You can see this in the Artificial Analysis benchmarks - GPT 6.1 Sol on medium thinking scores 10 points higher than DeepSeek 4.1, but is actually cheaper per task, as it only outputs 15 million tokens instead of 250 million.

Though for longer sessions I think DS4.1 would still come out cheaper... it's hard to beat that 98% cache discount

I thought this was very interesting - there's the map upfront, but you can also drill into individual countries and regions with some nice tables. Some "obvious" things here, though a few surprises:

* The solar potential of a gloomy, northern locale like the UK is "only" half of somewhere sunny like Australia. Though I can imagine the swings between seasons are greater.

* Northern Canada has a surprising amount of potential, compared to European regions at the same latitude

Canada compared to Mongolia of similar latitude, has lower potential though. Apparently that is because Mongolia is a desert and has no forests, very dry and high in elevation. The cold is no issue. The panels there produce 1.8 times more. The reason that Northern Europe is worse than Canada of the same latitude is apparently because of maritime proximity that increases humidity that diffuses light. Australia doesn't have maritime climate.

Cold actually increases PV output.

I wonder if the map accounts for this,

Without accounting for it, the maps wouldn't look like that. Cold, dry and high altitude increase the performance of solar.

Dry and high altitude increase the performance because those increase the amount of sunlight hitting the ground. The effect of temperature would be on top of those effects.

This happens with everything LLMs write - code comments, readmes, user facing UI text, etc. It just can't help but include every irrelevant detail from the context.

It's definitely gotten worse over time as well. Looking at older README files I generated using Opus/Sonnet 4.5, its actually quite shocking how clear and concise they are. Whereas eg, Opus 5 will write paragraphs and paragraphs about random things I prompted, or a partial solution that was then scrapped for something else.

My theory is that the increasing strength in performing long-context tasks has led to this particular quirk.

>It just can't help but include every irrelevant detail from the context.

The worst is when it walked through a dozen or so questions, and it decides to litter comments with "Per answer 7, ..." with no further context.

I have in my Claude.md that comments must be diegetic, if needed at all. This seems to have helped. And it is then something which gets checked by agents in code review.

Codex does this too, I've noticed. LLMs seem to love documenting irrelevant implementation details in what should be user-facing text. I guess because the context is overwhelmingly the technical side of things, so without a later directive to steer it, that's what's most likely to come out.

The user-facing string stuff is definitely the thing that drives me the most nuts. I guess it's nice that it defaults to gag-inducingly transparent, but I don't think the backend structure should be written out in the title element of the app.

What' especially annoying is when you as a code reviewer have to go through tons of such comments that keep distracting you because the author could not bother reading those comments and assessing that they contain references to prompts / instructions that are completely irrelevant to he target audience