OpenAI Releases O1 Reasoning Model via API

OpenAI Introduces o1 Reasoning Model to its API

OpenAI is now making its o1, a sophisticated “reasoning” AI model, available through its API. However, initial access is limited to a select group of developers.

Tier 5 Access Requirements

The rollout of o1 begins Tuesday, and is restricted to developers within OpenAI’s “tier 5” usage category. Qualification for this tier necessitates a minimum expenditure of $1,000 with OpenAI.

Furthermore, accounts must be at least 30 days old since the first successful payment was processed.

o1 Replaces o1-preview

This new release of o1 effectively supersedes the previously available o1-preview model within the API.

The Advantage of Reasoning Models

Unlike many conventional AI systems, reasoning models such as o1 possess the ability to self-verify information. This intrinsic fact-checking capability helps mitigate common errors encountered in other models.

It’s important to note that this enhanced accuracy often comes at the cost of processing time, as these models require more deliberation to formulate responses.

Cost Considerations

Reasoning models are computationally intensive, resulting in higher operational costs. OpenAI’s pricing structure for o1 reflects this, at $15 per approximately 750,000 words analyzed and $60 for every 750,000 words generated.

This represents a sixfold increase in cost compared to OpenAI’s latest “non-reasoning” model, GPT-4o.

Enhanced Customization Options

The o1 model integrated into the OpenAI API offers significantly greater customization than its o1-preview predecessor.

New features include function calling, enabling connection to external data sources, developer messages for controlling tone and style, and image analysis capabilities.

Additionally, the “reasoning_effort” API parameter allows developers to fine-tune the model’s deliberation time before responding to queries.

New Post-Trained Version

OpenAI has released a “new post-trained” version of o1 for both the API and, soon, its ChatGPT chatbot platform.

Designated as “o1-2024-12-17,” this iteration incorporates improvements based on user feedback, enhancing overall model behavior.

Incremental Rollout

OpenAI is implementing access to o1 gradually. This phased approach allows the company to expand access to additional usage tiers and optimize rate limits.

Improved Response Quality

According to OpenAI, the latest o1 model delivers “more comprehensive and accurate responses,” particularly in areas like programming and business-related inquiries.

It also exhibits a reduced tendency to inappropriately decline requests.

Updates to Realtime API

Alongside o1, OpenAI announced new versions of its GPT-4o and GPT-4o mini models for the Realtime API. This API is designed for building applications requiring low-latency, AI-powered voice responses.

The new models (“gpt-4o-realtime-preview-2024-12-17” and “gpt-4o-mini-realtime-preview-2024-12-17”) offer improved data efficiency, enhanced reliability, and reduced usage costs.

Realtime API Enhancements

The Realtime API, currently in beta, has gained new functionalities, including concurrent out-of-band responses. This allows for background processes, such as content moderation, to operate without disrupting user interactions.

The API now also supports WebRTC, an open standard for real-time voice applications across various platforms.

WebRTC Integration and New Hire

Notably, OpenAI recently hired Justin Uberti, the creator of WebRTC, in early December.

OpenAI stated that its WebRTC integration is engineered to ensure seamless and responsive interactions, even under fluctuating network conditions. It manages audio encoding, streaming, noise suppression, and congestion control.

Additional Updates

Finally, OpenAI introduced preference fine-tuning to its fine-tuning API. This technique compares model responses to refine its ability to identify preferred answers.

The company also launched an “early access” beta for official software developer kits in Go and Java.