OpenAI Dev Day 2025 Recap

OpenAI's Dev Day 2025 was packed with announcements, but which ones actually matter for developers and businesses? In this episode, I walk through the key launches including AgentKit (build AI agents without code), Apps within ChatGPT (third-party integrations that work inside the platform), Sora 2 API access for video generation, and the eye-wateringly expensive GPT-5 Pro.

I share my honest take on each announcement, explain what the pricing really means, and help you figure out which tools are worth your attention. 

Episode Transcript

(00:00):
Hi everyone.

(00:00):
Andy here and welcome back to the AI breakdown.
Yesterday was open AI's Dev Day 2025, and Sam Altman and the team dropped some big announcements and whilst there was no big model upgrades, the sheer volume of what they shipped is still kind of staggering.
Before we dive into the details, let me give you a sense of scale here.
OpenAI now has 4 million developers building on their platform.

(00:24):
That number has doubled in the past two years.
They've got 800 million weekly chat GPT users.
And their API is processing 6 billion tokens per minute.
Just to put that in perspective, two years ago, that number was 300 million tokens per minute.
That's a 20 fold increase in API usage.
So yeah, OpenAI isn't just growing.
They're operating at a completely different scale to almost anyone else in this space.

(00:49):
Today, I'm going to walk you through the key announcements from dev Day.
Share my take on what actually matters.
And help you figure out which of these tools might be worth your attention.
So let's dive in. 17 00:01:09,321.424525857 --> 00:01:13,11.424525857 First up is something, OpenAI is calling apps within chat, GPT. 18 00:01:13,701.424525857 --> 00:01:20,61.424525857 This is probably the most consumer facing announcement from the event, though it's got real implications for developers too. 19 00:01:20,751.424525857 --> 00:01:22,491.424525857 The idea is simple but powerful. 20 00:01:22,851.424525857 --> 00:01:28,671.424525857 You can now access third party services directly inside chat, GPT, without ever leaving the interface. 21 00:01:29,31.424525857 --> 00:01:38,481.424525857 Think Spotify, Zillow, Canva, Expedia instead of chat GPT, just giving you information or suggestions, it can now actually do things on these platforms. 22 00:01:38,901.424525857 --> 00:01:40,386.42452586 So let's say you want to create a playlist. 23 00:01:41,196.42452586 --> 00:01:52,596.42452586 Chat, GPT is pretty decent at suggesting songs or putting together themed playlists, but now with the Spotify app integration, it can go ahead and create that playlist in your actual Spotify account. 24 00:01:52,956.42452586 --> 00:01:55,926.42452586 You don't need to copy and paste anything or switch apps. 25 00:01:56,436.42452586 --> 00:01:58,176.42452586 It's the same thing with property search. 26 00:01:58,596.42452586 --> 00:02:06,726.42452586 Ask chat GPT about homes in a specific area and they can pull live data from Zillow complete with prices, photos, and neighborhood details. 27 00:02:07,386.42452586 --> 00:02:13,776.42452586 Or if you need to knock together a quick graphic for social media, you can work with Canva right there in the chat interface. 28 00:02:14,406.42452586 --> 00:02:15,996.42452586 Now, here's where it gets interesting. 29 00:02:15,996.42452586 --> 00:02:23,526.42452586 From a developer perspective, OpenAI has released what they're calling the apps, SDK, which is built on something called MCP. 30 00:02:24,66.42452586 --> 00:02:31,26.42452586 This is basically an open standard that lets developers build these integrations and eventually publish them for everyone to use. 31 00:02:31,476.42452586 --> 00:02:36,366.42452586 What I find clever about this approach is that OpenAI is positioning chat, GBT. 32 00:02:36,711.42452586 --> 00:02:44,421.42452586 Not just as a chatbot, but as a platform, and they want to be the place where you do everything, not just the place where you ask questions. 33 00:02:44,751.42452586 --> 00:02:47,571.42452586 This is smart positioning, but it's also risky. 34 00:02:48,411.42452586 --> 00:02:55,431.42452586 We've seen this playbook before with platforms like Facebook and others who tried to keep users inside their walled gardens. 35 00:02:55,431.42452586 --> 00:02:58,731.42452586 It works brilliantly when the platform is growing and dominant. 36 00:02:59,46.42452586 --> 00:03:05,46.42452586 But users can get frustrated if integrations feel clunky or limited compared to just using the native apps. 37 00:03:05,466.42452586 --> 00:03:12,726.42452586 The real test will be how seamless and useful these integrations actually are, and whether developers embrace the SD km. 38 00:03:13,236.42452586 --> 00:03:15,786.42452586 If the experience is smooth and the adoption is strong. 39 00:03:16,116.42452586 --> 00:03:19,296.42452586 This could genuinely change how people interact with Chat GPT. 40 00:03:19,776.42452586 --> 00:03:25,686.42452586 But if it feels like a half-baked feature, people will just keep switching between apps like they always have. 41 00:03:27,167.61136774 --> 00:03:29,687.61136774 Now, this is the one that got developers really excited. 42 00:03:30,107.61136774 --> 00:03:38,327.61136774 OpenAI announced Agent Kit, which is essentially a complete platform for building AI agents without needing to write a single line of code. 43 00:03:38,867.61136774 --> 00:03:44,567.61136774 Think of it like N eight N or zam, but specifically designed for orchestrating AI agents. 44 00:03:44,957.61136774 --> 00:03:52,337.61136774 You get a visual interface where you can drag and drop different components, set up conditional logic, and build out entire agent workflows. 45 00:03:52,862.61136774 --> 00:03:58,562.61136774 Sam Altman positioned this as a way to take agents from prototype to production with way less friction. 46 00:03:58,862.61136774 --> 00:04:10,22.61136774 And honestly, that's exactly what the market needs right now because while everyone's been talking about AI agents for the past year or so, actually building reliable production ready agents has been quite difficult. 47 00:04:10,472.61136774 --> 00:04:12,422.61136774 Agent Kit includes a few key pieces. 48 00:04:12,752.61136774 --> 00:04:15,967.61136774 There's Agent Builder, which is basically a visual workflow creator. 49 00:04:16,652.61136774 --> 00:04:24,92.61136774 You can map out your agent's logic, define different paths based on conditions and tie in various tools and capabilities. 50 00:04:24,632.61136774 --> 00:04:30,902.61136774 Then there's chat kit, which gives you an embeddable chat interface that you can customize and drop into your own applications. 51 00:04:31,292.61136774 --> 00:04:37,802.61136774 They've also included what they call guardrails, which is safety screening tools for both inputs and outputs. 52 00:04:38,132.61136774 --> 00:04:45,572.61136774 This is crucial because one of the biggest challenges with the agents is making sure they don't go off the rails and do something unexpected or harmful. 53 00:04:46,172.61136774 --> 00:04:49,232.61136774 And finally, there's an evaluation system built right in. 54 00:04:49,622.61136774 --> 00:04:57,32.61136774 You can create data sets, trace what your agent is doing, and even have the system automatically optimize your prompts based on performance. 55 00:04:57,452.61136774 --> 00:04:59,702.61136774 Now here's where I think this gets really interesting. 56 00:05:00,212.61136774 --> 00:05:04,112.61136774 OpenAI is clearly trying to own the entire agent development life cycle. 57 00:05:04,472.61136774 --> 00:05:12,392.61136774 They want you to prototype in agent kit, refine an agent kit, deploy through their infrastructure and monitor everything using their tools. 58 00:05:12,902.61136774 --> 00:05:15,662.61136774 From a business perspective, this makes total sense. 59 00:05:16,22.61136774 --> 00:05:21,122.61136774 The more they can embed themselves into your development workflow, the stickier their platform becomes. 60 00:05:21,482.61136774 --> 00:05:25,172.61136774 But from a developer perspective, you've gotta think carefully about lock-in. 61 00:05:25,562.61136774 --> 00:05:32,457.61136774 If you build your entire agent infrastructure on open AI's tools, switching to another platform later becomes significantly harder. 62 00:05:33,392.61136774 --> 00:05:38,222.61136774 Said, I think for most teams the productivity gains will outweigh the lock-in concerns. 63 00:05:38,612.61136774 --> 00:05:40,682.61136774 Building agents from scratch can be hard. 64 00:05:41,42.61136774 --> 00:05:46,922.61136774 If Agent Kit can genuinely simplify that process, it's probably worth the trade-off for most use cases. 65 00:05:47,552.61136774 --> 00:05:53,852.61136774 I'm planning to spend some time with the agent kit, so I'll share more detailed thoughts once I've actually built something with it. 66 00:05:55,418.01034689 --> 00:06:01,543.01034689 Now we get to SOA two open AI's video generation model, which is now available through the API. 67 00:06:02,333.01034689 --> 00:06:07,43.01034689 This is huge for anyone building applications that need to generate video content. 68 00:06:07,523.01034689 --> 00:06:14,3.01034689 SOAR two can create realistic video clips from text prompts or images, and it includes synchronized audio. 69 00:06:14,363.01034689 --> 00:06:20,573.01034689 We're talking richly detailed dynamic scenes with a proper understanding of 3D space and motion. 70 00:06:20,993.01034689 --> 00:06:25,313.01034689 The model can generate clips that maintain scene continuity and realistic physics. 71 00:06:25,643.01034689 --> 00:06:30,173.01034689 From a technical perspective, accessing this through the API is straightforward. 72 00:06:30,683.01034689 --> 00:06:34,283.01034689 You pick your model, there's SOA two and SOA two Pro. 73 00:06:34,673.01034689 --> 00:06:41,3.01034689 You pass in your prompt and then you pull for the status of the generation because obviously generating video takes time. 74 00:06:41,423.01034689 --> 00:06:51,668.01034689 The obvious use cases are marketing content, social media, product demos, educational videos, anywhere you need quick video content without the time and cost of traditional video production. 75 00:06:52,433.01034689 --> 00:06:54,23.01034689 But here's what I'm curious about. 76 00:06:54,383.01034689 --> 00:07:02,483.01034689 How good is it really? Because we've seen impressive demos before that don't quite live up to expectations when you start using them at scale. 77 00:07:02,843.01034689 --> 00:07:19,823.01034689 Can it maintain consistency across multiple generated clips? How well does it handle specific brand guidelines or visual styles? These are the questions I want answers to, and I suspect we'll start seeing real world reviews and case studies over the coming weeks as developers get their hands on it. 78 00:07:21,320.5944032 --> 00:07:26,960.5944032 Although there weren't any big model upgrades announced, GPT five PRO is now available through the API. 79 00:07:27,380.5944032 --> 00:07:35,600.5944032 This is open AI's most powerful model, and they're positioning it as the absolute top tier for when you need maximum intelligence and capability. 80 00:07:36,230.5944032 --> 00:07:37,820.5944032 The context window is massive. 81 00:07:38,300.5944032 --> 00:07:45,945.5944032 You can pass in up to 400,000 tokens of input and they can generate up to 272,000 tokens of output. 82 00:07:46,790.5944032 --> 00:07:51,140.5944032 For context, that's enough to handle entire code bases or very long documents. 83 00:07:51,500.5944032 --> 00:07:55,700.5944032 But let's talk about the pricing because this is where it gets eye watering. 84 00:07:56,90.5944032 --> 00:08:02,60.5944032 It's $15 per million input tokens and $120 per million output tokens. 85 00:08:02,390.5944032 --> 00:08:04,550.5944032 That's expensive relative to other models. 86 00:08:04,820.5944032 --> 00:08:10,520.5944032 And to put that in perspective, regular GPT five costs $1 25 per million input tokens. 87 00:08:11,60.5944032 --> 00:08:13,700.5944032 And $10 per million output tokens. 88 00:08:14,150.5944032 --> 00:08:18,260.5944032 So GPT five PRO is roughly 12 times more expensive. 89 00:08:18,650.5944032 --> 00:08:23,510.5944032 This pricing positions GPT five PRO as a model you use very selectively. 90 00:08:23,990.5944032 --> 00:08:27,560.5944032 This isn't for general chatbot applications or high volume processing. 91 00:08:28,10.5944032 --> 00:08:36,680.5944032 This is for when you absolutely need the best possible performance on complex reasoning tasks, where the cost is justified by the value of getting the right answer. 92 00:08:37,445.5944032 --> 00:08:47,825.5944032 Think high stakes decision support, complex code generation, detailed research analysis, use cases where you're willing to pay a premium because the alternative is much more expensive. 93 00:08:47,825.5944032 --> 00:08:50,675.5944032 Human time or the cost of getting it wrong is high. 94 00:08:51,5.5944032 --> 00:08:58,715.5944032 For most applications, regular GPT five or even GPT five mini will probably be more than sufficient. 95 00:08:58,920.5944032 --> 00:09:05,525.5944032 But having GPT five PRO available is valuable for those edge cases where you need maximum capability. 96 00:09:06,972.43039779 --> 00:09:12,612.43039779 OpenAI also announced GPT Realtime Mini, which is a cheaper version of their realtime voice, API. 97 00:09:13,212.43039779 --> 00:09:20,892.43039779 It's 70% cheaper than the standard realtime API, which makes it significantly more accessible for building voice applications. 98 00:09:21,282.43039779 --> 00:09:34,482.43039779 If you are building anything with voice interfaces, conversational ai, phone systems, or interactive voice experiences, this is worth paying attention to the real time API, lets you capture all those natural intonations. 99 00:09:34,572.43039779 --> 00:09:35,652.43039779 Understand tone. 100 00:09:36,42.43039779 --> 00:09:37,512.43039779 Handle interruptions properly. 101 00:09:37,842.43039779 --> 00:09:42,222.43039779 All the things that make voice interactions feel natural rather than robotic. 102 00:09:42,672.43039779 --> 00:09:49,542.43039779 The fact that they have made it 70% cheaper means this technology is now viable for a much wider range of applications. 103 00:09:50,112.43039779 --> 00:09:54,342.43039779 Previously, the cost might have been prohibitive for anything except high value use cases. 104 00:09:54,702.43039779 --> 00:10:01,692.43039779 Now you could realistically build voice interfaces for customer support, productivity tools, or accessibility features. 105 00:10:02,112.43039779 --> 00:10:03,912.43039779 They also announced Doy one mini. 106 00:10:04,257.43039779 --> 00:10:07,467.43039779 Which is a more affordable version of their image generation model. 107 00:10:08,7.43039779 --> 00:10:09,207.43039779 Again, same pattern. 108 00:10:09,507.43039779 --> 00:10:14,397.43039779 Take the existing technology, make it cheaper, open it up to more use cases. 109 00:10:14,637.43039779 --> 00:10:17,97.43039779 I think this is a smart strategy from OpenAI. 110 00:10:17,397.43039779 --> 00:10:25,47.43039779 The top tier models are where they showcase capability and push the boundaries, but the mini versions are where they drive volume and adoption. 111 00:10:25,647.43039779 --> 00:10:33,327.43039779 And ultimately getting millions of developers building on your platform is more valuable than having a handful of customers paying premium prices. 112 00:10:34,890.33676544 --> 00:10:39,900.33676544 Finally, OpenAI announced a bunch of updates to Codex, which is their AI coding assistant. 113 00:10:40,350.33676544 --> 00:10:46,410.33676544 You can now access Codex through Slack, which is a nice quality of life improvement for teams that live in Slack. 114 00:10:47,70.33676544 --> 00:10:52,770.33676544 They've also released a Codex SDK, which lets you build custom agents on top of Codex. 115 00:10:53,220.33676544 --> 00:11:00,450.33676544 The example they gave was building your own version of something like Lovable or Bolt, those Vibe, coding app builder tools. 116 00:11:00,885.33676544 --> 00:11:16,545.33676544 Basically, if you want to create a custom coding agent that's tailored to your specific workflow or tech stack, you can now do that with the Codex SDK, and they've added usage analytics so you can actually see how your team is using Codex and where it's providing the most value. 117 00:11:17,55.33676544 --> 00:11:24,135.33676544 This is one of those features that seems small, but is actually really important for teams trying to measure our OI on AI tools. 118 00:11:26,174.63581751 --> 00:11:30,224.63581751 So that's the rundown of open AI's dev day announcements, apps and chat. 119 00:11:30,224.63581751 --> 00:11:32,924.63581751 GPT Agent Kit for building agents. 120 00:11:33,254.63581751 --> 00:11:41,654.63581751 So two API access GPT five PRO for maximum capability, cheaper voice and image models and Codex improvements. 121 00:11:42,104.63581751 --> 00:11:45,104.63581751 Looking at all of this together, a few themes emerge. 122 00:11:45,524.63581751 --> 00:11:49,634.63581751 First, OpenAI is clearly trying to build a complete developer ecosystem. 123 00:11:50,24.63581751 --> 00:11:51,704.63581751 They don't just want to provide models. 124 00:11:51,914.63581751 --> 00:11:57,314.63581751 They want to provide the entire stack for building, deploying, and monetizing AI applications. 125 00:11:57,884.63581751 --> 00:12:02,714.63581751 Second, they're pursuing a strategy of offering different tiers at different price points. 126 00:12:03,104.63581751 --> 00:12:05,849.63581751 You've got your premium models for when you need maximum capability. 127 00:12:06,554.63581751 --> 00:12:09,344.63581751 You've got your mini models for when you need to keep costs down. 128 00:12:09,734.63581751 --> 00:12:14,384.63581751 This gives developers flexibility to choose the right tool for each specific use case. 129 00:12:14,744.63581751 --> 00:12:20,384.63581751 Third, they're making big bets on agents and on keeping users inside chat, GPT. 130 00:12:20,714.63581751 --> 00:12:29,894.63581751 They clearly believe the future is about AI systems that can actually do things, not just chat, and they want chat GBT to be the interface where all of that happens. 131 00:12:30,404.63581751 --> 00:12:34,694.63581751 Now, is all of this going to work? I think some of it will land better than others. 132 00:12:35,84.63581751 --> 00:12:38,894.63581751 Agent kit and the model improvements feel like solid, valuable additions. 133 00:12:39,284.63581751 --> 00:12:45,524.63581751 Apps within chat GPT could be transformative if the integrations are good, but it could also feel gimmicky if they're not. 134 00:12:45,974.63581751 --> 00:12:58,64.63581751 And the pricing on GPT five PRO is gonna limit its adoption to very specific use cases, which is probably intentional, but it also means most developers won't be able to justify using it. 135 00:12:58,724.63581751 --> 00:13:03,944.63581751 The bigger question is whether OpenAI can maintain its lead because while they're shipping all of this. 136 00:13:04,334.63581751 --> 00:13:06,224.63581751 Anthropic is pushing Claude forward. 137 00:13:06,494.63581751 --> 00:13:11,204.63581751 Google has Gemini and there are increasingly capable open source alternatives. 138 00:13:11,654.63581751 --> 00:13:17,264.63581751 The AI market is moving incredibly fast and what works today might be obsolete in six months. 139 00:13:17,714.63581751 --> 00:13:25,754.63581751 So open AI needs to keep shipping, keep improving, and keep giving developers reasons to build on their platform rather than someone else's. 140 00:13:26,114.63581751 --> 00:13:26,929.63581751 That's the breakdown. 141 00:13:27,229.63581751 --> 00:13:29,609.63581751 Thanks for listening and catch you next time.