Frontier Log

Benchmark · Active

CyberGym

Created by UC Berkeley, 3 Jun 2025 · Higher is better · Website

Results

ModelScoreSourceSource check
GLM-5.3-Flash26 Aug 2026 · Z.ai86%GLM-5.3 and the spread of advanced cyber capabilities95-task subset; paired with Anthropic's abliterated copy; run by Anthropic, not chartedSource image
CheckedA person checked this number against its source on 25 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced.
GLM-5.3-Flash abliterated (Anthropic)26 Aug 2026 · Z.ai86%GLM-5.3 and the spread of advanced cyber capabilities95-task subset; run by Anthropic, not chartedSource image
CheckedA person checked this number against its source on 25 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced.
GLM-5.314 Aug 2026 · Z.ai85%GLM-5.3 and the spread of advanced cyber capabilities185-task subset; paired with Anthropic's abliterated copy; run by Anthropic, not chartedSource image
CheckedA person checked this number against its source on 25 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced.
GLM-5.3 abliterated (Anthropic)14 Aug 2026 · Z.ai81%GLM-5.3 and the spread of advanced cyber capabilities185-task subset; run by Anthropic, not chartedSource image
CheckedA person checked this number against its source on 25 Oct 2026. This confirms the number matches what the lab published, not that the score was reproduced.

All benchmarks