minimaxir

Show Hacks:

I've been holding off because I will need to do experimentation with local models at some point...but not yet and won't be for a bit but I am afraid prices will rise given component cost trajectories.

I've been tempted to just get the M5 Max instead and just innovate with smaller models myself.

I've had my eye on the normal Mac Studio M5 Ultra (normal plebeian 96GB RAM) and the release (edit: ship) date there keeps slipping, now into January. I'm surprised Apple is pushing the backlog even more.

> Sidenote: I should also add that you can see this is an anomaly when you use Anthropic’s own “dollar of revenue per compute” number. For Q2 2026, it brought in $2 of revenue per $1 of compute spend, up from $1.39 in Q1 2026. This number dropped to $1.70 in Q3. I think it’s a stupid number as it doesn’t really say very much about the business, but even based on Amodeiconomics, this quarter feels like the exception rather than the rule.

...making money is a pretty important part of the business?

Is this the first time VendingBench was used as a serious benchmark result in a model release? That's kinda funny.

This bullet point is funny:

> Draw. Sketch and annotate directly onscreen.

Their demo example for the touch interface is piano keys? That's a 2010 Scott Forstall level of skeuomorphism that doesn't translate well for a laptop screen because it's not convenient.

Similar 75% price cut trajectory as GPT-5.6 Sol -> GPT-6.1 Sol. Overall, Terra is still half the price of Sol.

But like the original release of Terra, it's still a weird middle child between the small-model-that-got-substantially-better (Luna) and the large-model-that-got-significiantly-cheaper (Sol). The TerminalBench 4.0 cost-per-task results are a bit confusing and I'm not sure why you would pick Terra over Sol in that case. Overall this release seems like a flex on Sonnet 5.5: a model which is having its own middle child problems between the godly Opus 5.5 and the surprisingly capable Haiku 5.5.

For me, I used Zed Agent with Codex as it has native integration with Zed's ACP so it's pretty convenient.

However, Anthropic is pickier about the harnesses you can use with Claude Code and I noticed discrepancies when using Zed's External Agent feature with Claude Code, so I just use the Claude app. It's fine.

Calling your reply aggressive isn't against the rules, bad faith or not (multiple people on HN clearly think it is aggressive). What is in the Guidelines, however:

> Be kind. Don't be snarky. Converse curiously; don't cross-examine. Edit out swipes.

Simon's comment apparently got flagkilled; I vouched for it.

If enough people are flagging it, we may be at a point where the pelican benchmark is truly overdone.