GLM-5.3: Coding, API Pricing and Local Setup

Choose GLM-5.3 for a text-based coding or agent workflow where you want to evaluate Z.ai’s larger model. The vendor reports stronger long-horizon coding results than GLM-5.2, but a benchmark score does not tell you how many attempts your repository change will need.

For screenshot-driven frontend work or lower token spend, start with GLM-5.3-Flash. GLM-5.3 accepts text only. Supply extracted text or choose the multimodal model when the source material is visual.

Z.ai lists $1.40 per million input tokens, $0.26 for cached input and $4.40 per million output tokens. An uncached request with 20,000 input and 3,000 output tokens costs 0.02 × $1.40 + 0.003 × $4.40 = $0.0412. A thousand such requests cost $41.20 before tools and other charges.

Cached-input discounts apply to cache hits, not all repeated-looking text. Cache storage is separately marked free for a limited time. Z.ai’s built-in web search adds $0.01 per use; a coding agent can also incur execution and external-service costs.

Both offer a 1M-token context and up to 128K output tokens in Z.ai’s model guides. Their input modalities, rates and weight licenses differ. The example below assumes the same token counts, no cache hits and no tool charges; it compares billing arithmetic rather than task quality.

Count the full agent session, including follow-up prompts, reasoning output recorded in usage, tool-result input and retries. Record whether the task passed its tests. Cost per accepted patch is more useful than cost per first response.

To call GLM-5.3, use a Z.ai API key on your server and the pay-as-you-go Chat Completions endpoint below. A GLM Coding Plan uses its own subscription allowance and endpoint; configure the client for the product you purchased. A subscription quota is not an API dollar balance.

Start with reasoning_effort=low and a bounded output limit for a small integration check. These models require reasoning to stay enabled. Increase the budget and effort for a longer task after verifying that your client handles reasoning, tool calls and completion limits correctly.

In an October 16 engineering report, Kong describes using GLM-5.3 and Flash as default code reviewers after comparing blocking-defect recall, precision and review cost on four pull requests in a synthetic repository. This is an operator’s internal evaluation, with a small test set.

Kong separately reports that triaging reviews by complexity lowered median review cost from $3.93 to $2.47. Model selection and routing are separate savings: reproduce your own acceptance checks before applying either result to another repository.

GLM-5.3 supports low, high and max reasoning effort, with max as the default. Disabling thinking causes requests to fail. Explicitly choose effort instead of inheriting a client’s setting from a different model.

If an agent stops during reasoning, inspect its output limit, finish reason and timeout before repeating the same request. Long reasoning can consume the response budget before a useful final answer. Increasing a limit can also increase spend; compare completed tasks at the settings you intend to use.

The official GLM-5.3 download is FP8, with a separate BF16 checkpoint. Choose the matching runtime and format before estimating memory. The calculator provides a capacity and running-cost estimate; it does not measure throughput or guarantee that a maximum-context configuration fits.

SGLang documents eight-way tensor-parallel FP8 serving on supported AMD accelerators. Its guide notes that the roughly 1.51 TB BF16 checkpoint needs more room than a single MI300X node can provide once cache is included. This is a server-scale deployment rather than a desktop installation.

A community issue reports a sustained-load hang with FP8, eight H200 GPUs and EAGLE on SGLang revision 7760d52. Treat it as a configuration-specific reliability report: include long-context and sustained-load checks in qualification, alongside the initial smoke request.

GLM-5.3 uses a custom GLM-5.3 License, separate from Flash’s MIT terms. It permits use, modification and distribution with its notices and conditions.

For a licensee or affiliate operating a Model as a Service business, aggregate revenue above US$10 billion over any consecutive twelve months triggers a Z.ai security-review condition before commercial use. The license defines that business category and excludes certain embedded end-user features and simple request relaying. Read that definition for a hosted offering.