← Back to feed

Z.ai published the full GLM-5.3-Flash benchmark details on August 26, the same day it confirmed the model’s identity as the stealth ox-alpha deployment that topped OpenRouter usage charts earlier this month. The identity confirmation was the headline. The numbers behind it tell a more specific story.

The Price-Performance Position

GLM-5.3-Flash scores 57 on the Artificial Analysis Intelligence Index v4.1.1 — the same index that ranks Claude Opus 5 (max) and Opus 5 (xhigh) at the top of the frontier — at a cost of $0.045 per successful task on the discounted pricing tier. Z.ai describes this as pushing the Pareto frontier: intelligence previously only available at roughly ten times the cost.

The practical implication: any team running inference at volume and targeting the 55-60 range on the AA Index has a new option that undercuts the prior cheapest-equivalent by a factor of ten. That is a meaningful gap, not a rounding error in the cost comparison.

Benchmark Numbers

On six coding and agentic benchmarks, GLM-5.3-Flash improves sharply over GLM-5.2:

Benchmark	GLM-5.3-Flash	GLM-5.2	Change
DeepSWE v1.1	63.4	46.2	+37%
AutomationBench	48.8	26.2	+86%
Z.ai Code Bench (max effort)	29.0	—	vs Opus 4.8: 29.5

The Code Bench comparison is vendor-generated and should be read with the usual caveats around in-house evaluations. Still, 29.0 versus Claude Opus 4.8’s 29.5 on the same harness is close enough to be a relevant data point.

The DeepSWE and AutomationBench jumps are larger than typical generation-over-generation gains. A 37% improvement on agentic coding and an 86% improvement on automation tasks, on a model in the same price tier as the predecessor, is the architectural story Z.ai is betting on.

Architecture: Fewer Active Parameters, Hybrid Attention

GLM-5.3-Flash uses 320 billion total parameters with 18 billion active — a ratio of roughly 1:18 sparse activation. For comparison, GLM-4.5 used 355B total with 32B active. The reduction in active parameters, combined with a smaller layer count (45 vs 92), is what drives the inference cost down.

The model introduces two architectural changes not present in the GLM-4.5 series. First, a hybrid architecture combining sparse and linear attention: linear attention captures local dependencies through state modeling, while sparse attention retrieves global context through a lightweight indexer. A component called IndexPool compresses four indexer key vectors into one at million-token context lengths, reducing memory and latency overhead at scale.

Second, Manifold-Constrained Hyper-Connections (mHC), a technique for improving scaling efficiency. Combined with a 30-trillion-token multimodal pre-training corpus — the first in the GLM-5 series to be natively multimodal — Z.ai is claiming more intelligence per activated parameter than its prior generation.

The Chinese Chips Angle

GLM-5.3-Flash’s inference runs on Chinese domestic AI chips. Z.ai noted in the launch post that ox-alpha’s OpenRouter traffic, which drove it to the most popular model of the week, was served entirely on domestic hardware.

This is significant for a different reason than the benchmark numbers. US export restrictions on Nvidia H100 and H200 hardware have pushed Chinese AI labs to accelerate domestic chip adoption. GLM-5.3-Flash demonstrating competitive frontier-adjacent performance on domestic chips under real production load — not a controlled benchmark run — is a proof point that Chinese inference infrastructure can handle commercial-scale AI workloads without US hardware.

That distinction matters for anyone tracking AI infrastructure independence trajectories, not just raw model performance.

What This Adds to the Landscape

The ox-alpha disclosure period, August 21 through August 26, generated significant coverage of the stealth deployment angle. The model is now named and benchmarked. What the numbers show is that Z.ai’s GLM series has moved from “open-weight option in the top ten” to “price-performance outlier competing with premium inference tier models.”

GLM-5.3-Flash is available on HuggingFace (zai-org/GLM-5.3-flash) and via the Z.ai API. Pricing at $0.045 per task reflects a discounted tier; standard pricing is not listed in the launch materials.