life Updates
OpenAI also rolled out smaller but impactful updates to enhance the developer experience. For example, the process of obtaining an API key has been streamlined, and they've made API changes to make it easier to use function calling and guardrails. They've also made all sessions (30 videos) from their 2024 Dev Days available on YouTube.
These updates represent significant steps forward in making AI technology more accessible and practical for real-world applications. For businesses considering AI integration, the combination of improved performance, lower costs, and easier implementation could make now the right time to explore these new capabilities.
To round things out, the OpenAI API team and presenters hosted an AMA on the official OpenAI Developer Forum, addressing questions about the new updates and sharing insights.
So overall, Day 9 of OpenAI’s Shipmas was a big win for developers building AI-powered applications. Whether it’s real-time tools with WebRTC, preference fine-tuning for tailored responses, or the production-ready o1 model, the improvements set a strong foundation for the next generation of AI apps.