[Bug] Claude Opus 4.8 High: Reasoning accuracy degradation on context window quantification
Bug Description
On Opus 4.8 high. Get feeling that high is slow (understandably), what concerns me i have feeling it's kinda more stupid than Opus4.8 med (don't claiming that one is stupid). Think there's still some problem. Brief: where comprehension broke down (for your bug/feedback report)
┌─────┬──────────────────────────────────────────────────────┬─────────────────────────────────────────────┬────────────────────────────────────────────┐
│ # │ What happened │ Root cause │ Bug or inherent? │
├─────┼──────────────────────────────────────────────────────┼─────────────────────────────────────────────┼────────────────────────────────────────────┤
│ │ I claimed "on a 1M model at 150K you're not │ I used "% of window" as a health proxy, │ Bug-ish (model reasoning). More deliberate │
│ 1 │ unhealthy." You corrected me — │ silently conflating two distinct axes │ thinking catches this — the kind of thing │
│ │ attention/token-to-token load is absolute, a bigger │ (headroom vs recall quality). A genuine │ higher reasoning effort / "ultrathink" │
│ │ cap doesn't refresh recall. │ reasoning miss. │ exists for. Worth filing. │
├─────┼──────────────────────────────────────────────────────┼─────────────────────────────────────────────┼────────────────────────────────────────────┤
│ │ │ My AUQ options ("% of 1M vs % of 200K") │ Bug-ish. Lesson encoded: when an AUQ keeps │
│ 2 │ The 1M-coloring question took 3 AUQ rounds; your │ were built on the flawed model from #1, so │ missing, the framing is wrong — jump to │
│ │ answers didn't map onto my options. │ the right answer wasn't even an offered │ concrete previews sooner (which is what │
│ │ │ option. │ finally worked). │
├─────┼──────────────────────────────────────────────────────┼─────────────────────────────────────────────┼────────────────────────────────────────────┤
│ │ │ Markdown chat can't render ANSI color, so I │ Inherent medium limit (chat ≠ terminal). │
│ 3 │ You had to say "no emoji, use ASCII blocks ▓/░" │ reached for emoji to show hue — but you │ Mildly my fault for not using labeled │
│ │ twice. │ care about the literal artifact, not the │ blocks from the start. Low-value bug. │
│ │ │ chat's convenience. │ │
├─────┼──────────────────────────────────────────────────────┼─────────────────────────────────────────────┼────────────────────────────────────────────┤
│ │ You caught that 5 cells/band after compaction washes │ I under-specced legibility after compaction │ Not a bug — normal design iteration; you │
│ 4 │ out the threshold colors; then simplified my │ and over-engineered the fix. │ improved it. │
│ │ 2-width fix to one BAR_WIDTH=20. │ │ │
└─────┴──────────────────────────────────────────────────────┴─────────────────────────────────────────────┴────────────────────────────────────────────┘
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗