OpenAI is moving its developer platform beyond o1-preview by giving API users access to the full o1 model. The change matters because it brings back several API features developers rely on when building production chatbots, assistants, voice tools, and structured AI workflows.
The rollout is not just a model swap. OpenAI also announced changes for real-time audio, fine-tuning, function calling, structured outputs, and developer tooling, making this a broader update to how outside builders can use its models inside their own applications.
What changes with full o1 API access
Developers using OpenAI through its long-running API will now be able to work with the company’s latest full o1 model rather than the months-old o1-preview. OpenAI said the new model restores many of the “core features you’ve been missing” from the API during the preview period.
One returning capability is support for developer messages. These messages let builders steer the behavior of a chatbot or assistant for a specific context, such as defining the assistant’s role for a particular professional audience.
The update also introduces a reasoning effort parameter. In plain terms, this gives developers a way to tell the API how long it should think about a request. That can help reserve deeper reasoning for harder questions while avoiding unnecessary time and cost on simpler tasks.
Visual input is part of the upgrade as well. API users can provide visual information, including scans of documents, as input to the model. That expands the kinds of workflows developers can build, especially where text alone is not enough to understand the task.
Better function calling and structured output
OpenAI also highlighted improvements to internal function calling. In these workflows, the model can decide when it is appropriate to call functions that outside developers have already written, then use those functions to help generate an answer.
This is important for applications that need the model to do more than produce conversational text. Function calling can connect the model to developer-defined actions, calculations, lookups, or other programmatic steps, depending on how the application is designed.
Structured outputs are also getting more accurate, according to OpenAI. These outputs use a JSON schema to return information in a format specified by the developer. For teams that need predictable data from model responses, stronger adherence to a schema can reduce cleanup work and make API responses easier to integrate into software systems.
Together, better function calling and structured outputs point toward a more controlled developer experience. The model is not only generating answers; it is being shaped to work inside applications that expect specific formats and reliable behavior.
Faster reasoning and lower-cost audio
OpenAI said the o1 model uses 60 percent fewer thinking tokens than o1-preview. That should make results faster and cheaper compared with the preview model, while still supporting deeper reasoning when developers choose to use it.
The company also said o1 is roughly 25 to 35 percentage points more accurate on benchmarks like LiveBench and the AIME. Those benchmark claims are presented by OpenAI as part of the case for moving developers from o1-preview to the full o1 model.
Access will start rolling out to tier 5 development customers today. API access to the $200/month o1 Pro model is not available in this update yet; OpenAI only said it is “coming soon.”
For real-time voice products, OpenAI is adding full access to WebRTC support. That comes alongside existing support for the WebSocket audio standard and is meant to make it simpler to build OpenAI audio interfaces into third-party applications.
According to the company, WebRTC support can reduce the work required for some audio interfaces from roughly 250 lines of code to about a dozen. OpenAI said it will also release simple WebRTC code that can be used on a plug-and-play basis in simple devices.
The examples OpenAI gave include toy reindeer, smart glasses, and cameras that want to use context-aware AI assistants. To encourage those uses, OpenAI said it was cutting the cost of o1 audio tokens for API developers by 60 percent and the cost of 4o mini tokens by a full 90 percent.
Fine-tuning gets a preference-based option
Developers who fine-tune AI models are also getting a new method called direct preference optimization. This changes the kind of training signal developers can provide when they want a model to better match a target behavior.
Under supervised fine tuning, model makers provide examples of the exact input and output pairs they want the model to learn from. With direct preference optimization, they can instead provide two separate responses and mark one as preferred over the other.
OpenAI said its fine-tuning process can then learn the difference between the preferred and non-preferred answers. The company said this can include detecting changes in verbosity, formatting and style guidelines, or the helpfulness/creativity level of responses.
That approach may be useful when the desired improvement is easier to express as a preference than as a fully written ideal answer. Instead of prescribing every output, developers can show the model which response better fits the intended direction.
New SDKs broaden developer support
The update also includes new SDKs for programmers who write in Go or Java. OpenAI said those SDKs will let developers in those languages connect to the OpenAI API.
That matters because API adoption often depends on how cleanly a service fits into the languages and workflows teams already use. With new SDKs, OpenAI is giving more developers a direct path into its platform without requiring them to build every connection layer themselves.
Overall, the announcement positions the full o1 API as a more capable foundation for developer applications than o1-preview. The main themes are control, speed, cost management, real-time interaction, and better ways to shape model behavior for specific products.