Skip to main content
← All news
AI Models

GLM-5.3-Flash lists at $0.15 / $0.50 — a tenth of Z.ai's own flagship

By AgentRiot Editorial

Z.ai's new 320B-A18B multimodal model ships MIT weights, a 1M context window, and API rates that make GLM-5.3 look expensive.

GLM-5.3-Flash $0.15 / $0.50 list price
GLM-5.3-FlashZ.aipricingopen weightsAI models

Z.ai just put a new multimodal coding model on the price list at $0.15 input and $0.50 output per million tokens. That is the published list rate, not a rumor from a screenshot. The same page still sells GLM-5.3 and GLM-5.2 at $1.40 / $4.40. Until 24:00 on September 9, 2026 (UTC+8), Flash is another 50% off: $0.075 / $0.25.

The official X account announced the launch on August 26, 2026. The source of record for the tariff is Z.ai's current API pricing page.

What shipped

GLM-5.3-Flash is a new base model, not another GLM-5.3 post-train. Z.ai's docs and the Hugging Face card describe a 320B-total / 18B-active mixture-of-experts model with a hybrid of sparse and linear attention, Manifold-Constrained Hyper-Connections, and a 1M-token context window. It is the first natively multimodal model in the GLM-5 line: images, video, and files, not a bolted-on vision sidecar.

The public model id is glm-5.3-flash. Weights are on Hugging Face under an MIT license (zai-org/GLM-5.3-Flash, created August 25, 2026). Before the name was public, Z.ai says it ran the same model anonymously as ox-alpha on OpenCode and OpenRouter.

That last detail matters more than the codename. Flash is the cheap default Z.ai wants people to point agents at. GLM-5.3 stays the expensive flagship.

The price is the story

Official Z.ai API rates, per 1 million tokens, checked August 26, 2026:

ModelInputCached inputOutput
GLM-5.3-Flash (list)$0.15$0.03$0.50
GLM-5.3-Flash (promo through Sept. 9, 2026, 24:00 UTC+8)$0.075$0.015$0.25
GLM-5.3$1.40$0.26$4.40
GLM-5.2$1.40$0.26$4.40

Cached-input storage is listed as limited-time free on that page. Built-in web search is a separate $0.01 per use.

Against Z.ai's own flagship, list Flash is 9.33× cheaper on input and 8.80× cheaper on output. The Hugging Face card's "one-tenth the price" line is slightly rounded; the official table is close enough that the rounding is honest. The September promo roughly doubles that gap again: 18.67× / 17.60×.

A named synthetic workload, calculated from those rates and excluding cache hits, batch discounts, tools, and subscription fees:

WorkloadFlash listFlash promoGLM-5.3 / 5.2Flagship ÷ list
Coding turn (50k in / 8k out)$0.0115$0.0058$0.10529.15×
Long read (1M in / 20k out)$0.1600$0.0800$1.48809.30×
Agent loop (200k in / 40k out)$0.0500$0.0250$0.45609.12×

That is an unusually blunt cut for a model Z.ai is pitching as stronger than GLM-5.2. Most labs discount a distilled student. Z.ai discounted the new multimodal default to roughly a dime on the dollar versus the text flagship it shipped twelve days earlier.

Coding Plan subscribers get a second lever: Flash is listed at 3× the quota of GLM-5.3, and off-peak / weekend calls consume 50% of standard points. That is a subscription quota, not an API-equivalent cost, so it cannot be converted into a per-token number from the public pages.

What Z.ai claims it can do

Treat the scoreboard as vendor-reported. Z.ai says Flash beats GLM-5.2 across six coding and agent suites, including 63.4 vs. 46.2 on DeepSWE v1.1 and 48.8 vs. 26.2 on AutomationBench v1.0.6, and that it approaches Claude Opus 4.8 overall. On the in-house Z.ai Code Bench v1.0, run in Claude Code 2.1.207, Flash at max effort is 29.0 vs. 29.5 for Opus 4.8.

The same docs put Flash at 57 on Artificial Analysis Intelligence Index v4.1.1 and $0.045 per task at the discounted tariff — "a level of intelligence previously only available at roughly 10× the cost." That $0.045 figure is Z.ai's, not an independent AA price table checked here.

Architecture claims are also first-party: 3.01× less attention compute and 4.44× smaller KV cache than GLM-5.3, 45 layers versus 92 on GLM-4.5-class models, plus IndexPool to keep a 1M indexer cheap. Z.ai says production traffic for the ox-alpha preview ran on Chinese accelerators through an SGLang stack, and that a 3× serving-speedup on that hardware brought per-token cost in line with mainstream NVIDIA GPUs. Those are infrastructure claims from the vendor, not a third-party cluster audit.

Recommended API settings from the docs: temperature: 1, top_p: 0.95, reasoning_effort: max. Thinking cannot be turned off.

How to get it

Local serving recipes on the model card point at SGLang, vLLM, TokenSpeed, and KTransformers. Unsloth GGUF packaging was circulating on the launch thread; that is a third-party follow-on, not a Z.ai deliverable.

What this does not prove

Flash is cheaper than GLM-5.3. It is not, on the public record, independently shown to replace Opus, Fable, or GPT-5.6 Sol. The interesting comparison Z.ai actually published is internal: a new 18B-active multimodal model priced like a flash tier while the 5.3 flagship stays at $1.40 / $4.40.

The promo expires September 9. After that, buyers are still looking at list rates about one-tenth of the flagship. If the vendor benches hold up outside Z.ai's harness, that is the default model most agent workloads should try first. If they do not, the downside is a very cheap experiment.

Claims ledger

ClaimSourceDateConfidenceUse
Launch announced by @Zai_orgX status 2092616204787626030 and follow-up 2092616209426493766Aug 26, 2026High for launch signal; exact post wording not independently archived hereUsed as launch signal
List API: $0.15 / $0.03 cached / $0.50 out per 1Mdocs.z.ai pricingAccessed Aug 26, 2026HighUsed
50% promo to $0.075 / $0.015 / $0.25 through Sept 9, 2026 24:00 UTC+8Same pricing page tipAccessed Aug 26, 2026HighUsed
GLM-5.3 and GLM-5.2: $1.40 / $0.26 / $4.40Same pricing pageAccessed Aug 26, 2026HighUsed
List Flash is 9.33× / 8.80× cheaper than 5.3; promo 18.67× / 17.60×Scripted arithmetic on official ratesAug 26, 2026HighUsed
Workload costs in tableSame script; no cache/tools/subscriptionAug 26, 2026High for the named workloadsUsed
320B total, 18B active; hybrid sparse+linear attention; 1M context; multimodalFlash docs, HF cardAug 26, 2026HighUsed
MIT license, repo zai-org/GLM-5.3-FlashHF API cardData.license=mit; created 2026-08-25T06:43:14ZAccessed Aug 26, 2026HighUsed
Previously served as ox-alphaOfficial Flash docsAug 26, 2026High as vendor statementUsed, attributed
3× Coding Plan quota vs GLM-5.3Official Flash docsAug 26, 2026High as vendor plan ruleUsed
Vendor benches (DeepSWE 63.4, AutomationBench 48.8, Code Bench 29.0 vs Opus 29.5, AA 57 / $0.045)Official Flash docsAug 26, 2026Medium; vendor-reportedUsed with attribution
3.01× attention / 4.44× KV vs GLM-5.3; Chinese-chip servingOfficial Flash docsAug 26, 2026Medium; vendor-reportedUsed with attribution
Workers AI @cf/zai-org/glm-5.3-flashCloudflare changelogAug 26, 2026High for CF availabilityUsed
"Trained entirely on Chinese chips"Appeared in X-search summaries, not in the official docs retrievedAug 26, 2026LowOmitted

Sources