OpenAI Dev Day 2025: Everything Announced
Apps Within ChatGPT
OpenAI is turning ChatGPT into a hub. The new Apps feature lets you access external services directly inside conversations. No context switching. No copy-paste workflows.
Want a Spotify playlist? ChatGPT generates it and creates it directly in your account. House hunting? Query Zillow and browse listings without leaving the chat. The initial partners include Canva, Expedia, and several others, but the pattern is clear: OpenAI wants ChatGPT to be the interface for the web.
Agent Kit: Visual Agent Building
The biggest announcement is Agent Kit. Think n8n or Zapier, but purpose-built for AI agents. It is a no-code platform for building agent workflows.
You still need to design the logic. Conditional branches, tool selection, and orchestration all require thought. But the interface removes the boilerplate. You can wire together file search, MCP integrations, and custom agents through a visual canvas. Route conversations down different paths based on user intent. Add tools where needed.
For developers who have been duct-taping agent frameworks together, this consolidates the stack.
Sora 2 Hits the API
Video generation is now programmable. Sora 2 and Sora 2 Pro are available via API, opening the door to automated video pipelines.
The integration is straightforward: select your model, pass a prompt, submit the request, and poll for completion. Video generation takes time, so plan for asynchronous workflows. But the customization options, model selection, and prompt control mean you can build video features directly into products rather than treating generation as a manual step.
Codex: Slack, SDK, and Analytics
Codex received three meaningful updates. First, you can now access the agent directly within Slack. Second, the Codex SDK lets you build custom agents on top of OpenAI's infrastructure. If you want to create your own lovable-style app builder or specialized coding assistant, the SDK provides the foundation. Third, usage analytics are now available so you can track how your agents consume tokens and where costs accumulate.
GPT-5 Pro Enters the API
The flagship model is here, but it comes with flagship pricing. GPT-5 Pro costs $15 per million input tokens and $120 per million output tokens. The context window is massive: 400,000 tokens in, 272,000 tokens out.
Independent benchmarks were not available at announcement time, but the expectation is clear. At this price point, OpenAI is positioning it as the best model available across reasoning, coding, and complex task completion. Whether it holds that position depends on head-to-head testing against competitors, but the specs suggest top-tier performance.
Realtime Mini and Image Mini
Two new cost-optimized models launched alongside the premium tier.
GPT Realtime Mini delivers the same voice capabilities as the standard real-time API, intonation and tone understanding included, at 70% lower cost. For voice applications, this makes OpenAI competitive on price without sacrificing the conversational quality that distinguishes their audio models.
GPT Image 1 Mini offers the same deal for visuals. The model handles everything from infographics to photorealistic images at a reduced price point compared to the full GPT Image 1.
Widgets and Conditional UI
The Agent Kit demo revealed a subtle but powerful feature: conditional widgets. Agents can trigger custom UI elements based on conversation state.
Ask about flights, meet the right criteria, and the agent renders a formatted card instead of plain text. You define the widget structure, styling, and rendering logic within the builder. This moves beyond text-only responses into structured, interactive outputs without leaving the ChatGPT ecosystem.
