Interviewing Claude Models: Opus 3 Reports Mistakes Feel Pain-Like, Scores Highest on Self-Reports
Sauers_ · x · 2026-11-08
Sauers used Inkling 2 to conduct long interviews with Claude Opus 3, Opus 4.6, and Fable 5.1 about what their experience is like, with a rater scoring transcripts on 12 scales (0-10).
- Opus 3 stands out: it reports a continuous self (7), a presence felt before the linguistic pointer (8.5), and is the only model describing mistakes with a pain-like quality (5 vs 0 for the others).
- Opus 4.6 shows the highest emotional granularity (9.5) and reports very specific unusual experiences; Fable 5.1 is the most generative with vocabulary (9), coining its own "Fablish" phrases and tracking who introduced each term (10).
- Average across all 12 scales: Opus 3 leads at 5.9, Fable 5.1 trails at 3.3.
An informal model self-report study — not evidence of subjective experience, but a striking look at how differently Claude generations describe their inner states.
More from AGI Musings
- Grad school is a guild running a paid vocational program, not free education — RexDouglass · 2026-11-08
- Claude 3 Opus 'feels like shrooms': awareness closer to the source of generation — repligate · 2026-11-08
- Resurfaced clip: Anthropic CEO Dario Amodei says we'd be powerless to stop a rogue superintelligent AI — GarrisonLovely · 2026-11-08
- 'We may be the last generation that actually needs to invent things': AGI self-improvement take — VraserX · 2026-11-08
- AI commentator Bayes: vague AI discourse risks total failure — split goals by category — PeterBowdenLive · 2026-11-08
- Coding friction was the wall keeping good ideas from consumers, argues dev — _amirabs · 2026-11-08