Augment claims 33% token savings with new semantic code‑retrieval engine

Key Points

• Augment’s context engine matches Claude Code accuracy with 33% fewer tokens, per Terminal‑Bench test.
• Company spent 18 months developing retrieval and embedding models for large code bases.
• Founder attributes efficiency to precise code snippet selection, not model changes.
• Anthropic says Claude Code shows no measurable gain from added semantic tools.
• Difference traced to varying retrieval systems; not all context engines perform alike.
• Implementation details can make or break perceived benefits of RAG setups.