Reasoning models, which dedicate time to thoroughly thinking through prompts, are at the front of the generative AI pack. OpenAI announced on March 20 o1-pro, a heavy-duty version of one of its reasoning models.
Developers on any of OpenAI’s paid tiers can find o1-pro in the OpenAI API starting today. Those with higher usage tiers will have higher request and batch queue limits.
The price reflects greater compute power: o1-pro costs $150 for 1 million input tokens and $600 for 1 million output tokens. Compare that to $15.00 for 1 million tokens and $60 per 1 million output tokens when using the base model of OpenAI o1.
Patrocinado por TecnoArtesanos ¿Tu empresa ya está usando IA? Automatizamos procesos, integramos asistentes inteligentes y conectamos tus sistemas. Descubre cómo →See: OpenAI Recap: o3 Model Wraps 12 Days of Announcements
What capabilities does $600 per one million tokens unlock?
The price point makes OpenAI o1-pro one of the priciest AI models. It takes the crown from API access to the Silicon Valley darling’s own GPT-4.5. (GPT-4.5 costs $75 per million input tokens and $150 per million output tokens.)
What do you get for that cash? OpenAI is probably hoping for o1-pro to bring in a deep-pocketed handful of researchers, engineers and other professionals using generative AI for science, medicine, or technology. While o1-pro is optimized for reasoning tasks, OpenAI offers separate models for transcription or moderation. A smaller, cheaper model can probably meet users’ needs for those.
For the price of o1-pro you get a 200,000 context window and 100,000 max output tokens.
OpenAI said o1-pro can interpret data from machine vision, with text and image input. However, it only produces text output. In the API it also supports:
- Function calling.
- Structured Outputs, which ensure the model’s response will conform to the developer’s JSON Schema.
- Integration with the Responses API, which can be used to create AI agents that can interact with the web in various ways, including running searches.
- Integration with the Batch API, which allows for asynchronous requests, providing lower costs and higher rate limits for jobs with a 24-hour turnaround time.
OpenAI has not publicly specified the exact knowledge cutoff date for o1-pro, but previous models had real-world knowledge of up to late 2023.
AI companies compete to dominate the “reasoning model” landscape
OpenAI introduced o1 as “Strawberry” in September 2024. It competes with DeepSeek’s R1, Anthropic’s Claude Sonnet 3.7, Grok 3, Google’s Gemini 2.0, and other reasoning models. Meta is working on “theory of mind reasoning,” but that’s a slightly different matter: Meta says it is an adversarial project for evaluating advanced AI models.
