Announcement
Ampersand announces multi-agent coding mode for its Agent API
The mode comes with company-reported scores of 92.2% on SWE-bench Pro v2 HARD-51 and 77.9% on DeepSWE v1.1.
TLDR
Ampersand introduced a multi-agent coding mode available exclusively through its Agent API. It reports a 92.2% score on SWE-bench Pro v2 HARD-51 and 77.9% on DeepSWE v1.1, and claims the first result matches Fable 5.1 at 70% lower cost. The company says it gave the model no extra coding training. Its post does not explain the benchmark methodology or cost calculation.
Combined views
—
3 Sources, first seen ago
