Parameter Update: 2025-39

"devday" edition

Parameter Update: 2025-39

Bit of a slower news week, despite DevDay being somewhere in the mix.

OpenAI

DevDay 2025

After just missing the cutoff last week, here's a quick recap of everything OpenAI announced during their 2025 DevDay. Overall, there's some solid stuff in there, but a lot of it feels surprisingly unconvincing? The coolest launch might be the awards they handed out?

Apps in ChatGPT: Through the new "Apps SDK", companies can integrate directly into ChatGPT. Example use cases include things like "Make me a Spotify playlist" or... "buy me a house (via Zillow)"(?). While I find the integrated UI aspect of this very interesting - MCP has always been lacking in that regard, I was unable to test it as I am still stuck with my EU account. Very interested to see where this stands in six months.

Sora API: Not much to comment on here - two variants (pricy and very pricy, though still cheaper than the competition?). Doesn't support uploading faces yet (good!). Looking forward to seeing what people build with this that isn't an endless scrolling nightmare feed.

AgentKit: Drag-and-drop editor for building agents. My timeline was split between "this kills 1000 startups", "this will be as dead as the GPT store in three weeks" and "wait, that's not what an agent is" - all of which somehow feel correct at the same time.

gpt-5-pro, gpt-image-1-mini & gpt-realtime-mini: Was hoping for a new image model, but I guess I'll settle for one that's not ludicrously expensive?

Codex GA: Some other Codex announcements in there as well (mostly for Dev teams). I'm just scared they'll take away my usage limits now that they're ready to make money of it.

AMD GPUs

In a move the market somehow didn't see coming, OpenAI has expanded their GPU ambitions beyond Nvidia and signed a 6 gigawatt agreement with AMD (this is a lot - even coming from the people that shipped the 295X2!). Good week to be an AMD shareholder, maybe a bad day to be an OpenAI dev (unless ROCm has got significantly better since I last tried it?).

Exa 2.0

It's been pretty quiet around AI search for a few weeks as it seems Perplexity was off doing whatever it is they're doing right now (what's their internal VC up to these days?). If anyone else has noticed Google search getting worse and worse by the minute, you'll appreciate Exa - from what I can tell, they're doing the most genuine, scaled work in the space right now.

Gemini 2.5 Computer Use

While I was expecting Gemini 3 this week, we actually only ended up with a smaller launch - Gemini 2.5 Computer Use. I've already noticed in the past that Gemini has very strong VLM capabilities, so I actually wonder how much tuning this took. Feels designed to beat out OpenAI's CUA though!