GLM-5.3-Flash API Pricing 2026: Z.ai’s 320B Hybrid-Attention Multimodal Model

On August 26, 2026, Z.ai (the international arm of Beijing’s Zhipu AI) surfaced a new model on its API and on OpenRouter: GLM-5.3-Flash. Two numbers make it worth a hard look even in a week crowded with releases: it is a 320B-total / 18B-activated model, and it is the first open-source frontier model to pair sparse attention with linear attention in a hybrid architecture. Z.ai’s official documentation claims this design cuts attention computation by 3.01× and KV-cache size by 4.44× relative to GLM-5.3 — which is exactly the kind of serving-cost math an API pricing site like this one exists to interrogate. This review covers the verified per-million-token pricing on docs.z.ai (captured 2026-08-27), the 50% launch discount window, the multimodal surface, and how Flash fits against GLM-5.3 and the other low-cost reasoning APIs in the current market.

Why the architecture matters for API cost

Most "Flash" or "mini" models cut cost by shrinking the whole model. GLM-5.3-Flash takes a different path: it keeps 320B total parameters but activates only 18B per token, and it replaces the quadratic self-attention with a hybrid of sparse attention and linear attention. Per Z.ai’s docs, versus GLM-5.3 this reduces attention computation by 3.01× and KV-cache size by 4.44×. For a developer the practical consequence is a lower floor on serving cost and long-context memory, which is why Z.ai prices it at a fraction of the flagship. The same architecture is what lets the hosted API stay cheap while keeping "precise long-context capabilities" instead of the context-truncation you often get from a genuinely small model.

GLM-5.3-Flash pricing 2026: verified from docs.z.ai

The official pricing table on docs.z.ai/guides/overview/pricing (captured 2026-08-27) lists GLM-5.3-Flash per 1M tokens:

• Input: $0.15 per 1M tokens — $0.075 during the 50% launch promotion.
• Output: $0.50 per 1M tokens — $0.25 during the promotion.
• Cached input: $0.03 per 1M tokens — $0.015 during the promotion.
• Cached-input storage: limited-time free.

The strike-through prices are the list prices, and the 50% discount runs until 24:00 on September 9, 2026 (UTC+8, Singapore time). On OpenRouter the model is live as z-ai/glm-5.3-flash at $0.075 input / $0.25 output per 1M with a 1.25M-token context window — matching the discounted Z.ai rate, so the promo is what the market actually sees today.

GLM-5.3-Flash vs GLM-5.3: a clean two-tier story

Z.ai is now shipping two distinct GLM-5-class models, and the price gap is the fastest way to read them:

*Strikethrough = the list price; the un-struck figure (e.g. $0.075) is the current discounted price. GLM-5.3 is Z.ai’s reasoning flagship for complex software engineering and long-horizon agent tasks; GLM-5.3-Flash is the cost-efficient multimodal sibling. At one-ninth the input price with 18B active parameters, Flash trades the deepest chain-of-thought for throughput and unit economics — the right trade for high-volume, vision-in-the-loop coding agents where a cheaper token budget matters more than a marginally higher reasoning ceiling.

Native multimodal: visual coding in the loop

GLM-5.3-Flash is the first GLM-5-series model with native multimodal input: images, videos, and files go through the same Chat Completion API as text. You add an image via a content block with type: "image_url", passing either an image URL or a Base64 data URL. Because vision is baked into the coding loop rather than bolted on, the model can observe interfaces, rendered results, and interaction feedback, then test and improve its work — spanning frontend development, game creation, Blender 3D scenes, and GUI-driven operation via browser and computer-use tools. For teams that would otherwise stitch together a vision model plus a coding model, this collapses two vendors into one call.

Context, thinking mode, and recommended settings

Text parameters match GLM-5.3 with a 1M-token context window (OpenRouter reports 1,310,720 tokens). Recommended settings from Z.ai: temperature: 1, top_p: 0.95, reasoning_effort: max. Notably, thinking.type only supports enabled — you cannot turn thinking off. For streaming, Z.ai recommends enabling both stream: true and tool_stream: true. The model also supports function calling, context caching (the discounted cached-input tier), structured output, and full streaming.

Calling GLM-5.3-Flash: curl and Python

The endpoint is OpenAI-compatible, so the official OpenAI Python SDK works by swapping base_url. On Z.ai the model code is glm-5.3-flash; on OpenRouter it is z-ai/glm-5.3-flash.

curl https://api.z.ai/api/paas/v4/chat/completions \
  -H "Authorization: Bearer $ZAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.3-flash",
    "messages": [
      {"role": "user", "content": [
        {"type": "text", "text": "Describe this diagram and list the API endpoints it shows."},
        {"type": "image_url", "image_url": {"url": "https://example.com/architecture.png"}}
      ]}
    ],
    "temperature": 1,
    "top_p": 0.95,
    "reasoning_effort": "max"
  }'

from openai import OpenAI

client = OpenAI(
    api_key="$ZAI_API_KEY",          # or OpenRouter key
    base_url="https://api.z.ai/api/paas/v4",  # or https://openrouter.ai/api/v1
)

resp = client.chat.completions.create(
    model="glm-5.3-flash",           # or z-ai/glm-5.3-flash on OpenRouter
    messages=[{"role": "user", "content": [
        {"type": "text", "text": "Explain the cost of running this 320B/18B model at scale."},
    ]}],
    temperature=1,
    top_p=0.95,
)

print(resp.choices[0].message.content)

GLM Coding Plan: 3× quota, off-peak half-price points

GLM-5.3-Flash is now available on Z.ai’s GLM Coding Plan (Personal and Team subscriptions), and it carries 3× the quota of GLM-5.3 on the same plan. The plan uses a points-based quota system with transparent limits, and model calls made during off-peak hours — including all day on weekends — consume only 50% of the standard points. That off-peak pricing hook makes Flash the default pick for tool-addicted developer workflows (Claude Code, Cursor, Cline, Kilo Code and similar clients that can point at a custom OpenAI-compatible endpoint) where token burn is the dominant cost variable.

Who GLM-5.3-Flash is for

Three use cases stand out:

1. Vision-in-the-loop coding agents: one API call handles text + image + video, and the model iterates on rendered UI — collapsing a vision model and a coding model into a single endpoint.
2. High-volume, cost-sensitive multimodal workloads: at $0.075/M input during the promo (and $0.15 list), with 18B active parameters and linear attention, the per-token cost is a fraction of a frontier flagship — ideal for broad scraping, document and chart analysis, and long-context RAG.
3. Self-hosted long-context inference: as an open model, GLM-5.3-Flash’s weights can run on your own GPUs (or on Z.ai’s Chinese-AI-chip serving), which matters for mainland-China residency and strict data sovereignty.

It is a weaker pick when you need the deepest reasoning ceiling (GLM-5.3 — and OpenAI/Anthropic flagships — still hold that), when you cannot tolerate the mandatory thinking pass (thinking.type cannot be disabled), or when you need very long single outputs beyond 128K.

Bottom line

GLM-5.3-Flash is the most interesting "cost-efficient frontier" release of the week because its cost advantage is architectural, not just a smaller model. The hybrid sparse+linear attention design (3.01× less attention compute, 4.44× smaller KV cache vs GLM-5.3) is what lets Z.ai price a 320B-parameter multimodal model at $0.075/$0.25 per 1M during a promotion that runs to September 9, 2026, and settle at $0.15/$0.50 after. Verified on docs.z.ai (2026-08-27) and cross-checked on OpenRouter (live at the same discounted rates), it is a legitimate lower-cost sibling to GLM-5.3 with native image/video input and a 1M+ context window. If your budget is open-model-shaped and your workload is vision-in-the-loop coding or high-volume multimodal RAG, put GLM-5.3-Flash on the shortlist now — while the 50% promo is still live.

If you route GLM-5.3-Flash (or any of the open/multimodal models above) behind one OpenAI-compatible key with cross-region failover, FreeModel is the simplest handoff: one dashboard, one billing relationship, and routing controls without glue code.