OpenAI launches o3 and o4-mini: Smarter and faster reasoning AI models built to think like agents

Less than a week after supercharging ChatGPT’s memory in a major leap toward AGI, OpenAI on Wednesday unveiled two new AI models, o3 and o4-mini, calling them the company’s “smartest and most capable models to date.” Both bring a new level of performance across reasoning tasks and tool integration, making them the most capable releases in the ChatGPT lineup so far.

These models aren’t just about improved output—they signal a shift in how AI works across tools. They can reason through problems by using web search, code execution, image interpretation, file reading, and image generation—all within a single session.

No need for manual switching or step-by-step prompting. The models handle everything on their own, moving between tools as needed.

Built to Think Like Agents

The most talked-about feature is what OpenAI refers to as “agentic” behavior. That means these models don’t just spit out answers—they reason through problems, moving between tools inside ChatGPT like search, code, and image analysis, as needed.

OpenAI said it best in its post on X:

“For the first time, our reasoning models can agentically use and combine every tool within ChatGPT, including web search, Python, image analysis, file interpretation, and image generation.”

Introducing OpenAI o3 and o4-mini—our smartest and most capable models to date.

For the first time, our reasoning models can agentically use and combine every tool within ChatGPT, including web search, Python, image analysis, file interpretation, and image generation. pic.twitter.com/rDaqV0x0wE

— OpenAI (@OpenAI) April 16, 2025

This means the models can now process a file, look something up, analyze a photo, and explain everything in one session, without needing users to guide them step-by-step. This kind of workflow makes AI more autonomous and useful, especially for people who rely on it for deep or complex tasks.

o3: The New Standard for Multimodal AI

Of the two, o3 is the flagship. It performs well across coding, math, science, and anything that requires multi-step logic. Based on posts from OpenAI and early testers on X, it’s outperforming other models in various benchmarks.

One major step forward is its visual reasoning. Unlike earlier versions, o3 doesn’t treat images as isolated inputs. It works with them in context, seeing, interpreting, and connecting visual information to written queries. That opens up new use cases in education, science, design, and research.

o4-mini: Smaller Size, Surprising Strength

o4-mini is built for speed, affordability, and efficiency. It’s smaller than o3 but still capable, and it’s priced for scale. OpenAI says it delivers strong performance and reasoning without the heavier usage limits, making it perfect for real-time support, high-frequency queries, or budget-conscious projects and developers who need to run a lot of queries without racking up big bills.

Despite its size, o4-mini is already proving to be more than just a compact version. It has been praised on X for holding its own in tasks like code generation, data parsing, and anything that benefits from fast response times. Think of it as the go-to for high-volume or real-time needs.

Access and Availability

Both models are now available to ChatGPT Plus, Pro, and Team users. o3, o4-mini, and o4-mini-high are replacing older models like o1 and o3-mini in the model selector. Enterprise and Edu users will get access in the next week. Rate limits won’t change for now.

Developers can use both models through the Chat Completions API and the newer Responses API. The API now supports features like reasoning summaries and token preservation, which can help with more consistent function calls. OpenAI also teased some upcoming API upgrades for developers.

Mixed Reactions Online

The AI crowd on X has been buzzing. Some users called o3 “a beast” and said it “kicks ass on every benchmark.” Others highlighted o4-mini as a surprising performer, punching above its weight.

That said, some users flagged concerns about the evaluation process. External assessors like METR reportedly only had two weeks to test the models before launch. That short window raised eyebrows among critics, calling for more independent review time before rollout.

What’s Next

The launch of o3 and o4-mini signals OpenAI’s ongoing push to improve what its models can do out of the box. With tool usage now deeply baked into the models, OpenAI is betting that AI systems don’t just need better answers—they need better ways to think.

Codex is also making a return, now with a command-line interface. That’s likely to please developers who’ve missed its coding capabilities.

As OpenAI keeps building, all eyes will be on how these models perform in the wild—and how the company handles transparency and safety questions moving forward. But if early results are anything to go by, o3 and o4-mini just raised the bar for what AI is expected to do.