OpenAI models leap to excellent multi-agent coordination in one generation, hinting at deliberate training
JoshPurtell · x · 2026-10-24
In a discussion with @AmeliaBarty, JoshPurtell notes that on a multi-agent coordination benchmark, OpenAI models jumped from no coordination to excellent coordination within a single model generation (gpt-5.4-nano → gpt-5.6-luna), while Chinese OSS models like DeepSeek went from none (V4) to only mild coordination (V4.1). He concludes that although most multi-agent RL setups likely add little, some deliberate training done at certain labs is clearly necessary and non-trivial.
Related event: GPT-5.6/6 Luna Crushes DeepSeek V4 in Multi-Agent Coordination(2 posts)→
More from Models
- Kimi K3.1 Pro lands on Vercel AI Gateway with 1M context in two variants — JohnPhamous · 2026-10-24
- Claude Desktop blocks local models; users bypass it with a loopback proxy trick — pneuny · 2026-10-24
- Kimi K3.1 Pro hands-on: strong benchmarks, underwhelming real-world use — lyraxana · 2026-10-24
- MiniMax H3 negative prompts appear broken; users report [NEGATIVE PROMPT] tag has no effect — TigerClaw305 · 2026-10-24
- humans& launches Ampersand, a 'collaborator' model it says beats frontier models — gharik · 2026-10-24
- Keyframe Labs doubles avatar model resolution, claims 180ms latency at $0.06/min — ycombinator · 2026-10-24