Microsoft has rapidly incorporated OpenAI’s just-announced o3 and o4-mini artificial intelligence models into its core cloud and developer platforms, making the advanced reasoning capabilities available through Azure and GitHub. This quick integration underscores Microsoft’s strategy of embedding OpenAI’s latest technology across its services, offering users access to models characterized by improved performance and what the company calls “early agentic behavior.”
Unlike previous generations, o3 and o4-mini can autonomously determine and initiate the use of internal tools—such as web browsing, code execution, or file analysis—to address user requests without explicit step-by-step instructions.
This “agentic” approach represents a shift towards AI systems that can plan and act independently. These models also feature vision processing capabilities, enabling them to interpret images, and support for advanced functionalities like function calling, structured outputs, and long-context handling up to 200,000 tokens.
The rollout comes after OpenAI released the API only GPT-4.1 series of models and activated a “recall” memory feature in ChatGPT earlier in April, allowing context persistence across sessions. Alongside the launch of the o3 and o4-mini models, OpenAI also released Codex CLI, a free, open source tool for model-agnostic AI coding that lets developers use AI directly in the terminal.
Agentic AI Lands on Microsoft Platforms
Microsoft confirmed the models are live within the Azure OpenAI Service via Azure AI Foundry, initially available in the East US2 and Sweden Central Azure regions. Alongside these, new audio models (GPT-4o-Transcribe variants and Mini-TTS) also debuted on Azure Foundry in East US2. Azure pricing reflects the models’ different positionings: o3 costs $10/million input tokens ($2.50/M for vision) and $40/million output tokens, while the efficient o4-mini is $1.10/million input tokens ($0.275/M for vision) and $4.40/million output tokens, according to OpenAI.
Developer Tools and Access Controls
On the developer front, GitHub is rolling out public previews for the new models on GitHub Copilot and GitHub Models. GitHub frames o3 as being “ideal for deep coding workflows and complex technical problem solving,” with o4-mini combining “low latency with high-quality output, full tools support, and multimodal inputs.” Access is tiered: o4-mini reaches all paid Copilot plans, while o3 is limited to Enterprise and subscribers of the new Pro+ plan.
Selection occurs via model pickers in Visual Studio Code or GitHub Copilot Chat online once the rollout completes, although Copilot Enterprise requires admin enablement. GitHub also highlighted its data policies, assuring users of a zero data retention agreement with OpenAI and default non-training on business data for the Copilot integration. Within GitHub Models, developers can benchmark these newcomers against models from Meta, Cohere, Microsoft, and others in a shared playground.
Capabilities, Cost Efficiency, and Context
Benchmark data released by OpenAI suggests performance gains over the older o1 and o3-mini models in reasoning, coding, and multimodal tasks. While o3 often leads in complex operations, o4-mini demonstrates strong capabilities, particularly notable given its lower operational cost – a factor reflected in its Azure pricing.
This positions o4-mini as a potentially cost-effective option for many applications. Beyond Azure and GitHub, OpenAI is also broadening access through ChatGPT itself; Free users can sample o4-mini, while Enterprise and Edu users gain full access shortly, with an ‘o3-pro’ variant anticipated for the Pro tier.
Safety Dialogue Continues
This rapid deployment follows reports and discussions surrounding AI safety practices. OpenAI used a significantly shortened internal safety testing period for o3 and OpenAI’s update to its Preparedness Framework potentially allowing adjustments to safety protocols based on competitor actions.
A former employee quoted expressed concern over testing intermediate model checkpoints, stating, “It’s bad practice to release a model which is different from the one you evaluated.”
In contrast, Microsoft’s announcement emphasized that o3 and o4-mini offer “significant improvements on quality and safety” and feature “the next level of safety improvements within the o-series,” attributed to a “deliberative alignment” training strategy.
This reflects the tension between rapid advancement and safety considerations within the AI industry, contrasting with different approaches like Google DeepMind’s proposed global framework and Anthropic’s interpretability tools. Microsoft had previously integrated the o3-mini-high model into its free Copilot tier in March, indicating a continued push to integrate OpenAI’s reasoning line.
