Z.ai released GLM-5.3-Flash on August 26, 2026 — the first natively multimodal model in the GLM-5 series. It combines open weights under an MIT license, Z.ai-reported gains of 17 to 22 percentage points on coding and agent benchmarks over GLM-5.2, and API pricing at roughly one-ninth the rate of its predecessor at list prices. Taken together, these properties move GLM-5.3-Flash from a routine model update to a genuine evaluation decision for teams running AI infrastructure on coding, agent, and long-context workloads.

The release confirmed what many developers had already inferred. For the week before the official announcement, GLM-5.3-Flash had been served anonymously as "Ox Alpha" on OpenRouter and OpenCode. By August 25, multiple independent outlets — including The Next Web, the Straits Times, and LLM Rumors — reported it had reached first place among 558 models in OpenRouter's rolling-week popularity. OpenRouter reported approximately 23.2 trillion tokens processed in the prior week; OpenCode reported approximately 134,000 distinct users in its first week. Access during the preview period was free or heavily discounted. Pre-release adoption at zero cost demonstrates real workload exposure — it does not independently validate performance against paid models in controlled conditions.

What Z.ai released

GLM-5.3-Flash is a mixture-of-experts model with 320 billion total parameters and 18 billion activated per inference pass. Z.ai describes it as the first open-source frontier model to adopt a hybrid architecture combining sparse attention and linear attention — a design intended to reduce long-context serving costs while maintaining context capability. The architecture also incorporates Manifold-Constrained Hyper-Connections (mHC). Z.ai reports training on a 30-trillion-token multimodal corpus. The technical report is available as arXiv:2602.15763.

The model supports a maximum context of 1,048,576 tokens, text and image input, function calling, structured output, context caching, and streaming. Weights are publicly available at huggingface.co/zai-org/GLM-5.3-Flash under the MIT license, which provides a permissive basis for commercial use and modification. Local serving is supported via SGLang, vLLM, TokenSpeed, and KTransformers.

What the benchmarks show — and what they measure

Z.ai reports the following comparisons between GLM-5.3-Flash and GLM-5.2. All figures are Z.ai-reported. Each benchmark uses a distinct evaluation setup:

Benchmark	GLM-5.3-Flash	GLM-5.2	Delta
DeepSWE v1.1	63.4	46.2	+17.2 pp
AutomationBench v1.0.6	48.8	26.2	+22.6 pp
Toolathlon Verified	78.4	59.9	+18.5 pp
Terminal Bench 2.1	84.3	81.0	+3.3 pp

Evaluation setup (Z.ai): DeepSWE used the mini-swe-agent harness with a 400,000-token context limit and a six-hour timeout. AutomationBench v1.0.6 includes a null-type handling fix (PR #13). Toolathlon Verified reports pass@1 averaged over three independent runs via the official evaluation service. Terminal Bench 2.1 was evaluated inside Claude Code 2.1.207 with a 65,536-token generation limit. Each benchmark uses distinct harness configuration, context limits, and generation settings. These results measure performance on specific setups and are not interchangeable.

Coding and agent benchmarks show large gains over GLM-5.2. Terminal Bench shows a smaller improvement of 3.3 percentage points. Z.ai separately states the model approaches Claude Opus 4.8 on its internal coding benchmark — that comparison uses Z.ai's own methodology, not a neutral third-party assessment.

The Artificial Analysis Intelligence Index v4.1.1 lists GLM-5.3-Flash at 57. Artificial Analysis independently evaluates at least one index component (GDPval-AA v2). The full index incorporates nine evaluations.

What the economics look like

Z.ai's official pricing page lists the following rates per million tokens. A promotional 50% discount runs through September 9, 2026 (24:00 UTC+8):

	GLM-5.3-Flash (promotional)	GLM-5.3-Flash (list)	GLM-5.2
Input	$0.075	$0.15	$1.40
Cached input	$0.015	$0.030	$0.26
Output	$0.25	$0.50	$4.40

At list prices, GLM-5.3-Flash input costs approximately 9× less than GLM-5.2; output approximately 8.8× less. At promotional rates, those gaps widen to roughly 19× and 18× respectively. Z.ai describes the relationship as "one-tenth the price," which is approximately accurate at list prices across input and output combined. These are comparisons between Z.ai's own models on their own API — not comparisons with other providers.

Architecture efficiency: a different baseline

The architectural efficiency figures Z.ai publishes use GLM-5.3 — the full non-Flash model — as the comparison point, not GLM-5.2. Compared to GLM-5.3, Z.ai reports the hybrid attention design reduces attention computation by 3.01× and KV cache size per layer by 4.44×. These are attention-module measurements. They do not translate directly into 3× lower total inference cost or 4.4× lower total GPU memory requirement.

Z.ai also reports a 3× improvement in end-to-end serving throughput on its Chinese-chip infrastructure, compared to its own initial serving baseline on the same hardware. Z.ai states the optimized stack achieves per-token economics comparable to mainstream NVIDIA GPU deployments. This comparison applies to Z.ai's own serving infrastructure; it is not a direct hardware benchmark against NVIDIA GPUs.

Who should evaluate it

Existing GLM-5.2 users. The combination of Z.ai-reported capability improvements and a substantially lower API price creates a clear evaluation case. Even if benchmark gains only partially transfer to production workloads, the economics alone may justify testing.

Teams evaluating open-weight coding or agent models. MIT-licensed weights, active serving framework support (SGLang, vLLM, KTransformers), and strong Z.ai-reported coding and agent benchmark results create a concrete evaluation target. The pre-release Ox Alpha usage — with approximately 134,000 OpenCode users and 23.2 trillion weekly tokens on OpenRouter — provides evidence of substantial real-world workload exposure, though access was free or discounted throughout the preview period.

Cost-sensitive inference workloads. For high-volume agentic or output-heavy workloads, the output price differential ($0.25 vs $4.40 per million tokens at current promotional rates) is material. Teams should model their actual usage patterns against both the promotional and list prices before assuming the economics hold post-promotion.

Self-hosting teams. Public weights and multiple serving framework integrations make self-hosted deployment feasible. However, 320 billion total parameters is a large model. The 18 billion active parameter figure describes the per-token compute in the MoE routing — not the model's total footprint in memory. Self-hosted inference at scale requires infrastructure capable of hosting a 320B parameter model.

What is not yet established

All capability comparisons in this article are Z.ai-reported as of the release date. Independent evaluation has already begun. Artificial Analysis currently reports an Intelligence Index score of 57 for GLM-5.3-Flash (see Artificial Analysis evaluation). However, independent evidence remains much narrower than Z.ai's release-day benchmark suite: the DeepSWE, AutomationBench, Toolathlon, and Terminal Bench comparisons discussed above are still Z.ai-reported and have not yet been comprehensively reproduced by neutral third parties.

For teams making infrastructure decisions, GLM-5.3-Flash is a strong evaluation candidate — not yet a proven replacement for models already validated in production environments.