Head to head

Beam vs GLM 5.2

8:5

Beam wins 8 of 13 shared benchmarks

Verdict

This is the matchup Reflection chose for its headline claim, and Beam holds up: it wins 8 of the 13 shared benchmarks, including SWE-Bench Pro v1, MCP Atlas, AutomationBench and IFBench, while GLM 5.2 is ahead on hard reasoning (AIME, HLE, GPQA, CritPt) and narrowly on Terminal-Bench. Reflection says Beam reaches comparable reasoning scores with three to four times less inference compute.

Benchmark by benchmark

Benchmark	Beam	GLM 5.2	Difference
DeepSWE v1.1	44.4	44.0	+0.4
SWE-Bench Pro v1	65.5	62.1	+3.4
Terminal-Bench v2.1	80.1	81.0	-0.9
AIME 2026	97.8	99.2	-1.4
Humanity's Last Exam (no tools)	36.2	40.5	-4.3
CritPt (AA)	16.3	20.9	-4.6
GPQA Diamond	90.5	91.2	-0.7
AutomationBench (public)	37.0	26.2	+10.8
MCP Atlas	78.7	77.8	+0.9
τ³ banking	38.0	37.1	+0.9
AA-LCR (long context)	79.3	78.3	+1.0
LongBench v2	65.5	64.0	+1.5
IFBench	79.7	73.3	+6.4

All scores are from Reflection's announcement (October 8, 2026 revision). Scores for GLM 5.2 are as cited by Reflection from Artificial Analysis and DataCurve.

The two models

	Beam	GLM 5.2
Developer	Reflection AI	Z.ai (Zhipu AI)
Based in	United States	China
Parameters (total / active)	501B / 23B	753B / 40B
Weights	Apache 2.0, promised for later in October 2026	MIT
Released	October 5, 2026	June 16, 2026
How to access	Waitlisted beta API; weights coming	Hugging Face, Z.ai API, OpenRouter