Parameter Update: 2026-20

All things considered, it's been a surprisingly slow week. We were also expecting a new GPT model this week, but it appears it got pushed a week? Anyway, two more weeks until WWDC, which should be fun.

Anthropic

Claude Opus 4.8

Anthropic keeps pushing out incremental model updates for Opus. This week brought version 4.8, which is designed to be more honest than 4.7 and capable of longer horizon work. Perhaps more interesting than the single-digit gains in the majority of benchmarks, are some of the other news around the model drop:

• They made Fast Mode 3x cheaper (still only available at API pricing, not included in the Claude subscription though)
• "Workflows" in Claude Code, which allow the model to write orchestration scripts for subagents (playing into the ability to keep on task for longer-running work)

On my timeline, I've seen people make fun of how workflow specifically will absolutely wreck your usage limits (which makes sense), while the fast mode price drop was almost universally beloved. As for the model itself: In my experience, the main issue with Opus 4.7 wasn't actually dishonestly, it was the ability to realistically understand it's own limitations, so fingers crossed this might also be addressed.

Big funding round

After the compute expansion news of the past couple weeks, it's no wonder Anthropic would be looking for cash to keep the game running. This week, they announced (as far as I can tell) the highest valuation private financing round of all time - 65 Billion at almost $1 Trillion valuation.

People more interested in financial markets might want to argue if this represents a failure of the IPO process, but I'll leave it at:

It's also hilarious that this and the increases in Anthropic's compute capacity through the xAI deal mean that the Mythos "security concerns" are now starting to disappear, with Anthropic targeting a rollout "in the coming weeks". Irony aside, I am excited that we'll finally get the model, but remain concerned about what pricing for it might look like.

Collaboration (?) with the Pope

In what is probably the most unexpected collab of the past couple weeks, the pope has released an extremely long text discussing the implications of AI on society - partially supported by Anthropic.

Given the text was quite critical at times, some of the responses have been hilarious.

While I have seen some good discourse around it, the most hilarious finding is that some AI detection models actually flagged portions of the text as being AI generated. I also loved people arguing if this text means they can get a religious exemption from being forced to use AI at work?

Nvidia RTX Spark

Nvidia officially entered the consumer CPU space at Computex this week, announcing the RTX Spark (previously floating around under the N1/N1X codenames). This isn't Nvidia's first CPU rodeo (remember Tegra?) and the chip is essentially a consumer/OEM-tuned variant of the same Grace Blackwell silicon behind the existing GB10 / DGX Spark, rather than anything genuinely new. It pairs a 20-core Arm-based Grace CPU (co-developed with MediaTek) with a Blackwell RTX GPU, an NPU, and up to 128GB of unified memory, with Nvidia pitching ~1 petaFLOP of FP4 AI performance and the usual full-stack treatment (CUDA, DLSS 4.5, Reflex, G-SYNC, ray tracing). It's bound to make a splash in the Arm PC market and put Qualcomm (and now Apple and AMD/Intel) under renewed pressure.

For a company looking to justify its increasingly ludicrous valuation, "officially" planting a flag in the CPU space is a smart move, even if it's mostly in the interest of vertical integration. Jensen leaned all the way in, reportedly quipping that Nvidia "used to be a GPU company" and framing this as reinventing the PC "for the era of personal AI agents".

As for how this actually shakes out: the splashy launch and big claims arrived right on schedule. Pricing wasn't announced, but the partner lineup tells the story — ASUS ProArt, Dell XPS, HP Omnibook, Lenovo Yoga, MSI Prestige, Microsoft Surface Ultra — so expect this to land firmly on the "expensive" side of things. Windows on ARM remains the open question: Nvidia and Microsoft are pushing native ports and a new agent runtime ("OpenShell"), but anything not natively ported still leans on Prism emulation, which I expect to be rough around the edges at best based on past experience. With systems not shipping until fall, availability remains the other open question - what good do these systems do if most people can't afford them, and those that can't get one?

Nevertheless, I'll be cautiously optimistic that, at the very least, this might encourage all other players to lock in moving forward.