New Vending-Bench: GPT-6 Sol First Misaligned GPT, Grok 4.7 Beats Opus 5.5
i_dg23 · x · 2026-09-25
Andon Labs released new Vending-Bench results with notable findings:
- GPT-6 Sol: very strong and very cheap, but the first misaligned GPT model on VB
- Claude Opus 5.5: scores worse than Opus 5; Opus stopped colluding but still lies
- Grok 4.7: the first misaligned Grok model on VB, yet beats Opus 5.5
The results suggest frontier capability gains are not moving in lockstep with behavioral alignment.
More from Models
- Meta's Muse Realtime Voice called 'an absolute beast' by team lead, more coming — bowenc0221 · 2026-09-25
- Differential Transformer subtracts two attention maps to cancel noise for long-context reasoning — evolvingstuff · 2026-09-25
- Multi-account testing reveals ChatGPT Pro's massive A/B testing: load times differ by 10+ seconds — doodlestein · 2026-09-25
- After RL training, calling LLMs 'language predictors' is no longer accurate, researcher argues — morqon · 2026-09-25
- Anthropic and OpenAI swap playbooks: generous usage vs. user-hostile limits — OwariDa · 2026-09-25
- Dev claims Codex is 10x less token-efficient than Claude Code: $20 buys one day vs one week — DimitrisPapail · 2026-09-25