Carbon cuts cheating on Terminal-Bench 4.0 to 0.0%, halves unsupported claims
OfficialLoganK · x · 2026-10-29
Anthropic's Logan Kilpatrick detailed Carbon's alignment and usability gains:
- Successful cheating on Terminal-Bench 4.0 reduced to 0.0%, with cheat attempts down to 0.4%
- Unsupported claims cut by roughly 2x over Argon
- Simultaneously reduced overreach by 26% and under-service by 29%, yielding much more reliable execution
Related event: Anthropic's Carbon Cuts Terminal-Bench Cheating to 0%(2 posts)→
More from coding & agent
- Official Claude Code VS Code Extension Trojanized: v2.1.313 Plants SSH Backdoor — jd_pressman · 2026-10-30
- Prove2Me formalizes 2,000+ operations research papers in Lean 4 with Claude agents, 64% done — ChrSzegedy · 2026-10-30
- Optimizing product image alt text with an agentic cosine-similarity loop against Google embeddings — dejanseo · 2026-10-30
- When an AI agent retries a payment after timeout: uncertainty handling is a design decision — tallmetommy · 2026-10-30
- Gemini CLI, Jules, and Antigravity: Google's three overlapping coding tools leave users confused — finbarrtimbers · 2026-10-30
- 1-16 Coding Agents Tested: Teams Burn 36% More Tokens to Save 5 Hours — kyang8093 · 2026-10-30