Ampersand launches co-coding model scoring 92.2% on SWE-bench Pro HARD-51 at 70% lower cost
niloofar_mire · x · 2026-11-12
Humansand introduced Ampersand and Ampersand Tutor, pitching "Co-Coding" — a model trained to work alongside humans. Its multi-agent mode matches frontier coding models without extra coding training: 92.2% on SWE-bench Pro v2 HARD-51 (matching Fable 5.1 while 70% cheaper, beating Astra, Muse Wonder and Kimi K3) and 77.9% on DeepSWE v1.1, on par with Gemini 4 Argon. The multi-agent mode is available exclusively via their Agent API, with a $250 credit for the next week.
More from coding & agent
- Alchemy author demos Z3-backed Prove/Explore for TypeScript domain invariants — samgoodwin89 · 2026-11-12
- SteerCamp user racks up 1,182 agent memory items in 2 weeks with natural recall — pramodk73 · 2026-11-12
- 3 GitHub repos worth reading for the code: nanoGPT, tinygrad, llama.cpp — cneuralnetwork · 2026-11-12
- Reddit comparison of 7 agent messaging approaches: A2A, email, Slack, custom protocols — StrongCoffee85 · 2026-11-12
- Dev feeds new Claude a Call of Duty video, ships a playable browser tactical shooter — TAbrodi · 2026-11-12
- Second V8 patch lands: TypedArray indexOf/includes up to ~30x faster on Uint8Array — cnakazawa · 2026-11-12