Zhipu AI

GLM-5.3-Flash

Released Aug 26, 2026 · knowledge cutoff unpublished

Status: Unknown
Location: China
Modality: Multimodal
Context window: 1M
Max output: 131K
Speed (max effort): 50 tok/s · 1.49s to first answer token
Price ($/MTok in / out): $0.15 / $0.50
Cost per task (max effort): $0.090
Open weights: Yes

Benchmarks2 of 10 reported

Terminal-Bench 2.1

84.3%

LMArena Elo

1475

MMLU-Proretired

—

GPQA Diamond

—

SWE-bench Verified

—

SWE-bench Pro

—

AIME (latest)retired

—

Humanity's Last Exam

—

GDPval-AA v2

—

ARC-AGI-2

—

Strengths & weaknesses

Strengths

Combines sparse and linear attention in one open checkpoint, cutting attention compute and KV-cache cost several-fold while keeping the million-token window usable rather than degrading toward the edges
Handles native multimodal tool calls — text, image, video and file input in the same turn — without a vision adapter bolted onto a text-only base
Ran anonymously on OpenRouter as "Ox Alpha" for a week under heavy public load before Zhipu claimed it, serving at real production volume before anyone knew whose model it was

Weaknesses

Streams roughly half as fast as its own flagship sibling GLM-5.3 — the Flash name doesn't guarantee low latency, so time-sensitive use still needs its own benchmarking
Default responses aren't automation-ready for strict structured output — needs explicit response-format constraints where some rivals return clean JSON out of the box
Reasoning is always-on with only a low/high/max budget dial, not an off switch — every call carries at least some thinking overhead

320B-total/18B-active MoE, Zhipu's first natively multimodal model in the GLM-5 line. Terminal-Bench 2.1 (84.3%) is Z.ai's own reported figure; GPQA Diamond, HLE, SWE-bench Verified, AIME and ARC-AGI-2 are left null (LMArena Elo was listed later, below) — Z.ai's own table reports non-standard benchmarks instead (DeepSWE v1.1 63.4%, AutomationBench 48.8%, Toolathlon Verified 78.4%), and a single-source claim of an Artificial Analysis GPQA Diamond (91.2%) and HLE (55.3%) score could not be corroborated across independent sources on re-check. costPerTask ($0.09) and speed (49.8 tok/s, 1.49s to first answer token) are Artificial Analysis' max-effort measurements — the model has no medium-effort variant (only low/high/max). Pricing is Z.ai's standard $0.15/$0.50 rate; a 50%-off launch promotion runs through 2026-09-09. predecessorId is left null since Zhipu's own materials frame this as a new capability tier rather than an explicit replacement for GLM-5.2 or GLM-5.3. LMArena Elo 1475 is the "glm-5.3-flash" text-arena listing (arena.ai, 2026-09-13 update, rank 29). On 2026-09-18 Z.ai added GLM-5.3-FlashX (`glm-5.3-flashx`), a hosted high-speed tier of the same weights at up to 200 tok/s for $0.37/$1.25 — no separate benchmarks or checkpoint, so it is recorded here rather than as its own entry; FlashX is not on the GLM Coding Plan.

For developers

API model strings

zhipuglm-5.3-flash

Licence

MIT License · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor