OpenAI DevDay 2025: GPT-5 Pro API, AgentKit, and 5 Developer Features That Change How You Build with AI

Six billion tokens per minute. Eight hundred million weekly users. The numbers Sam Altman dropped at OpenAI DevDay 2025 on October 6 weren’t just impressive — they were a statement. At Fort Mason in San Francisco, with over 1,500 developers in attendance, OpenAI unveiled a stack of tools that signals a fundamental shift in how developers will build with AI: GPT-5 Pro API, AgentKit, a ChatGPT Apps SDK, Sora 2 in the API, and a fine-tuning overhaul. Here’s everything that matters — and what it means for your workflow.

GPT-5 Pro API: $15/$20 Per Million Tokens — Who Is This For?

The headline announcement was the general availability of GPT-5 Pro API. Priced at $15 per million input tokens and $20 per million output tokens — roughly 10x the cost of standard GPT-5 — this isn’t a model for casual chatbots. OpenAI explicitly positioned it for finance, legal, and healthcare, sectors where reasoning accuracy justifies the premium. Early adopters have reported notable improvements in complex contract analysis and medical diagnostic support, though community feedback flagged uneven latency and consistency in production environments.

Alongside GPT-5 Pro, OpenAI rolled out two budget-friendly models: gpt-realtime-mini, a low-latency voice model priced 70% cheaper than the standard real-time offering, and gpt-image-1-mini, which cuts image generation costs by 80%. The strategy is clear — premium reasoning at the top, cost-efficient specialized models at the bottom, and developers pick their price-performance sweet spot.

AgentKit: Build AI Agents Without Writing a Single Line of Code

If there was one announcement that generated the loudest applause, it was AgentKit. This is OpenAI’s comprehensive toolkit for building, deploying, and monitoring AI agents, and it ships with four key components:

• Agent Builder — A visual, drag-and-drop canvas for designing agent workflows without code
• Connector Registry — MCP-based tool access management with enterprise-grade data governance
• ChatKit — Embeddable chat UI components for shipping agent experiences directly in your product
• Eval & Tracing — Automatic scoring of multi-step agent decisions with trace-level observability

For developers who prefer code over canvases, OpenAI also released the open-source Agents SDK for Python and TypeScript. It exposes primitives for agents, handoffs, guardrails, and sessions with built-in tracing — enabling what TechCrunch described as a “design-and-operate” workflow on custom infrastructure. The dual approach is smart: AgentKit lowers the barrier for non-technical teams while the Agents SDK gives power users the control they need.

ChatGPT Apps SDK: Full-Stack Apps Running Inside a Conversation

Remember ChatGPT plugins? The Apps SDK is what plugins always wanted to be. Built on the Model Context Protocol (MCP), it lets developers build full-stack, interactive applications that run directly inside the ChatGPT interface. We’re not talking about text responses with links — users can operate a design tool, search real estate listings, or create music playlists without ever leaving the conversation.

Launch partners tell the story: Canva for design, Zillow for real estate, Spotify for music, and Figma for interface design. During the keynote demo, Zillow showed conversational property filtering, and Spotify demonstrated AI-driven playlist creation — both running natively in ChatGPT. OpenAI confirmed plans to open a public app store with monetization and an approval process later in 2025. If this sounds familiar, it should — OpenAI is building its own App Store moment.

Fine-Tuning and Codex: Production-Ready Upgrades

Reinforcement fine-tuning hit general availability on o4-mini and entered private beta on GPT-5, now featuring custom tool calls and custom graders. This means you can fine-tune models not just on text quality but on how well they use specific tools and follow multi-step workflows — a meaningful upgrade for anyone building production agent systems.

The evaluation platform also got a major expansion. Datasets, trace grading, and prompt optimizers joined the toolkit, and notably, third-party models can now be evaluated inside OpenAI’s Evals platform. The fact that OpenAI is inviting competitors into its evaluation sandbox suggests real confidence in GPT-5 Pro’s positioning.

And after what felt like an eternity, Codex officially reached general availability. With GitHub Copilot, Cursor, and Claude Code all competing aggressively in the AI coding space, Codex GA adds another serious contender to the mix.

Sora 2 API: Video Generation Goes Enterprise

Sora 2 is now available to developers through the API, delivering cinematic-quality video generation with synchronized sound capabilities. OpenAI emphasized that provenance tracking, consent frameworks, and content review were baked in from the design phase — a clear response to ongoing debates about AI-generated media. For enterprises looking to integrate video generation into commercial pipelines, this is a significant milestone.

The Developer Verdict: Excitement With Caveats

Developer reception was mixed. AgentKit and the Apps SDK earned praise as “tools you can actually ship with,” but one analyst noted the event “felt like doubling down on existing opportunities rather than frontier advancement.” The 10x price tag on GPT-5 Pro raised questions about cost-effectiveness, and early reports of inconsistent latency suggest the model needs more time to stabilize in production environments.

Still, the bigger picture is unmistakable. With 4 million developers and 800 million users, OpenAI is leveraging ecosystem scale to execute a platform lock-in strategy. AgentKit, Apps SDK, and the Evals platform form an integrated development environment that encourages developers to do everything inside OpenAI’s walls. DevDay 2025 wasn’t just a product announcement — it was a declaration of intent in the platform wars. Whether you’re building agents, shipping apps, or fine-tuning models for production, the tools OpenAI dropped in October 2025 deserve a serious look.