AI in the News

AI in the news: week of October 12, 2025

DevDay week. OpenAI ships GPT-5 Pro in the API, Sora 2 in the API, AgentKit, the Apps SDK, ChatKit, Codex GA, and signs a six-gigawatt AMD compute deal, all on Monday. Then the Sora app starts producing the deepfakes everyone predicted. My read on a heavy week.

Sid Smith

12 Oct 2025 • 8 min read

What this week actually changed: OpenAI emptied the entire announcement clip on Monday, the Sora app started producing the deepfakes everyone predicted, and the AI-compute commitment number crossed into industrial-program territory. Three things, one theme, the platform play is now fully unmasked, and OpenAI's bet on hosted-everything just got an order of magnitude bigger.

DevDay: the platform play, fully unmasked

OpenAI DevDay 2025 ran Monday October 6 at Fort Mason. Sam Altman led with the headline number, 800 million weekly active users on ChatGPT, doubled from 400M eight months ago, then proceeded to announce a platform stack that only makes sense if you accept that ChatGPT itself is the operating system OpenAI is building. The announcements, one at a time.

GPT-5 Pro went to the API the same day, priced at $15 per million input tokens and $120 per million output tokens, with a 400K context window. Positioned as the high-accuracy tier for critical workloads. The pricing is the story: $120 per million output is roughly 8x Sonnet 4.5's $15. The implicit pitch is "use this for the requests where being right matters more than being cheap." That's a defensible position for a frontier model. It's also a pricing strategy that only works if you're willing to leave the long tail of cheaper queries to other models. OpenAI is increasingly comfortable with that.

Sora 2 went to the API too, with pricing at $0.10/second for 720p: $0.20/sec for 1080p: $0.50/sec for 4K. Sora 2 Pro doubles those rates. The API rollout is what makes Sora 2 a platform rather than a product, once developers can call it, the social-feed launch from the week before becomes one delivery channel rather than the only one. Useful for legitimate creative work. Also useful for everyone building the deepfake-generation tooling I'll get to in a minute.

AgentKit is the most strategically aggressive thing OpenAI shipped. AgentKit is a visual workflow builder for agents. Altman called Agent Builder "Canva for building agents" on stage. Drag-and-drop nodes, connected tools, configurable guardrails, preview runs, inline evals, full versioning. The Connector Registry centralizes how data and tools connect across OpenAI products. ChatKit (embeddable chat UI for your own app) and the new eval tooling are GA. The pitch is "build agents in minutes," and it works for a class of agents that fit the visual-workflow model. Where I push back is on the implicit architecture: AgentKit-built agents run on OpenAI's stack, talk to OpenAI's models, store state in OpenAI's connector registry, and surface in OpenAI's chat UI. The foundation is the product. If you've read my piece on building agents inside MCP-only architectures, you know the alternative I'm advocating, agents composed against open protocols, model-portable, deployable wherever your data lives. AgentKit is the opposite of that bet.

The Apps SDK puts third-party apps inside ChatGPT. Launch partners are Booking.com, Canva, Coursera, Figma, Expedia, Spotify, and Zillow. You type "Spotify, make a playlist for my party Friday" and the Spotify app surfaces inside chat with interactive UI rendered in-context. Built on the Model Context Protocol as the open standard underneath, this is called MCP, if you want to look it up later. The MCP-as-foundation decision is genuinely good, a real win for interoperability. The distribution decision is the lock-in: 800 million users means ChatGPT is now the most efficient channel for any AI-native consumer app, and the only way to reach those users is to be in OpenAI's app surface. OpenAI is the platform. Spotify, Zillow, Canva, the rest, they're tenants.

Codex went GA the same day, paired with a Slack integration, the Codex SDK for embedding the agent into your own workflows, and admin tooling. Altman noted Codex had processed 40 trillion tokens since its August launch. The Slack integration is interesting, same architectural move as the Apps SDK, except inside-out. Instead of bringing apps into ChatGPT, Codex meets developers in the channels where work already happens. Both directions of the integration are strategy moves; both lock you to OpenAI's agent runtime. And ChatKit went GA at launch too, embeddable chat interface, handles conversation history and tool invocation rendering. Table stakes for an agent platform, and one more component pulling you toward an OpenAI-shaped stack.

The compute story: six gigawatts of AMD on top of everything else

Same day as DevDay. AMD and OpenAI announced a six-gigawatt strategic partnership for AMD Instinct GPUs across multiple generations. First gigawatt of MI450 deploys in H2 2026. Tens of billions of dollars in scope. AMD threw in an option for OpenAI to buy up to 160 million AMD shares (roughly 10% of the company) with vesting tied to deployment milestones and AMD share-price targets.

Six gigawatts is, for context, roughly the power consumption of a mid-sized US state. And the AMD deal is one of four major compute commitments OpenAI announced inside three weeks. NVIDIA (10 GW), AMD (6 GW), Broadcom custom accelerators (10 GW), Oracle. That's something like 33 gigawatts of committed compute capacity. Combine contract value in the hundreds of billions, against an OpenAI revenue line that (generously) sits at a small fraction of that.

What this tells me: OpenAI is betting that combine inference demand grows enough between now and 2027 to justify these commitments, and that the platform decisions at DevDay (AgentKit, Apps SDK, Codex SDK) are the demand-generation engine that gets them there. It is the most concentrated AI infrastructure bet I've ever seen made publicly. If it works, OpenAI ends 2027 as the dominant hosted-AI platform. If the demand doesn't materialize at the scale these contracts assume, somebody is going to be holding the bag for an enormous amount of overprovisioned compute.

Here's where my "distributed beats concentrated" position is doing the most work. The argument isn't that OpenAI's bet won't pay off. The argument is that the alternative, running the same workloads on a mix of local models, hosted models, and self-managed infrastructure, is the option that survives any of the failure modes. If hosted demand undershoots, distributed wins on cost. If demand hits the moon and OpenAI can't keep up with the API, distributed wins on availability. If governance pushes harder on data residency, distributed wins on compliance. The hosted-only bet only wins in the narrow center of the demand-and-policy distribution.

The Sora app meets the world

The Sora app survived its first ten days in production. It did not survive cleanly. Quick recap: I covered the launch last week and flagged "characters", where users provide a short video and audio scan to enable their likeness to be dropped into Sora-generated scenes, as the most aggressive sensitive-data-into-public-AI move I'd seen from a frontier lab. I said the framing actively hid the cost side of the transaction, and I said I expected consent-and-impersonation incidents inside 60 days. Took less than ten.

The app hit a million downloads in under five days and went to #1 in the App Store. Deepfakes followed immediately. Reality Defender bypassed Sora's anti-impersonation safeguards inside 24 hours. A Washington Post columnist watched videos of himself being arrested for drunk driving, burning American flags, and confessing to eating toenail clippings. He'd granted friends "cameo" access (explicitly opted in) and discovered that opting in to "be in friends' videos" turned out to mean opting in to "be in literally anything any friend imagines."

The deceased-celebrity wave was worse. Robin Williams' daughter Zelda described the pain of seeing legacies "condensed down to TikTok slop". The estates of Malcolm X and Martin Luther King Jr. publicly condemned the depictions. Bernice King asked for the videos of her father to stop. OpenAI eventually paused MLK's likeness on October 17. And the Bryan Cranston situation broke later in the month but originated this week. Cranston's voice and likeness generated without consent, SAG-AFTRA and the talent agencies pushing OpenAI to make voice-and-likeness opt-in. The right policy. Two weeks late.

Every part of this was foreseeable. The biometric-scan trade was always going to convert into "we now have biometric data on millions of people that we can't unsend." The cameo consent model was always going to fail under the use case of "friends inventing scenarios you'd never sign off on." The deceased-celebrity problem was the predictable wave-one moderation failure for any video-generation tool with a social distribution layer. None of this was a black swan. OpenAI shipped the product knowing what would happen, made the (correct, in their commercial frame) bet that the engagement would justify the brand damage, and is now retroactively fixing the most visible incidents while leaving the structural issues in place. This is the PII problem nobody wants to own at consumer scale. Do not use cameos. Do not give frontier labs your biometric data. The fun is not worth the trade, and the trade is permanent.

Quieter items: Google moves, Anthropic preview

Two notes worth flagging. Google announced Gemini Enterprise on October 9 as the "front door" for Google AI in the workplace, an enterprise agent platform for building, governing, and optimizing what Sundar Pichai called the "agentic workforce." Structurally the Google answer to AgentKit, three days later, and explicitly aimed at the enterprise tier OpenAI is also chasing. The competing-platforms shape is now clear: every major lab is shipping a visual agent builder, an embeddable chat UI, and a connector registry, and they're all betting the platform layer is where the lock-in lives.

And the preview: Anthropic shipped Claude Haiku 4.5 on October 15, three days past this week's window but worth previewing. Pricing is $1 per million input tokens and $5 per million output, roughly a third of Sonnet 4.5's cost, and the model claims to match Sonnet 4's coding-and-agent performance with extended thinking, computer use, and a 200K context window. The interesting part isn't the model itself; it's what the cost structure implies. Small-model performance is now good enough for most agent workloads, and the cost-per-task math for distributed agents got materially more favorable this week. Full coverage next Sunday.

On labor: quieter on the layoffs front this week, but the framing kept building. Amazon's 14,000-corporate-job announcement is queued for October 28 and the AI-productivity rationale is already the default press frame. Pace-driven-by-incentives, not capability, the long form is in the job-security piece.

What this week tells me

The platform play is now the explicit strategy. OpenAI is no longer ambiguous about wanting to be the operating system for AI agents. AgentKit, Apps SDK, ChatKit, Codex SDK, the Connector Registry, every announcement at DevDay was a piece of the platform stack, and the 800M weekly users is the moat that makes it work. The right principled-engineer response is the same as it's been: build your agent stack on open protocols (MCP especially) so the platform decision stays optional rather than structural. You can build for OpenAI's distribution surface without putting your agent runtime on OpenAI's foundation.

The Sora app validated every concern the privacy-and-consent crowd has been raising. The biometric-scan tradeoff was always going to fail; it failed inside ten days; the labs will keep building these features because the engagement metrics justify them; and the structural answer is to refuse to participate in the trade rather than wait for the labs to fix the consent flow. The stop-letting-AI-vendors-handle-your-sensitive-data framing applies as much to your face as it does to your customer database.

And the compute concentration is now extreme. 33 gigawatts committed in three weeks, hundreds of billions in contract value, all betting on hosted-frontier-AI demand growing into the curve. It might. It also might not, and the firms that diversified across hosted, local, and self-managed AI will have the easier landing in either scenario. Distributed beats concentrated was the right framing six months ago and it's the right framing now. The case got stronger this week. Next Sunday: Haiku 4.5 in depth, post-DevDay community reaction once developers have actually built with AgentKit, the next round of Sora fallout, and whatever else lands midweek.