AI Pulse by Inblix

OpenAI drops o3 and o4-mini: Reasoning models that actually use tools

OpenAI Blog · Jul 14, 2026 · 2 min read · Read original article →

Curated by the Inblix editorial team


Featured image for article: OpenAI drops o3 and o4-mini: Reasoning models that actually use tools

OpenAI isn’t just releasing new models. With o3 and o4-mini, they’re fundamentally changing how their AI reasons about the world by giving it hands. These are the first reasoning models trained to agentically use every tool in ChatGPT—web search, Python, image generation, and file analysis—and decide for themselves when to reach for which one. That’s a real shift from models that just think to models that act.

The flagship o3 is now the company’s most powerful reasoning model, setting new records on Codeforces, SWE-bench, and MMMU. External experts found it makes 20 percent fewer major errors than o1 on difficult real-world tasks, with particular strength in programming, consulting, and generating novel scientific hypotheses. It’s especially sharp with visual inputs like charts and diagrams. The smaller o4-mini is optimized for speed and cost, yet manages to be the top benchmarked model on AIME 2024 and 2025. Give it a Python interpreter and it hits 99.5% on AIME 2025—a number that demands attention even if tool access changes the comparison.

Under the hood, both models were trained with large-scale reinforcement learning that taught them not just how to use tools, but when. OpenAI says they’re seeing the same “more compute equals better performance” trend in RL that they saw in GPT pretraining, pushing an order of magnitude further and still finding gains. Let o3 think longer, and its performance keeps climbing. That’s a validation of the scaling approach, just applied to reasoning time instead of training data.

There’s also an update for Pro subscribers: o3-pro dropped on June 10, offering an even more deliberative version designed for maximum reliability. Both o3 and o4-mini should feel more conversational than their predecessors, referencing memory and past chats. The tool-use training means these models can tackle multi-step problems that would have stumped earlier versions—not because the reasoning was bad, but because they couldn’t grab the right instrument for the job.

💡 Key Takeaways

  1. o3 and o4-mini are the first OpenAI reasoning models trained to autonomously decide when and how to use ChatGPT's full toolkit, marking a shift from pure reasoning to agentic behavior.
  2. External evaluators found o3 makes 20% fewer major errors than o1 on difficult real-world tasks, with standout performance in programming, consulting, and scientific hypothesis generation.
  3. o4-mini achieves 99.5% on AIME 2025 with Python access, demonstrating that tool-augmented reasoning can dramatically outperform standalone model benchmarks.
  4. OpenAI confirms the RL scaling trend continues—letting o3 think longer reliably improves performance, with no sign of diminishing returns at current compute levels.

Keep reading: See related articles below for more coverage on this topic.

Get smarter about AI

The sharpest AI news, curated daily. Delivered free to your inbox.

Learn more

Glossary terms

← Back to all articles