Here are three things I found interesting in the world of AI in the last week.

Opus 4.8 finally admits when it isn't sure. Maybe.

Anthropic shipped Claude Opus 4.8 this week. I reckon it's a rushed response to problems in the 4.7 release, but they've also tried to fix Claude's most annoying habit.

I know a lot of devs who started switching to codex after the 4.7 release. Despite strong benchmarks it felt lazier. It wouldn't follow instructions properly, it would skip over steps in workflows and say things were done when they weren't. A big uptick in 'I'm sorry....' interactions on workflows that were decent under 4.6.

Anthropic says 4.8 is roughly 4x less likely than 4.7 to wave through flaws in code it wrote itself, and over 10x less likely to deliver a wrong answer with confidence. The system card claims it never once passed off a flawed result as fine in their honesty test, which they reckon no earlier Claude has managed.

I've been running it non stop since release and I dunno, maybe it's working. There have been a few instances where I felt like it was more cautious, but a few days isn't long enough to be sure. I'm damn glad they are trying though.

The worst part of this release is I that the rate of Claude using "here's my honest take" seems to have gone up. And it was pretty bad before. Meh. I guess I can live with it if it's actually more honest.

Google is closing its open-source CLI and people are not happy

At I/O last week Google announced that on June 18 it stops serving individual free, Pro and Ultra users on its open-source Gemini CLI and moves them to a closed-source replacement, the Antigravity CLI.

Hundreds of people sent Google free work on an open codebase, and the individual tier they built it for is the getting cut off. A contributor who'd just landed a 27-commit pull request the day of the announcement asked whether they'd been "essentially working for free on a code base that will only be used in enterprises." Hard to argue with that logic.

None of this is unique to Google. Claude Code keeps its harness closed too, shipping the real tool as a binary while the public repo is mostly docs. The difference is Anthropic never invited anyone to contribute, so there was no community left holding the bag.

As for the new tool, it's alright. Google released an update to desktop Antigravity at the same time and it's pretty comparable to Codex. It's definitely a bit raw and I find it frustrating to bump into bugs and not be able to inspect the source code or submit a PR.

I've also found that gemini wants quite different prompts and if I just give it my claude and openai ones it does stupid stuff. So take small steps and run lots of prompt retros if you're going to switch over existing workflows.

NZ green-lit its first AI factory, and it's a good fit for the country

Closer to home, New Zealand consented Datagrid's Southland campus near Invercargill a while back. I was pretty sceptical of it because of all the US data center shenanigans, but after digging in it looks pretty good to me.

It's a roughly $3.5 billion, 280MW build, and at full size it would be the country's second-largest electricity user after the Tiwai Point smelter, about 6% of national demand, with most of the compute sold to overseas customers down a new trans-Tasman cable. Ugh. But it looks like the long term contracts are being used to underwrite new, renewable generation.

A 15-year, 140MW agreement with Mercury works as an anchor customer that pays for new wind to get built, not a fight over the cheap hydro we already have. That's a real difference from the old Tiwai arrangement, which leaned on existing supply and reportedly added about $200 a year to household bills. And the compute runs near-zero-emissions on Southland renewables, where the same workloads would burn far more carbon almost anywhere else on the planet. I hope they get a premium for those sweet green tokens.

A lot of people are concenerd about water. The scare stories come from places like Phoenix and Las Vegas, deserts running huge volumes through evaporative cooling. Southland is the opposite. Invercargill gets about 1,277mm of rain a year (Phoenix gets around 200) and sits near 10 degrees. Datagrid's consented groundwater take is 220 million litres a year. It's about 8.5% of an under-allocated local zone, roughly seven dairy farms' worth. So, not nothing, but a much better scenario than in the US.

It isn't all upside. Once the 1,200-strong construction crew moves on it's about 50 permanent jobs and $60 million of GDP a year, a small low-value wetland on the site gets removed (with mitigation required), and the customers and most of the profit sit offshore. But the deal funds its own new generation, the environmental footprint is modest for the location, and ownership actually came home last year, NZ is now the largest shareholding bloc. For once the AI infrastructure story has us on the right side of it.

I get a lot of queries about data sovereignty and having frontier models in NZ would be a big boost to folks concerned about that.