OpenAI DevDay 2025 - What Hit What Missed

OpenAI’s DevDay 2025 keynote centered on four practical product moves: apps inside ChatGPT, a new Agent Kit for building agentic systems, Codex becoming generally available, and fresh API/model updates—especially for video and real-time voice. The biggest “why it matters” thread running through all of them is speed-to-build: fewer barriers for turning ideas into working AI experiences, whether that’s a workflow app, an internal agent, or a voice-enabled application.

First came “apps inside ChatGPT,” positioned as an apps SDK in preview, with a ChatGPT app-store rollout later. The pitch is straightforward: developers can build app experiences that run within ChatGPT, letting users trigger third-party capabilities without leaving the chat. Examples included using Canva to convert an outline into a deck and using Coursera to watch a course inside ChatGPT and then chat with the video content. The unresolved question is discovery—how a new app gets found and adopted when users can already access established partners like Canva, Expedia, and Zillow through their own channels and generic search.

Second, OpenAI introduced Agent Kit, a bundled toolkit for building, deploying, and optimizing agents. It includes Agent Builder, a drag-and-drop visual canvas for assembling agent workflows, plus a connector registry intended to standardize how tools and services connect—potentially including third-party connectors and MCPs. Agent Builder drew the most attention, with demos that leaned toward workflow-style agent apps but also suggested the system can support more complex, recursive agent behavior. The target audience appears to be teams that want agentic functionality without a dedicated AI engineering staff, enabling internal sharing of agent experiences. Notably, the discussion didn’t emphasize A2A communication with other agents or multi-agent systems, though the connector registry hints at broader integration.

OpenAI also paired Agent Kit with additional “guardrails” and agent evaluation (evals) support, aiming to reduce the pain of testing and measuring agent behavior—an area where many teams struggle with reliability and iteration.

Third, Codex was announced as generally available, with usage momentum highlighted: daily usage reportedly grew 10x since early August, and the first three weeks reportedly exceeded 40 trillion tokens. The emphasis here was less on flashy new features and more on scaling and expanding deployment, including integrations such as Codex in Slack and upcoming admin tooling. A demo also encouraged voice-based usage.

Finally, API updates focused on Sora 2 and real-time voice. Sora 2 is coming via API with controls like aspect ratio and video length, plus the ability to inject images as context. Pricing was framed as a lever for market competition: basic Sora at 10 cents per second, and Sora 2 Pro at 30–50 cents per second depending on resolution. OpenAI also announced a new “GPT real-time mini,” described as 70% cheaper, with lower costs for both audio input and audio output. Combined with falling image costs, the economics look more favorable for real-time voice applications that rely on function calling and related capabilities.

Overall, the keynote landed as a more polished rollout than earlier launches, with a clear emphasis on developer tooling, integration, and cost reductions that could make agentic and real-time experiences easier to build and deploy.