OpenAI supercharges agents with MCP, image gen and reasoning tools

OpenAI loaded its Responses API with a fresh batch of built-in tools this morning, and the move says a lot about where agentic applications are heading. The headline here is native support for remote Model Context Protocol servers. That single integration plugs GPT-4o, GPT-4.1, and the o-series reasoning models directly into a growing universe of third-party services—from Stripe and Shopify to HubSpot and Plaid. Developers can now wire their agents into the tools their users already live in with just a few lines of code. OpenAI didn’t just ship the feature; it also joined the MCP steering committee, signaling it’s serious about this open standard becoming the connective tissue for the agent economy.

There’s more under the hood for reasoning models specifically. The o3 and o4-mini models can now call tools and functions directly within their chain-of-thought in the Responses API. That’s a big efficiency unlock. Instead of losing the thread between reasoning steps and tool calls, the model preserves its reasoning tokens across requests. The practical upshot for developers is better intelligence on complex, multi-step tasks and noticeably lower latency. OpenAI says this tighter integration boosted performance on brutal benchmarks like Humanity’s Last Exam.

The API is also getting tools that make agents feel less like text predictors and more like actual digital assistants. A new Code Interpreter tool lets models do data analysis, tackle complex math, and even “think with images” by manipulating visual data step-by-step. The gpt-image-1 tool brings image generation directly into the API with real-time streaming previews and multi-turn editing, meaning an agent can refine a generated image through conversation rather than starting from scratch. File search now works with the reasoning models too, letting agents pull relevant information from a user’s own document troves.

Enterprise reliability got some attention as well. A new background mode handles long-running tasks asynchronously, so agents don’t time out when the work gets heavy. There’s also support for reasoning summaries and encrypted reasoning items, which should ease privacy and compliance headaches for larger deployments. Since the API’s debut in March, hundreds of thousands of developers have pushed trillions of tokens through it, powering tools like Zencoder’s coding agent and MagicSchool AI’s education assistant. With these additions, the barrier to building a genuinely useful, multi-step agent just dropped a few notches lower.