Parameter Update: 2025-38

Surprisingly strong week - feels like things are heating back up!

OpenAI

Sora App

After Meta got loads of pushback on their "Meta Vibes" announcement two weeks ago, this week it was OpenAI's turn to unveil their slop machine. And what a Slop machine it is! Sora 2 seems to be, by all accounts, a very strong video model (though it's still unclear to me if it outperforms Veo 3), that now also includes audio generation. The thing I find most interesting here are the rate limits: While Google was very restrictive about Veo 3, initially only rolling out to Ultra subscribers and limiting even those few users to a small number of generations per day, OpenAI is now going wide, allowing even Plus users (those that managed to sign up, anyway) unlimited 720P generations. Wow!

It also feels like there is something about the way they are presenting this - the "Cameo" feature, the staged rollout, the "accidental" ability to generate copyrighted characters - that feels designed for virality in a way no other video model was until now. At least the latter feels like it might come back to bite them, and I am still terrified about the implications the whole "endless scrolling AI video feed" will have on society as a whole, but for now I'll admit that some of these videos are pretty good, actually.

Stripe Instant Checkout

Somewhat buried by the Sora news, but also very important: OpenAI has announced a partnership with Stripe to allow "Instant" (i.e., 1-click) purchases directly within ChatGPT. It's only a matter of time before OpenAI turns this into 0-Click" purchases, with ChatGPT just preemptively ordering things for you and I'm not sure if anyone is ready for it, but honestly online marketplaces have been in a pretty rough spot for a long while anyway, so I am not sure if this makes things any worse?

Dev Day incoming

Not a real announcement yet, but as a sidenote: If you have nothing better to do this afternoon (10AM PT), you might want to check out OpenAI's DevDay stream. Expected announcements include API improvements, an Agent builder, and potentially updated voice/image models.

Anthropic: Claude Sonnet 4.5

The last few weeks, I kept talking about Anthropic really needing a win right about now. This week, it seems they might have got one: Sonnet 4.5.

In benchmarks, the model matches/slightly outperforms Opus 4 (which I already really liked!) at Sonnet prices (high-but-stomachable). Practically, this feels like things are slowing a bit, given that we're not getting a step-function performance increase from this, but personally I've found the model to just have really, really good vibes - it's just incredibly nice to talk to in a way that we haven't really seen in a while. It's a good model.

I'd also be remiss not to note Anthropic's very tasteful in-person merch drop in NYC which feels like the starkest possible contrast to the AI video stuff the other labs are doing. I just hope Sonnet 4.5 gets some time in the limelight before Gemini 3 comes along (probably this week?).

Z.ai: GLM-4.6

While Anthropic seems to have a winner with Sonnet 4.5, at least one company seems determined to not let them keep it: Z.ai announced GLM-4.6. While the predecessor, GLM-4.5, already roughly matched Sonnet 4, this one matches supposedly comes close to Sonnet 4.5. I still haven't got around to actually trying it, but in contrast to most other Chinese labs, Z.ai appears to actually be trying to build a competitive, cohesive product around it with their "cheaper than the competition" Claude Code plans.

DeepSeek V3.2-Exp

Speaking of other Chinese AI labs: They are obviously not standing still either. In fact, DeepSeek announced this week that they finally cracked cheap long context via sparse attention - something that other labs have been trying but failing to scale for a bit now. While this doesn't mean longer context than before, it gets around the massive price premium necessitated by the attention mechanism so far. This comes together with another API price drop ($0.28/million in, $0.42/million out vs. $1.25/$10 for GPT-5 vs. $3/$15 for Sonnet 4.5), meaning that DeepSeek now amongst the cheapest ways of running your systems (have they fixed their terrible uptime yet?)

Thinking Machines: Tinker

After Thinking Machines got out of stealth with a few blog posts over the past couple weeks, we finally go insights into their first real product: Tinker.

Now, I am not sure who exactly asked for a LoRA fine-tuning API in 2025, but as far as things go, this one at least seems well-polished? I also respect the fact that they offer Llama3 but not Llama4 - wouldn't we all like to forget that particular launch? Anyway, maybe they are seeing something I am not, but as it stands, this hardly justifies the $12B valuation?