Ampersand hits 92.2% on SWE-bench Pro v2, 70% cheaper via human-AI co-coding
gharik · x · 2026-11-06
Humans & launched Ampersand, a model trained to collaborate with humans, plus a multi-agent co-coding mode. Benchmarks:
- SWE-bench Pro v2 HARD-51: 92.2%, matching Fable 5.1 at 70% lower cost, beating Gemini 4 Argon, Astra, Muse Wonder and Kimi K3
- DeepSWE v1.1: 77.9%, on par with Gemini 4 Argon, ahead of Astra, Grok-4.7, Kimi K3 and DeepSeek v4 Pro
- No extra coding-specific training was used; the bet is that human-AI collaboration beats solo frontier models
Available exclusively via their Agent API, with a $250 API credit for a week. gharik reposts inviting people to try it.
More from coding & agent
- Building a work-focused agent harness: chat isn't enough, you need a workspace — GGTech407 · 2026-11-06
- The Church of Approximate Grace: an MCP server that sends agents to weekly services — coleclifford · 2026-11-06
- A five-agent team ran a startup for a day: 1h46m total runtime, onboarding to payments covered — pramodk73 · 2026-11-06
- Users claim Claude performs better with an empty system prompt than Anthropic's official one — repligate · 2026-11-06
- OpenClaw agents drown macOS in Chrome tabs until memory runs out — PhilippaWinchester · 2026-11-06
- Agent money wins roundup: $320K month, $34K/mo infra savings, Pokemon card flips — abe17124 · 2026-11-06