Z.AI GLM-5.3-Flash — the cost tier of the 5.3 generation. Documented by the vendor under its VLM section: text parameters match GLM-5.3 and it additionally accepts image_url content blocks, so it is multimodal on input. Reasoning is always on.
GLM-5.3-Flash is the cost tier of Z.AI's 5.3 generation. It carries the same 1M-token context and always-on reasoning as the GLM-5.3 flagship, but adds image_url input — making it the first model in the GLM-5.x line to accept vision alongside text — while priced at roughly one-ninth of GLM-5.3's token cost. The vendor documents it under the VLM section rather than a standalone page, and its text parameters are stated to match GLM-5.3 exactly.
The practical routing split between Flash and the flagship comes down to three factors: image input, token cost, and latency. GLM-5.3-Flash is the right route when the request includes an image — it is the only text-generation model in the GLM-5.3 generation that accepts image_url content blocks. For text-only tasks, it is the budget choice when the peak coding performance of the flagship is unnecessary. At $0.162/M input and $0.54/M output on TheRouter, GLM-5.3-Flash sits among the more competitively priced 1M-context multimodal models in the catalog.
Like GLM-5.3, the Flash variant requires reasoning to always be on. The thinking.type: "disabled" path is gone; applications that previously disabled thinking must switch to reasoning_effort: "low" before calling either model. The three effort levels — low, high, max — are available on both routes, and TheRouter exposes reasoning_effort in the supported_parameters list for direct pass-through.
reasoning_effort: "low" combined with Flash's lower price gives a cheaper, faster path than the GLM-5.3 flagship at the same effort levelthinking.type: "disabled" — reasoning cannot be turned off in GLM-5.3-Flash any more than in GLM-5.3; always budget for reasoning tokens, and use reasoning_effort: "low" to trim overheadZ.AI documents GLM-5.3-Flash under the VLM section with text parameters matching GLM-5.3 and additional image_url input support. Context is 1M tokens, reasoning is always on (effort: low/high/max). No separate Flash pricing page has been extracted.
TheRouter exposes zai/glm-5.3-flash through the OpenAI Chat Completion surface with 1M context, 131,072 max completion tokens, text+image input, text output, tools, tool_choice, response_format, and reasoning_effort. Pricing is $0.162/M input and $0.54/M output.
| Type | Rate |
|---|---|
| Input | $0.162 / 1M tokens |
| Output | $0.540 / 1M tokens |
| Context window | 1M tokensdocs.z.ai ↗ | verified |
| Maximum output tokens | 131,072 tokens on the active TheRouter routetherouter.ai ↗ | verified |
| Modalities | Text + image input, text output (vision-capable)docs.z.ai ↗ | verified |
| Image input | image_url content blocks accepted; documented in the Z.AI VLM sectiondocs.z.ai ↗ | verified |
| Reasoning | Always on; thinking.type is fixed at enabled. Three effort levels: low (lightweight), high (enhanced), max (deep, default). Disabling is not supported.docs.z.ai ↗ | verified |
| TheRouter pricing | $0.162/M input tokens; $0.54/M output tokenstherouter.ai ↗ | verified |
| Cost vs GLM-5.3 flagship | Approximately 8.6x cheaper on input, 8.1x cheaper on outputtherouter.ai ↗ | verified |
| API capabilities | Function calling, streaming, structured output, reasoning effort control, image_url vision inputdocs.z.ai ↗ | verified |
| Generation family | GLM-5.3 generation; text parameters match GLM-5.3 flagship; documented under VLM sectiondocs.z.ai ↗ | verified |
| Training data cutoff | Not publicly disclosed | unknown |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
DeepSWE v1.1 (GLM-5.3 flagship, generation reference) Vendor-reported score for the GLM-5.3 flagship from the Z.AI model page. No separate Flash score has been published; this row provides generation-context only and should not be treated as a Flash score. | 66.9 | docs.z.ai ↗ | |
Z.ai Code Bench Max (GLM-5.3 flagship, generation reference) Vendor-reported score for the GLM-5.3 flagship from Z.AI docs, representing a 50% improvement over GLM-5.2 on this in-house benchmark. No separate Flash score has been published. | 34.5%% | docs.z.ai ↗ | |
CyberGym vulnerability discovery (GLM-5.3 flagship, generation reference) Vendor-reported score for the GLM-5.3 flagship on vulnerability discovery, from Z.AI docs. Provided for generation-context; no separate Flash score has been published. | 84.5%% | docs.z.ai ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "zai/glm-5.3-flash",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'GLM-5.3-Flash accepts image_url content blocks alongside text. Pass the image as a URL or base64 data URI in the user message content array.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "zai/glm-5.3-flash",
"reasoning_effort": "low",
"messages": [
{
"role": "user",
"content": [
{"type": "image_url", "image_url": {"url": "https://example.com/diagram.png"}},
{"type": "text", "text": "Describe the architecture shown in this diagram."}
]
}
],
"max_tokens": 800
}'TheRouter's routing guide for the GLM-5.3 generation covers benchmark context, migration from GLM-5.2, API integration with Z.AI and DashScope, and how to position GLM-5.3 and Flash in a multi-model policy.
GLM-5.3-Flash adds image_url vision input — making it multimodal on input — and costs roughly 8–9x less per token than the GLM-5.3 flagship. The text parameters and 1M context are stated to match. The flagship is text-only and carries higher coding benchmark scores; the Flash is the budget and vision-capable route within the same generation.
No. Like GLM-5.3, the Flash variant requires reasoning to always be enabled. The thinking.type: "disabled" parameter is removed. If you previously disabled thinking, switch to reasoning_effort: "low" before updating your model id — this is the lowest-overhead reasoning mode available.
The route is active and listed in TheRouter's catalog, so it is technically available for production. However, no independent benchmark data for GLM-5.3-Flash's vision capability has been published. Validate image understanding quality on your representative inputs before committing to a high-volume pipeline.
Both are large-context, low-cost multimodal flash-tier models with image_url input. At $0.162/M input and $0.54/M output, GLM-5.3-Flash is competitively priced. The key differentiator is the always-on reasoning with selectable effort levels, which Gemini Flash does not expose in the same way. Choose based on ecosystem fit, independent vision benchmarks for your task type, and whether reasoning overhead is an acceptable latency tradeoff.
| source | URL | retrieved | |
|---|---|---|---|
| Context window | docs.z.ai ↗ | 2026-09-18 | verified |
| Maximum output tokens | therouter.ai ↗ | 2026-09-18 | verified |
| Modalities | docs.z.ai ↗ | 2026-09-18 | verified |
| Image input | docs.z.ai ↗ | 2026-09-18 | verified |
| Reasoning | docs.z.ai ↗ | 2026-09-18 | verified |
| TheRouter pricing | therouter.ai ↗ | 2026-09-18 | verified |
| Cost vs GLM-5.3 flagship | therouter.ai ↗ | 2026-09-18 | verified |
| API capabilities | docs.z.ai ↗ | 2026-09-18 | verified |
| Generation family | docs.z.ai ↗ | 2026-09-18 | verified |
| Training data cutoff | — | — | unknown |
| DeepSWE v1.1 (GLM-5.3 flagship, generation reference) | docs.z.ai ↗ | 2026-09-18 | verified |
| Z.ai Code Bench Max (GLM-5.3 flagship, generation reference) | docs.z.ai ↗ | 2026-09-18 | verified |
| CyberGym vulnerability discovery (GLM-5.3 flagship, generation reference) | docs.z.ai ↗ | 2026-09-18 | verified |
| GLM-5.3: Zhipu's Open-Source Flagship with 1M Context — API Integration and Routing Guide | TheRouter Blog ↗ | 2026-09-18 | verified |
| What is different between GLM-5.3-Flash and GLM-5.3? | docs.z.ai ↗ | 2026-09-18 | to verify |
| Can I disable reasoning in GLM-5.3-Flash? | docs.z.ai ↗ | 2026-09-18 | to verify |
| Is GLM-5.3-Flash suitable for production vision pipelines? | therouter.ai ↗ | 2026-09-18 | to verify |
| How does GLM-5.3-Flash compare to Google Gemini Flash for multimodal tasks? | therouter.ai ↗ | 2026-09-18 | to verify |