opus 5.5 is #1 in all six of artificial analysis's new job indexes. same week people say it feels nerfed
so i checked what's in them:
coding is 15% of one index (engineering). the other five have zero
most of every score is agent office work, knowledge recall and reasoning
v2 also leaned harder on tool use, and the scores are opus at max effort
matches my setup tbh. fable's been my lead since oct 15, opus still runs my tool-heavy, low-judgment orchestration
a #1 here says great tool user. it doesn't say better coder
1
57
83% of my claude code startup context is written by anthropic, not me
12.4k tokens before i type anything. 10.3k is their system prompt + tools. 2.1k is my memory files, 3 skills and 1 agent
last week they added two sentences to the opus 5.5 prompt: no recreating someone else's game, even renamed
it came down from their servers. nothing to update on my end and no setting to turn it off
the 17% that's mine is the only part i get to steer, so i keep it tight
when's the last time you ran /context?
6
3
122
openai is doing what claude already does: from december your chats and codex eat from one bucket
web, desktop and claude code have always shared one pool
openai now adds thinking-model chat to the codex pool. only instant chat stays free
neither one gives you a real number. openai says "expanded", claude says 5x or 10x of pro
my $100 openai sub is basically a codex review sub (astra + sol). so every long thinking chat now costs a review
the shared pool isn't the problem. not knowing how big it is, is
4
1
80
You can't trust benchmarks.
When Opus 5 dropped and was #1 on everything but broke almost everything it touched, I knew the numbers were pointless.
Always do your own testing. (Opus 5.5 is good tho)
xitter.gq/Cutzilla/status/210538…











