Head-to-Head: Clean Evals vs. Legacy Benchmarks

3 matched domains · positive boost means higher legacy scores
Capability Domain Clean Evals (Tiers S & A) Legacy Benchmarks (Tiers C & D) Public Lift (Boost)
Agentic & Enterprise27.8p (6)41.0p (2)+13.2p
Coding & SWE47.3p (4)47.6p (2)+0.4p
Math & Scientific Reasoning17.6p (3)25.0p (1)+7.4p
Non-comparable pre-release tests omitted from audited scoring (model released 2026-08-26): Terminal-Bench-Science 0.1 (2026-08-27).
Clean Evals Avg (Tiers S & A)
30.4p
15 private, leak-resistant tests
Legacy Benchmarks Avg (Tiers C & D)
37.9p
5 public or aging tests
Matched Public Lift (Boost)
+7.0p
Near zero or negative indicates clean generalization
Machine-Readable Export

Audited Benchmark Performance of GLM-5.3-Flash

Grouped by benchmark integrity so private capability signal stands apart from aging public tests.

Top-Tier Validated Tests (Tiers S & A) (15)

Private, leak-resistant tests that build real capability signal.

BenchmarkBenchmark Integrity StatusDomainRankPercentile
ITBench-AAPrivate / FreshCoding & SWE#3/1283.1p
AutomationBench-AAPrivate / FreshAgentic & Enterprise#28/5044.9p
AA-Briefcase v1.1Private / FreshAgentic & Enterprise#32/5036.7p
Harvey LAB-AA v1.1Private / FreshAgentic & Enterprise#18/2131.6p
MKA (Multimodal Knowledge Agent)Private / FreshAgentic & Enterprise#19/2331.2p
FrontierCode-AA 2.0Private / FreshCoding & SWE#35/4625.6p
EnterpriseOps-Gym-AAPrivate / FreshAgentic & Enterprise#11/1316.7p
FrontierMathPrivate / FreshMath & Scientific Reasoning#26/2713.2p
MLCR-AAPrivate / FreshMath & Scientific Reasoning#43/468.7p
CyberBench v1.1Private / FreshCoding & SWE#16/2541.1p
Code Migration v1.1Private / FreshCoding & SWE#15/2439.4p
CritPt v2Private / FreshMath & Scientific Reasoning#28/4030.8p
LiveBenchPrivate / FreshGeneral Knowledge & QA#23/2830.6p
AA-Omniscience v1.1Private / FreshGeneral Knowledge & QA#35/4727.3p
AA-AnalystAgentPrivate / FreshAgentic & Enterprise#24/245.4p

Mid-Tier Evaluations (Tier B) (2)

Partially exposed or transitional evaluations with moderate leakage risk.

BenchmarkBenchmark Integrity StatusDomainRankPercentile
HLE-DiamondSemi-ExposedGeneral Knowledge & QA#37/4925.0p
VIALS-AASemi-ExposedMath & Scientific Reasoning#18/2322.7p

Legacy & Vulnerable Benchmarks (Tiers C & D) (5)

Public or aging benchmarks prone to overfitting and training-data leaks.

BenchmarkBenchmark Integrity StatusDomainRankPercentile
GDPval-AA v2.1Public / AgingAgentic & Enterprise#18/4359.5p
SWE Atlas - Codebase QnAPublic / AgingCoding & SWE#10/1550.4p
Terminal-Bench 4.0Public / AgingCoding & SWE#27/5044.9p
GPQA DiamondPublic / AgingMath & Scientific Reasoning#13/1725.0p
GDP.pdfPublic / AgingAgentic & Enterprise#39/5022.4p