7 Major Announcements At OpenAI DevDay 2025

OpenAI Launches Interactive Apps in ChatGPT

At DevDay 2025, OpenAI announced that ChatGPT will now host apps directly within the chat interface, powered by a new Apps SDK (preview). By embedding interactive tools like Spotify, Zillow, Canva, and more inside conversation, ChatGPT shifts from just a chatbot to a dynamic platform where users can accomplish tasks without leaving the chat.

Key Points:

1. SDK & integration - Developers can build apps using the Apps SDK (based on the Model Context Protocol) which defines logic plus interactive UI components.

2. Contextual app suggestions - ChatGPT will surface or suggest relevant apps during conversation; users may also invoke apps by name explicitly.

3. Pilot partners & rollout - Early partners include Booking.com, Canva, Spotify, Zillow. Apps are available now outside the EU on Free, Go, Plus, Pro plans; EU and enterprise expansions to follow.

Conclusion

This move marks a paradigm shift: ChatGPT is no longer just a text assistant but becoming a host for interactive services. By opening access to developers, OpenAI is enabling an ecosystem of apps living inside conversations. Over time, this could reshape how users transition between tools—everything may happen inside chat.

AgentKit: OpenAI’s New Toolset to Build Production-Grade Agents

During DevDay 2025, OpenAI unveiled AgentKit, a unified suite of tools designed to simplify the creation, deployment, and optimization of intelligent agents. The announcement aims to address the fragmentation and complexity developers face when stitching together orchestration, connectors, UI, evaluation, and safety.

Key Points:

1. Agent Builder (visual canvas + versioning) - Enables drag-and-drop workflows that connect logic, agents, and guardrails, with built-in version control and previews.

2. Connector Registry & ChatKit - A centralized system to manage tool/data integrations, plus a toolkit to embed agentic chat UIs in external products.

3. Evaluation & reinforcement fine-tuning - Enhanced evaluation tools (dataset support, trace grading, prompt optimization) and early support for RL fine-tuning of agents.

Conclusion

AgentKit lowers the barrier to building robust agents by consolidating fragmented tools into a cohesive developer experience. For teams, this means faster iteration, better safety control, and more reliable agent deployment. Over time, AgentKit may become the de facto foundation for agent-based products.

Sora 2 Brings Video + Audio Generation into the API

At DevDay, OpenAI launched Sora 2, a next-generation multimedia model capable of generating synchronized video, speech, and ambient sound — now accessible via API. With better consistency, multi-shot support, and the ability to inject real-world cameo footage, Sora 2 pushes generative media into mainstream app building.

Key Points:

1. Video + audio output - Sora 2 can produce video sequences along with matching voice, sound effects, and ambient audio, enhancing realism in generated content.

2. Consistency & control - Improved handling of physics, shot-to-shot coherence, and user instructions across scenes.

3. Cameo & real-world injection - Users can insert themselves or real objects into generated scenes via short captured inputs.

Conclusion

By unifying video and audio generation into a single model and exposing it via API, OpenAI is making immersive multimedia creation more accessible. This opens doors for apps in marketing, storytelling, education, and entertainment - any domain that benefits from generated video. The challenge will lie in quality, cost, and safe, ethical use.

Codex Gets Smarter: Slack, SDK & Enterprise Controls Now Available

At DevDay 2025, OpenAI upgraded Codex with deeper integrations and governance features. Now generally available with Slack integration, a Codex SDK, and enterprise controls, Codex is evolving from a standalone coding assistant into a fully embedded part of developer workflows.

Key Points:

1. Slack integration — Codex can now respond to coding queries inside Slack threads, propose fixes, and interact with context from chat.

2. Codex SDK — Developers can integrate Codex features (code generation, reviews, automation) directly into their tools, CI/CD, or internal platforms.

3. Enterprise governance & logging — New dashboards, permissioning, auditing, and usage controls help organizations adopt Codex safely in production environments.

Conclusion

These enhancements deepen Codex’s role in real-world engineering environments, reducing friction between idea and code. With embedded integration and governance, Codex can become a core part of dev tooling - but reliability, security, and developer trust will determine how far adoption goes.

GPT-5 Pro Arrives: High-Precision Thinking for Critical Tasks

OpenAI used DevDay to introduce GPT-5 Pro, its most capable reasoning model, alongside gpt-realtime-mini, a lighter voice model optimized for cost and latency. Together, they represent a strategy of specialization: use heavyweight models when precision matters, and lean ones when efficiency does.

Key Points:

1. GPT-5 Pro reasoning & token limits - built for depth and complex tasks; supports much longer output ranges than earlier variants.

2. gpt-realtime-mini cost & latency gains - designed as a cheaper, faster voice model (~70% lower cost) for interactive applications.

3. Tiered model approach - encourages intelligent routing: use Pro for high-stakes tasks, mini for efficient modalities.

Conclusion

These model upgrades reflect maturation in AI infrastructure: it's not just about bigger models but right-sized ones. GPT-5 Pro gives depth where needed; gpt-realtime-mini makes voice practical at scale. The real test will be how well developers can balance usage, latency, accuracy, and cost in real deployments.

Cheaper Creativity: gpt-image-1-mini Reduces Image Costs by ~80%

In its DevDay 2025 lineup, OpenAI quietly added gpt-image-1-mini, a streamlined image generation model that can deliver outputs at a cost ~80% less than its full counterpart. This enables more economical experimentation and higher-volume use across applications.

Key Points:

1. Significant cost reduction - around 80% lower cost per image makes high-volume usage more viable.

2. Trade-offs in fidelity - while cheaper, outputs may show lower resolution, detail, or consistency in complex prompts.

3. Usage strategy - developers can use mini for drafts, variants, or early exploration before upgrading to full model for final output.

Conclusion

gpt-image-1-mini democratizes AI imagery by lowering financial barriers to entry. For many use cases, “good enough” may suffice, and this mini variant is likely to see broad adoption in prototyping, scalable pipelines, and environments where cost sensitivity matters. The balance between quality and affordability will be key.

OpenAI and AMD Strike a 6 GW Chip Deal to Power the AI Future

OpenAI announced a strategic partnership with AMD at DevDay 2025: commit to deploying 6 gigawatts of AMD AI chips across future generations, beginning with 1 GW in 2026, and receive warrants for up to 160 million AMD shares (~10%) tied to milestones.

Key Points:

1. Massive compute commitment - a multi-year deployment of 6 GW, starting with AMD’s MI450 in H2 2026.

2. Equity alignment via warrants - OpenAI gains options on AMD shares tied to delivery milestones and performance.

3. Strategic supply diversification - the partnership reduces reliance on any single hardware vendor in the AI compute stack.

Conclusion

This deal isn’t just about securing capacity - it’s about aligning incentives, differentiating infrastructure strategy, and anchoring OpenAI’s compute roadmap. With AMD as a strategic partner, OpenAI is positioning itself to avoid supply constraints and scale confidently; the success will depend on execution, performance, and integration across its broader model ecosystem.

ElevenLabs launches Agent Workflows feature

ElevenLabs’ Agent Workflows feature empowers users to build more sophisticated, branching conversation flows via a visual graph interface instead of linear scripts. With this, agents can adapt to user inputs dynamically, route through tools or subagents, transfer between human or automated agents, and tailor behavior (e.g. voice/tone, knowledge base, system prompt) at different phases of the conversation.

Key Points:

1. Node Types - Several specialized node types let you control conversation flow in detail:

   • Subagent Nodes let you override or modify parts of the agent’s behavior (system prompt, voice configuration, tools, knowledge base) at particular points.

   • Dispatch Tool Nodes are places where specific tools are invoked; they support branching based on success/failure of that tool.

   • Agent Transfer / Transfer to Number Nodes support handing off the conversation to another agent (could be human) or switching via phone systems.

2. Edges & Flow Control - The connections (“edges”) between nodes define how the conversation progresses. The routing can be unconditional or conditional (based on what the LLM determines), and there are forward and backward edges. This allows for adaptive paths rather than strict linear sequences.

3. Fine-grained Behavior Customization During Flow - Aside from routing, the workflow lets you adjust agent configurations mid-conversation: switch LLM models, change voice characteristics, update knowledge‐base or tools for a given phase or node. This supports context-sensitive behavior.

Conclusion

The Agent Workflows capability gives developers using ElevenLabs’ Agents Platform much more control over how conversations evolve, adapt, and respond based on user input or tool outcomes. It supports more realistic, flexible, and maintainable dialogue logic-especially useful for use cases with branching, conditional logic, or handoffs (e.g., customer support, educational agents, interactive storytelling). As workflows get more complex, these visual tools help avoid spaghetti logic and make maintaining and iterating on behavior more manageable.