OPENAI provides developers with models and many tools to develop AI agents capable of taking actions independently.
2025 will definitely be the year of agentic AI. From Google Cloud to Microsoft via AWS or Oracle, the main providers of Cloud and AI provide their customers with tailor -made tools to develop their own agents. Openai is no exception. The San Francisco start-up currently offers the most advanced tools and models to create its agents. Models, SDKs, tools… Here is the OpenAi starter pack available to developers.
4 reasoning models
It is the agent’s brain. The new reasoning models based on COT (Chain-OF-Thought) are to be favored for agent developments. By reasoning in stages, the model has a better understanding of the prompt several stages to be achieved. OPENAI offers, in April 2025, four major models of reasoning: O1, O3-mini, 01-mini and O1-Pro.
Models Complexity of the task to be made Latency PRIX An input ($ / 1M tokens) Prix en output ($/1M tokens) o1 average average 15 60 o1 mini simple weak 1,1 4,4 o1-pro advanced important 150 600 O3-Mini simple weak 1,1 4,4The O1 models were the first to be presented by Openai. The O3 family then arrived with a competitive performance / price ratio. For simple tasks, choose a small model with a low latency: O1-mini or O3-mini. With a preference for O3-Mini, the most recent model. For more complex tasks, O1 remains a very good choice with a higher pricing. Finally, for the most complex cases, O1-Pro is the ultimate weapon. It should be used sparingly because of its prohibitive price.
Des modes the text-to-speech it’s speech-to-text
To complete the reasoning models, Openai recently announced two new models dedicated to the voice (three in reality with the mini version). These allow vocal agents to be developed. For the transcription part, gpt-4o-transcribe Allows you to transcribe with great precision (greater than Whisper) of audio recordings. Finally its pendant part of reduced size GPT-4o-mini-Transcribe performs exactly the same function with reduced latency. The latter is therefore to be favored for aging configurations or an almost instantaneous vocal treatment is necessary. Finally, GPT-4o-Mini-Tts Allows you to generate complete vocal sequences with personalized diction.
Responsients: An API increased with agent tools
Responses is a specially designed API for agentic AI. It will replace the current OpenAi assistants API in 2026. The API makes it possible to use the OPENAI models by coupling them, if necessary, to the agental tools developed by the start-up. Currently, three tools are available.
- Web search To use web search capacities and obtain up -to -date information
- File search To give a Knowledge Base with context to a model
- Computer use To automate tasks in a web browser, on the principle of Operator
The tools can be used with models separately or jointly (example: Web Search and File Search). The API obviously supports multimodal entrances and the multitude.
SDK agent: orchestrate and monitor his agents
Finally to orchestrate everything, Openai now has a complete SDK developed in Python. The latter allows you to create, monitor and secure agents. Its installation (PIP Install Openai-Agents) and its use are thought of as the entire OPENAI ecosystem: simple and intuitive. In detail, SDK agent allows you to create three types of objects: agents, handoffs and guardrails.
THE agents are simply defined by a name and instructions. This is the simplest step. Example of the definition of two agents:
from agents import Agent math_tutor_agent= Agent( name="Math Tutor", instructions="You provide help with math problems. Explain your reasoning at each step and include examples", ) history_tutor_agent = Agent( name="History Tutor", handoff_description="Specialist agent for historical questions", instructions="You provide assistance with historical queries. Explain important events and context clearly.", )
More complex, the handoffs (without equivalent term in French) designate, by simplifying, the agent’s orchestration process. The mechanism allows an agent to transfer control of a task or a conversation to another more specialized agent or better equipped to process a specific request. Handoffs work as an intelligent relay system. The agent can thus decide at any time which new agent choose to progress in his task.
Example :
triage_agent = Agent( name="Triage Agent", instructions="You determine which agent to use based on the user's homework question", handoffs=[history_tutor_agent, math_tutor_agent] )
Finally the guardrails Simply make it possible to create filters to limit unwanted requests in input. They complement internal guardrails to the OpenAi models. Guardrails are, in reality, a different type of agent responsible for verifying the correct validation (or invalidation) of one or more conditions.
Example of the definition of a gardage:
guardrail_agent = Agent( name="Guardrail check", instructions="Check if the user is asking about homework.", output_type=HomeworkOutput, )
In summary, with its reasoning models adapted to different levels of complexity, its advanced vocal capacities via the text-to-speech and Speech-to-Text models, its API Responses enriched with agent tools, and its SDK facilitating the orchestration and securing of agents, Openai now provides all essential components to build intelligent and autonomous agents.
