OpenAI DevDay 2025 recap: apps in ChatGPT, AgentKit, Codex and Sora 2 for marketing teams

Updated 2 October 2026: prices are now shown in US dollars as OpenAI lists them, the UK availability of apps in ChatGPT is corrected from OpenAI’s launch post, and the box below records what OpenAI has retired or renamed since. The coverage that follows is as we wrote it in October 2025.

What has changed since DevDay 2025 (checked 2 October 2026)

Much of what was launched on this day has been replaced within a year. OpenAI’s deprecations page lists Sora 2 and the Videos API as shut down on 24 September 2026, Agent Builder as closing on 30 November 2026 with a migration guide to the Agents SDK, the DevDay snapshot of GPT-5 Pro as closing on 11 December 2026, gpt-image-1-mini on 1 December 2026 and gpt-realtime-mini on 20 January 2027. Codex itself carries on, and OpenAI’s help centre says that on 9 July 2026 the ChatGPT app directory was merged into the Plugins directory, so the apps in ChatGPT described below are now found and installed as plugins. For what OpenAI announced this year, including dots, GPT-6.1 Sol and Pro 500, read our OpenAI DevDay 2026 recap.

OpenAI’s third annual DevDay at Fort Mason in San Francisco on 6 October 2025 felt different from previous years. Less research showcase, more commercial proposition. The company unveiled tools that transform ChatGPT from a conversational interface into something closer to a platform for building modern applications. Sam Altman opened with numbers that are difficult to ignore: ChatGPT now serves 800 million people weekly, up from 100 million just two years ago. Four million developers are building on the platform, which has doubled since 2023. The API processes 6 billion tokens per minute, a thirtyfold increase from 200 million in 2023.

For digital marketing teams, the event signalled that AI integration is moving from experimental to operational faster than many agencies have prepared for.

ChatGPT becomes a distribution channel

The Apps SDK was the headline announcement, and it changes the terms of customer interaction. Third-party developers can now build full applications that run directly inside ChatGPT conversations. This creates a new distribution channel. Rather than driving users to websites or mobile apps, brands can deliver interactive experiences within the conversational flow where 800 million people already spend their time.

Think about what this means in practice. A travel brand displays bookable properties with pricing and availability directly in chat. An e-commerce retailer renders product catalogues with add-to-basket functionality. A media company serves personalised content recommendations with inline playback. The launch partners named in OpenAI’s announcement, Booking.com, Canva, Coursera, Expedia, Figma, Spotify and Zillow, show this is more than theoretical. These are live integrations handling real transactions.

The technical side builds on Anthropic’s Model Context Protocol, which OpenAI has extended to support component rendering with HTML. This allows developers to design both logic and interface in a single framework whilst maintaining cross-platform compatibility. From a business perspective, the value proposition is simple: develop once, deploy to an audience the size of Instagram’s early peak. The apps went live on 6 October to logged-in users on the Free, Go, Plus and Pro plans outside the European Economic Area, Switzerland and the United Kingdom, with OpenAI saying only that it expected “to bring apps to EU users soon”. UK users were not in the first wave, a pattern OpenAI repeated with dots a year later.

For agencies, this opens several service opportunities. Brands will need a strategy for conversational commerce, implementation expertise for app development, and optimisation capabilities to drive engagement within the ChatGPT ecosystem. The technical requirements include HTTPS endpoints with Server-Sent Events or Streamable HTTP transport, OAuth 2.1 authentication, and dynamic client registration. These are established web development standards, which means most technical teams can implement Apps SDK integrations without requiring entirely new skill sets.

The security considerations deserve attention, though. When you’re building apps that connect to multiple data sources simultaneously, the attack surface expands considerably. Clear data handling policies, regular security audits, and compliance with GDPR or industry-specific requirements become non-negotiable.

Nick Turley, OpenAI’s head of ChatGPT, told Axios the day after the keynote: “It’s not inconceivable to me that over time, you perceive ChatGPT to be a type of operating system.” That’s the pitch, anyway. Whether it becomes the AI-first platform OpenAI envisions remains to be seen, but the infrastructure is now in place for developers to find out.

Figma integration moves markets

Among the initial ChatGPT apps, Figma’s integration stood out enough to send its stock soaring. Sam Altman highlighted it during the keynote, and investors responded by driving Figma’s share price up as much as 15 percent intraday before it pared the gain. So what does the integration actually do?

Essentially, it links ChatGPT with Figma’s collaborative design tools, allowing users to generate diagrams and visuals from a conversation. You can convert ChatGPT conversations directly into FigJam diagrams (Figma’s online whiteboard product). Brainstorming text becomes flowcharts or wireframes in real time. Conversely, ChatGPT can also suggest using Figma when a visual aid might help, handing off part of the task to Figma’s interface and then continuing the dialogue.

This tight coupling of a design platform with an AI chatbot hints at how creative work may become more conversational and assisted by AI. A product manager could sketch out an idea in text and have ChatGPT and Figma translate it into a draft design, which can then be refined collaboratively. For non-designers, this lowers the barrier to creative planning considerably. The boundary between discussing an idea and designing it is disappearing.

The stock market reaction tells you something about the perceived value of this early platform integration. By being a launch partner, Figma gains credibility and a foothold in OpenAI’s app ecosystem, potentially expanding its reach beyond traditional design teams into broader brainstorming and ideation contexts. The move comes as its competitor Adobe pushes its own AI features.

For marketing teams, this suggests a case study worth studying. AI platforms can drive significant user engagement to services that integrate early, creating new growth channels. The AI handles the generative heavy lifting or tedious setup, whilst humans focus on higher-level direction and refinement.

AgentKit: building AI agents gets easier

AgentKit tackles something most agencies have already discovered: building AI agents is still more difficult than it should be. The toolkit bundles five components designed to reduce development time and improve reliability.

Agent Builder provides a visual drag-and-drop canvas for multi-agent workflows, in beta at launch. This allows developers to orchestrate complex sequences of AI actions with a clear visual overview, enabling faster iteration and collaboration across teams. Until now, creating an AI agent required stitching together various components (prompts, model calls, external tool APIs, custom logic). AgentKit streamlines this considerably.

ChatKit offers an embeddable chat UI with custom branding capabilities, generally available from day one. If you build an AI agent with OpenAI’s models, ChatKit provides the front-end components to integrate a ChatGPT-like chat experience into your product, skinned to your brand and needs.

The Connector Registry centralises APIs and data sources with administrative controls. This makes it easier to plug an agent into the software and data it needs, whilst maintaining security and admin control. It’s a hub to manage integrations with services like Dropbox, Google Drive, Slack, internal databases, and third-party MCP-based tools.

Guardrails is an open-source safety layer that can mask or flag personal data, detect jailbreaks and apply other checks before an agent acts.

Enhanced Evals for Agents adds step-by-step trace grading, datasets, and automated prompt optimisation. Developers can measure and improve an agent’s performance with reinforcement learning fine-tuning options.

The live demonstration at DevDay made the value proposition clear. An OpenAI engineer built a complete DevDay assistant with two functional agents in under eight minutes. Whilst live demonstrations are always optimised for impact, the early adopter examples OpenAI published suggest these efficiency gains translate to production environments. Ramp says Agent Builder turned months of orchestration work into a couple of hours and cut its iteration cycles by 70 percent. LY Corporation built and ran its first multi-agent workflow in under two hours. Klarna’s support agent handles two-thirds of all customer service tickets.

Canva, Evernote and Albertsons were among the first ChatKit users, and Box used the new Evals, according to OpenAI. For marketing applications, several use cases become immediately more viable.

Customer service automation becomes practical when you can build branded chatbots that handle common enquiries whilst maintaining brand voice and escalating complex queries to human agents. The ability to integrate with existing CRM systems, ticketing platforms, and knowledge bases creates a unified customer experience rather than yet another disconnected tool.

Lead qualification is another area where this makes sense. Deploy agents that engage with prospects, ask qualifying questions, and route high-intent leads to sales teams. The Connector Registry allows secure integration with marketing automation platforms, ensuring lead data flows into existing workflows rather than creating manual export-import processes.

Content workflows could benefit as well, though this is where agencies will need to be most careful about quality control. You can create agents that assist with content research, generate first drafts, optimise for SEO, and maintain brand guidelines. Multi-agent systems can coordinate research, writing, editing, and approval processes within a single workflow. Whether that actually saves time or just creates a different kind of overhead remains to be seen in practice.

Campaign management is perhaps the most immediately practical application. Develop agents that monitor campaign performance, identify anomalies, generate reports, and suggest optimisations based on historical data. Integration with Google Analytics, Meta Ads, and other platforms provides a lot of related data. This is the kind of repetitive analysis work that benefits from automation.

AgentKit carries no additional fees beyond standard API pricing, making it accessible for agencies and clients of various sizes. The Connector Registry is rolling out in beta to some API, ChatGPT Enterprise and Edu customers, and needs the Global Admin Console, which gives IT teams domain, single sign-on and multi-organisation controls. For agencies managing multiple client accounts, these administrative controls are essential for maintaining security and compliance standards.

Codex: the AI coding assistant goes mainstream

Codex, OpenAI’s software engineering agent, reached general availability with features that suggest the company is serious about enterprise adoption. OpenAI says Codex now reviews almost every pull request at the company and that its engineers merge 70 percent more pull requests each week, a striking testament to how far AI-assisted coding has gone internally.

At general availability the agent runs on GPT-5-Codex, a version of GPT-5 tuned for agentic coding that adjusts how long it thinks to the size of the task. It writes features, fixes bugs, runs tests, and proposes pull requests, all within isolated cloud sandboxes.

The development timeline from research preview in May to production readiness by October is notably fast. The generally available version includes Slack integration, allowing teams to assign tasks directly from conversations. There’s a Codex SDK for embedding the agent in your own tools and CI/CD pipelines. Administrative tools provide environment controls, monitoring and analytics dashboards for organisations juggling multiple projects.

For marketing agencies, Codex addresses several operational bottlenecks that technical teams will recognise immediately. Website and landing page development could be accelerated considerably. Companies like Cisco, Duolingo, Rakuten and Instacart are early adopters, and Cisco reports code reviews that are 50 percent faster.

Marketing technology integration is another practical application. Building and maintaining integrations between marketing platforms, CRM systems, analytics tools, and data warehouses is the kind of work that’s essential but rarely exciting. The ability to generate clean code from natural language descriptions reduces the technical barrier for marketing technologists who understand what needs connecting but don’t necessarily want to spend hours writing API integrations.

Automated reporting benefits as well. Developing custom reporting dashboards and data pipelines that consolidate metrics across multiple platforms is simple in concept but often tedious to implement. Codex can write scripts that extract, transform, and load data from APIs, maintaining these integrations as platforms inevitably change their specifications.

Technical SEO implementation is perhaps less obvious but potentially useful. Generating structured data markup, creating XML sitemaps, implementing canonical tags, and building tools for custom technical SEO audits all require attention to web standards. The model’s understanding of current best practices means you’re less likely to end up with deprecated markup or non-compliant implementations.

GPT-5-Codex served over 40 trillion tokens in the three weeks after its launch, which points to real production usage rather than experimental testing. From 20 October 2025, Codex cloud tasks count towards usage limits. This marks the shift from preview to metered commercial service, which is fair enough but requires agencies to do proper capacity planning and budget allocation. The efficiency gains often justify the costs, particularly for teams managing multiple client projects with recurring technical requirements. But “often” doesn’t mean “always”, and testing with real workloads before committing to production usage continues to make sense.

Four new models that actually matter

OpenAI introduced four models optimised for different use cases and price points, each with specific applications for marketing operations. The prices below are OpenAI’s US list prices for the API, in dollars, as published on its model pages.

GPT-5 Pro: enterprise reasoning

GPT-5 Pro targets enterprise workloads requiring high precision and deep reasoning. Priced at $15 per million input tokens and $120 per million output tokens, it represents a tenth of the price of the previous o1-pro at $150 and $600. For marketing applications, GPT-5 Pro suits complex strategic analysis, competitive research synthesis, and content planning that requires a nuanced understanding of brand positioning and market dynamics. Whether it outperforms o3-pro as claimed remains to be tested in practice, but the pricing alone makes it worth evaluating for high-stakes work.

Sora 2: video generation goes mainstream

Sora 2 is where things get interesting from a content production perspective. OpenAI describes it as “more physically accurate, realistic, and more controllable than prior systems”, with synchronised dialogue and sound effects, persistent world state across multi-shot sequences and detailed camera direction, in realistic, cinematic and anime styles. In the API, Sora 2 costs $0.10 a second for 720p output and Sora 2 Pro $0.30 a second, with 1080p on the Pro model.

The marketing applications are fairly direct. Generate short-form video content for Instagram Reels, TikTok, and YouTube Shorts without relying on stock videos or traditional video production resources. Create product demonstrations and feature explainer videos from text descriptions, reducing both time and cost whilst enabling rapid iteration based on performance data. Transform campaign concepts and storyboards into video mockups for client presentations, allowing stakeholders to evaluate creative directions before committing to full production budgets.

There’s also a “Cameos” feature that allows insertion of specific individuals into generated scenes following one-time identity verification. The potential for personalised video in account-based marketing campaigns is obvious, though there will be some trial and error around what feels compelling versus what crosses into uncanny valley territory.

The API went live the same day as the announcement, which is unusual for OpenAI and suggests confidence in its production readiness. Programmatic access supports workflow automation and integration with existing content management systems.

This opens the door for applications like dynamic video content creation, game development, or marketing, where an app can generate bespoke video clips on demand. Imagine generating concept commercials or visualising product designs via AI video. For creativity, individuals and small businesses could produce quality videos without big budgets. But it also intensifies concerns around deepfakes and misinformation if not used responsibly. OpenAI appears aware of this, rolling out features in Sora (the consumer app) for more user control and content moderation.

gpt-realtime-mini: voice commerce gets viable

The voice model, gpt-realtime-mini, is a smaller native speech-to-speech model that processes audio directly without text conversion. OpenAI prices it at 70 percent less than its large gpt-realtime model while, it says, keeping the expressiveness.

Marketing applications include voice commerce, where customers can make purchases and resolve issues through natural voice interactions. Voice-activated content that responds dynamically to listener questions. Customer service deployments that handle routine enquiries with natural conversation flow. Another potential use case would be accessibility features that ensure marketing content reaches audiences with visual impairments or reading difficulties.

The model responds over WebRTC, WebSocket or SIP connections, the last of which means phone calling. The latency improvements over traditional multi-step pipelines create noticeably more natural conversational experiences, which matters for customer-facing applications.

gpt-image-1-mini: scaled visual production

gpt-image-1-mini is priced at 80 percent less than the large gpt-image-1, at $8 per million image output tokens. On OpenAI’s model page a high-quality 1,024 by 1,024 image works out at about 3.6 cents, medium quality 1.1 cents and low quality half a cent.

For marketing teams managing substantial image requirements across social media, display advertising, email campaigns, and web content, the cost reduction enables scaled production of visual assets. The ability to generate variations quickly supports A/B testing of creative concepts. The transparent background option facilitates integration into various design contexts. Agencies can use gpt-image-1-mini for concepting and iteration, reserving the more expensive gpt-image-1 for final production assets requiring higher quality. That tiered approach to quality versus cost is probably how most teams will end up using these models in practice.

Infrastructure investment: the AMD partnership

OpenAI announced a strategic partnership with AMD to deploy 6 gigawatts of AMD Instinct GPUs, with OpenAI receiving a warrant for up to 160 million AMD shares that vests as deployment, share-price and commercial milestones are met. That’s roughly 10 percent of the company, which gives you some sense of the scale of infrastructure investment required. The first gigawatt, of MI450 GPUs, is due to begin in the second half of 2026. For context, 6 gigawatts is on the order of the output of several large power plants, highlighting the voracious compute needs of advanced AI models.

The arrangement is financially hefty. AMD expects the partnership to generate tens of billions of dollars in annual revenue, and more than $100 billion in new revenue over four years from OpenAI and other customers, Al Jazeera reported. News of the alliance sent AMD’s shares up more than 34 percent during the day, on track for their biggest one-day gain in more than nine years, adding roughly $80 billion to AMD’s market value.

This partnership ties OpenAI to one of Nvidia’s chief rivals in the AI chip space. Nvidia currently dominates AI hardware (and had itself agreed to invest in OpenAI recently), but OpenAI’s massive commitment to AMD’s forthcoming MI series GPUs indicates a desire to diversify its hardware base. In effect, OpenAI is ensuring it has enough cutting-edge chips to power future models. By securing a stake in AMD, it could benefit if AMD’s AI market share grows.

For agencies relying on OpenAI’s platform for client deliverables, improved infrastructure stability reduces the operational risk of API unavailability. One of the limiting factors for AI adoption has been the scarcity and expense of GPU compute. By locking in 6GW of compute, OpenAI is gearing up to train larger models (think GPT-6 and beyond) and serve more complex queries, which could translate to better AI capabilities available via its API and ChatGPT. It might also reduce the risk of service capacity crunches.

Priority and flex processing tiers

The priority processing announcement matters for anyone running production systems. GPT-5 API requests run about 40 percent faster on the priority processing tier compared to the standard tier. There’s a new Service Health Dashboard that provides real-time monitoring of uptime, request time, token velocity, and time to first token. Priority processing delivers SLA-backed, predictably low latency even during peak demand, with Enterprise access and premium per-token pricing. The alternative “flex” tier offers 50 percent cheaper processing with increased latency for o3, o4-mini, and gpt-5 models.

For agencies managing client campaigns with time-sensitive requirements, the performance tiers enable appropriate service level selection. Standard tier suffices for batch processing and non-urgent tasks. Priority tier ensures reliable performance for real-time applications like chatbots, voice interactions, and live customer service. Flex tier provides cost savings for development environments and testing workflows where latency tolerance is higher. In practice, most agencies will probably use a combination of all three tiers depending on specific workload requirements.

The Service Health Dashboard is one of those features that seems minor until you actually need it. When API performance degrades, being able to quickly identify whether issues stem from OpenAI’s infrastructure or internal systems reduces diagnostic time considerably. For agencies with service level agreements, this visibility supports more accurate performance reporting.

What this actually means for marketing agencies

The announcements at DevDay 2025 suggest several considerations for marketing agencies evaluating AI integration, though it’s worth approaching this with some healthy scepticism alongside the enthusiasm.

OpenAI’s evolution from a research organisation to a platform provider does reduce adoption risk. With 800 million users, 4 million developers, and established enterprise customers, the platform demonstrates commercial viability beyond experimental implementations. That said, platform maturity doesn’t automatically translate to marketing ROI. The technology exists, but a successful application still requires strategy, implementation expertise, and realistic expectations about what AI can and cannot do effectively.

The Apps SDK, AgentKit, and Codex create several marketing service opportunities. Clients will require strategy development for conversational commerce, technical implementation of custom applications, and ongoing optimisation as the platform evolves. Agencies that develop expertise early gain a competitive advantage in an emerging market. However, it’s also worth remembering that we’ve been here before with other platforms that promised revolution and rather just delivered evolution. Early adopter advantage exists, but so does early adopter risk.

Whilst new models offer significant price reductions compared to predecessors, production usage at scale can be more expensive than initial estimates suggest, particularly for multimodal work where image, audio and video tokens add up quickly. Testing with limited traffic before full deployment is essential for avoiding surprises. Monitor token consumption patterns closely and adjust implementation accordingly.

Security and compliance become more complex as AI agents access multiple data sources simultaneously. For anyone managing client data across CRM systems, marketing platforms, and analytics tools, this includes clear data handling policies, regular security audits, and compliance with GDPR or industry-specific requirements.

OpenAI competes directly with Anthropic’s Claude, Google’s AI tools, and Microsoft’s GitHub Copilot. The Apps SDK’s foundation on Anthropic’s Model Context Protocol suggests increasing interoperability, which potentially reduces platform lock-in risk. That’s encouraging, though cross-platform compatibility in practice often proves more complicated than in theory.

An ecosystem that evolves

OpenAI’s DevDay 2025 announcements collectively paint a picture of a company expanding its reach on all fronts. It’s not just launching a new model here or a feature there, but building an entire ecosystem. On the software side, turning ChatGPT into an app platform and launching AgentKit means OpenAI wants to be the foundation upon which AI-powered applications are built and delivered, much like iOS for mobile apps or Windows for desktop software.

This creates opportunities for developers (new channels, tools, and revenue streams) whilst giving users a more integrated AI experience. On the model side, OpenAI is pushing the envelope in both capabilities (GPT-5 for coding, Sora 2 for video) and accessibility (making these available in products and APIs, not just research). On the hardware side, it’s securing the raw power needed to fuel those ambitions at scale.

For consumers, these developments promise AI that is more useful, ubiquitous, and integrated into the apps we already use (and perhaps some new ones). Ask your chatbot to book travel, sketch a UI, or even generate a short video, and it can oblige by tapping specialised services behind the scenes.

For the tech industry at large, OpenAI’s moves signal both consolidation and competition. The company is solidifying its role as a platform player, not just an AI model provider. This could concentrate AI activity around OpenAI (and by extension Microsoft, its major backer and partner), raising questions about ecosystem lock-in and control, much as big tech platforms have in the past. At the same time, the partnership with AMD and support for open standards like MCP show that OpenAI is willing to collaborate and shape industry standards, possibly to avoid bottlenecks and regulatory scrutiny.

Practical next steps

For agencies considering OpenAI’s platform following DevDay, a measured approach makes sense. Start by identifying specific client challenges that align with the announced capabilities. Customer service automation, content production workflows, and technical marketing operations represent practical starting points with measurable ROI. Avoid the temptation to find problems that fit the solution. Focus on real operational bottlenecks that these tools might address more effectively than current approaches.

Begin with contained pilot projects that deliver value quickly whilst limiting exposure. A ChatGPT app for a single client, an AgentKit-powered customer service bot, or Codex integration for landing page development provides learning opportunities without betting the entire agency strategy on unproven implementations. Learn what works in practice rather than what sounds impressive in theory.

Establish a clear understanding of token consumption patterns for intended use cases before committing to production deployment. Test with representative data volumes. Ensure budget allocations reflect actual usage rather than theoretical estimates. The gap between expected and actual costs can be substantial, particularly for implementations that process large volumes of content or maintain persistent conversational context.

Conduct a thorough assessment of data handling requirements, particularly for implementations that connect multiple client systems. Establish clear policies for data access, storage, and transmission that comply with relevant regulations and client security standards. Document these policies clearly and review them regularly as implementations evolve.

Establish metrics for measuring implementation success beyond anecdotal evidence. This includes technical performance (response time, error rates, availability), business outcomes (cost savings, efficiency gains, customer satisfaction), and comparative analysis against previous solutions. Without baseline measurements, determining whether implementations actually deliver value becomes largely subjective.

The DevDay announcements represent a substantial evolution in OpenAI’s commercial offering. For marketing agencies, the strategic question is not whether to integrate AI capabilities but how to do so effectively while managing costs, security, and complexity. Several new tools now exist for production deployment. The challenge lies in applying them to deliver measurable client value.

The shift from research showcase to production platform is complete. What happens next depends on whether agencies can move from experimentation to operational integration without losing sight of what actually drives results for clients. The technology is ready. The question is whether marketing teams are.

A year on

Twelve months later the pattern is clear: OpenAI ships a platform at DevDay and replaces much of it within the year. Sora 2 has been switched off, Agent Builder is closing, and the apps have become plugins. The lesson for a UK team is to pilot on what is sold today and to budget for a migration, which is why our DevDay 2026 recap leads with prices and availability rather than demos. If you want help deciding which of this year’s launches to test, talk to AIWIZ.