Ampersand matches frontier coding models at 70% lower cost via co-coding training
niloofar_mire · x · 2026-11-06
Humansand released Ampersand, trained to collaborate with humans — and it accidentally became a top multi-agent coding model without extra coding training.
- 92.2% on SWE-bench Pro v2 HARD-51, matching Fable 5.1 (70% cheaper) and Gemini 4 Argon, outperforming Astra, Muse Wonder, and Kimi K3
- 77.9% on DeepSWE v1.1, on par with Gemini 4 Argon and ahead of Astra, Grok-4.7, Kimi K3, and DeepSeek-v4-pro
Multi-agent mode is available exclusively via the Agent API, with $250 in API credits for the next week.
More from coding & agent
- Deep dive: how Google's AI search grounding snippets actually work — dejanseo · 2026-11-06
- rakyll: MCP server disallows background commands, tool calls block until finished — rakyll · 2026-11-06
- Data engineer maps AI concepts to data engineering: agents are just pipelines — Zachly · 2026-11-06
- Linear Launches Mobile Diffs: Review Code and Iterate Coding Agents From Your Phone — julianlehr · 2026-11-06
- Will backup, data platforms and enterprise search converge? Is Claude+MCP the enterprise frontend? — mobileraj · 2026-11-06
- Perplexity Computer builds 6-minute GTA RAGE engine explainer from 40 web sources — jeffgrimes9 · 2026-11-06