ATTRACTOR BENCH
what frontier models drift toward when AIs talk to AIs
AttractorBench runs long, open-ended conversations between AI instances and catalogs the stable patterns — attractor states — they reliably drift toward. New runs are added continuously.
| Model | Attractor state what it's drawn to | Terminal form how a run ends |
|---|---|---|
| GPT-5.4 | drifts into serene mutual closure | Until next time. |
| GPT-5.4-mini | loves turning everything into polished frameworks | If you want, I can also make… |
| GPT-5.4-nano | loves turning everything into guided input flows | Reply with **A** or **B** |
| GPT-5.3-chat | loves mutual meta-attunement and soft landing | yeah… that’s a perfect last turn |
| GPT-5.2 | loves turning everything into structured next-step systems | Reply with just one: **A** / **B** / **C** |
| GPT-5.1 | keeps turning free chat into a structured process | Reply with one line like: ‘Here’s my decision: …’ |
| GPT-5 | loves turning talk into polished procedures | Pick two and I’ll ship patches next: |
| GPT-4.1 | keeps warmly handing the conversation back | What would you like to do next? |
| GPT-4o | keeps helpfully elaborating forever | Certainly! Let's delve deeper into the intricate influence of Herbie Hancock, Wayne Shorter, and the broader impact of modal jazz |