GLM-5.3-FlashX: What It Is, How Fast It Runs, and What It Costs

What is GLM-5.3-FlashX?

GLM-5.3-FlashX is the speed-optimized endpoint of the GLM-5.3-Flash family. Zhipu announced it on September 18, 2026, describing it as faster and smoother for enterprises and developers, with the API and the official experience center both open at launch.

Two things are worth separating:

• The model — GLM-5.3-Flash, a 320B-A18B native multimodal model that Zhipu open-sourced on August 26, 2026. Architecture, weights, context length, and capabilities are shared.
• The serving tier — FlashX, an inference configuration tuned for throughput. Zhipu reports speeds up to 200 tokens/s and attributes the gain to infrastructure work rather than a different checkpoint.

That distinction matters when you evaluate it. Benchmarks, context limits, and multimodal behavior described for GLM-5.3-Flash apply to FlashX because they are the same model. Latency and price do not: those change with the serving tier.

Model specifications at a glance

Specification	GLM-5.3-FlashX
API model ID	glm-5.3-flashx
Sibling model ID	glm-5.3-flash
Released	September 18, 2026
Total / activated parameters	320B / 18B
Layers	45
Context window	1M tokens
Maximum output tokens	128K
Input modalities	Video, image, text, file
Output modality	Text
Published peak speed	200 tokens/s
Thinking mode	thinking.type: enabled only; it cannot be disabled
Core API features	Streaming, function calling, context caching, structured output, tool streaming

Zhipu recommends temperature: 1, top_p: 0.95, and reasoning_effort: max. For multi-turn work it recommends keeping the thinking history rather than clearing it (clear_thinking: false), and for streaming requests it recommends enabling both stream: true and tool_stream: true.

GLM-5.3-FlashX vs GLM-5.3-Flash

Dimension	GLM-5.3-Flash	GLM-5.3-FlashX
Underlying model	GLM-5.3-Flash	GLM-5.3-Flash
Published peak speed	Standard tier	Up to 200 tokens/s
List input price / 1M	$0.15	$0.37
List output price / 1M	$0.50	$1.25
Context and output limits	1M / 128K	1M / 128K
Multimodal input	Yes	Yes
Open weights	Yes, since August 26, 2026	Same base model
GLM Coding Plan	Included, with 3× the GLM-5.3 quota	Not included at launch

The request format is identical. Moving from one tier to the other is a change to the model field, not a rewrite of your integration — which makes FlashX a reasonable A/B candidate for workloads where time per turn is the bottleneck.

What "200 tokens/s" does and does not tell you

Zhipu publishes 200 tokens/s as a maximum. Read it as a ceiling on generation speed, not a promise about your workload:

• It is a peak figure. Real throughput depends on prompt length, cache hits, concurrency, and the selected provider.
• It is not time to first token. A fast decode rate still feels slow if the model thinks for several seconds before emitting anything. For interactive products, measure both.
• It is not total task latency. An agent turn that calls three tools is bounded by tool execution, not by token rate.
• Long context changes the picture. At 1M tokens, prefill and KV-cache behavior dominate. That is exactly what the Flash architecture targets.

The practical rule: benchmark Flash and FlashX on your own prompts, and compare time to first token, output tokens per second, and cost per completed task rather than a single speed number.

Where the speed comes from

Zhipu attributes the FlashX speedup to infrastructure investment rather than a new architecture, built on top of the 100,000 domestic AI accelerators already serving GLM-5.3-Flash traffic.

• A dedicated inference engine on SGLang, written for this architecture instead of adapted from a general-purpose stack.
• Memory optimization for 1M-token contexts, including ReplaySSM, W8A8 quantization, hybrid INT8/FP8/BF16 cache quantization, and Layer Split.
• An Encode–Prefill–Decode (EPD) disaggregated serving architecture that separates multimodal encoding, prompt prefill, and token-by-token decoding into independently scheduled worker pools, so each stage scales on its own.
• A 3× end-to-end serving improvement versus the initial baseline on the same hardware, which Zhipu says brings per-token cost to a level comparable with mainstream NVIDIA GPUs.

One detail from that work is worth noting on its own: Zhipu says a GLM-5.3-powered infrastructure agent helped engineers optimize kernels, diagnose bottlenecks, and improve the serving stack — the model participating in the system that serves it.

Benchmarks: what Zhipu reports for the base model

All of the following numbers are published by Zhipu for GLM-5.3-Flash and apply to FlashX because the model is the same. They are vendor-reported results, not independent reproductions.

Benchmark	GLM-5.3-Flash	GLM-5.2
DeepSWE v1.1	63.4	46.2
AutomationBench	48.8	26.2
Z.ai Code Bench v1.0, max effort	29.0	Claude Opus 4.8: 29.5

Zhipu also reports that GLM-5.3-Flash scores 57 on the Artificial Analysis Intelligence Index v4.1.1 at $0.045 per task under its discounted pricing — a level it says was previously available only at roughly 10× the cost — and that its coding performance is comparable to Claude Opus 4.8 on the company's internal Z.ai Code Bench.

Treat these as direction, not warranty. Vendor benchmarks are usually run on vendor-favorable harness versions; the Z.ai Code Bench line above was measured on Claude Code 2.1.207. Re-run your own evaluation before moving production traffic.

Architecture: why the Flash tier is cheap to run

FlashX inherits the design that made Flash inexpensive to serve, which is the reason a faster tier can exist without a price jump to flagship levels.

• Hybrid sparse and linear attention. Zhipu describes it as the first open-source frontier model to combine the two. Linear attention captures local dependencies through state modeling; sparse attention retrieves relevant global context through a lightweight indexer.
• IndexPool. Four indexer key vectors are compressed into one by weighted pooling, cutting the indexer's latency and memory overhead at a 1M-token context.
• Manifold-Constrained Hyper-Connections (mHC) for better scaling efficiency.
• Lower per-token cost than GLM-5.3. Zhipu reports 3.01× less attention compute and a 4.44× smaller KV cache than GLM-5.3. Against GLM-5.3, the activated parameter count drops from 32B to 18B and layers from 92 to 45.
• A 30-trillion-token multimodal pre-training corpus.

Zhipu is candid that the KV cache is still slightly larger than Kimi-K3's and DeepSeek-V4-Flash's and calls that a direction for future work — a useful reminder that architecture comparisons are per dimension, not a single ranking.

Ox-Alpha: the anonymous model that tested in public

Before the Flash launch, Zhipu ran GLM-5.3-Flash anonymously as Ox-Alpha on OpenCode and OpenRouter. According to Zhipu, it became the most popular model of the week and set usage records on both platforms, with all of that traffic served on Chinese AI chips.

For developers, the interesting part is not the leaderboard position but the validation method: real traffic, real coding agents, an unknown model name, and no brand pull. It is a stronger signal than a benchmark table, though still a vendor-controlled rollout with no published methodology.

Native multimodal and visual coding

GLM-5.3-Flash is the first GLM-5-series model built as multimodal from the start, and FlashX keeps that. Images are passed as a content block with type: image_url inside messages[].content[], using either a URL or a Base64 data URL, and multiple blocks can be combined in one message. Video and file input are documented alongside image and text.

The capability that matters most in practice is visual coding: the model inspects a rendered interface, compares it with the target, and iterates. Zhipu documents workflows for rebuilding an application from screenshots or recordings, building Blender scenes, producing Godot game prototypes, running browser and computer-use agents, and reproducing CAD parts — each with the model checking its own rendered output rather than only generating code.

The same loop extends to document work. Zhipu demonstrates PPTX, PDF, DOCX, and XLSX deliverables, financial research with traceable sourcing, contract review with tracked annotations, and legal drafting, with the model rendering and visually inspecting its own output for overflow, misalignment, and inconsistent styling.

One example Zhipu gives is unusually concrete: with no external assets, GLM-5.3-Flash ran autonomously for 16 hours to build a roughly 400 m² chef's residence and test kitchen as a Blender scene.

Pricing: what FlashX costs

Official list prices per 1M tokens, from Zhipu's pricing pages (checked September 18, 2026):

Token type	GLM-5.3-FlashX	GLM-5.3-Flash	GLM-5.3
Input (USD)	$0.37	$0.15	$1.40
Cached input (USD)	$0.075	$0.03	$0.26
Output (USD)	$1.25	$0.50	$4.40

Cached input storage is listed as limited-time free across the three tiers.

Cost example. For 10M uncached input tokens and 1M output tokens, list prices give:

Workload	GLM-5.3-FlashX	GLM-5.3-Flash	GLM-5.3
10M input + 1M output	$4.95	$2.00	$18.40
Same volume, all input from cache	$2.00	$0.80	$6.60

FlashX is roughly 2.5× Flash on input and output, and still well below GLM-5.3. Whether that is worth it depends entirely on what a slow turn costs you — for a batch pipeline it rarely is, and for an interactive agent it often is.

How to call GLM-5.3-FlashX

FlashX uses the same chat-completions interface as Flash, so migrating is a one-line change.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_APIMASTER_KEY",
    base_url="https://apimaster.ai/v1",
)

response = client.chat.completions.create(
    model="glm-5.3-flashx",
    messages=[
        {
            "role": "user",
            "content": "Refactor this module for testability and show the diff.",
        }
    ],
    temperature=1,
    top_p=0.95,
    extra_body={
        "reasoning_effort": "max",
        "thinking": {"type": "enabled", "clear_thinking": False},
    },
)

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

For streaming, add stream: true and tool_stream: true as Zhipu recommends. Note that thinking cannot be switched off on this model, so budget for reasoning tokens in your cost estimate and in your client timeouts.

A practical migration path: keep the model ID configurable, send a representative slice of production traffic to both glm-5.3-flash and glm-5.3-flashx, and compare completion rate, time to first token, and cost per finished task before switching a workload over.

Choosing between FlashX, Flash, and GLM-5.3

• Choose FlashX for interactive products, IDE-side completions, and agent loops with many short turns, where wall-clock time per turn is the constraint and the workload still fits Flash-class capability.
• Choose Flash for batch processing, offline generation, and high-volume pipelines where cost per task matters more than latency — and for open-weight or self-hosted deployments, since Flash is the tier with published weights.
• Choose GLM-5.3 when you need the flagship's ceiling rather than the Flash tier's cost profile.
• Do not assume the speed gain applies evenly. Prefill-heavy, long-context requests and tool-bound agent turns may see much less benefit than short generation tasks. Measure the workload you actually run.

Get started with the GLM family on APIMaster.ai

APIMaster gives you one OpenAI-compatible key for the GLM family alongside Claude, GPT, DeepSeek, Qwen, Gemini, Kimi, and other models — with pay-as-you-go pricing from $1 and up to 70% off official rates on selected routes.

Because FlashX keeps the same request format as Flash, teams already calling GLM through APIMaster can evaluate the faster tier without touching their integration beyond the model ID.

1. Create an APIMaster account.
2. Add pay-as-you-go credit, starting from $1.
3. Create an API key in the APIMaster console.
4. Point your OpenAI-compatible client at https://apimaster.ai/v1 and set the model ID you want to evaluate.

Register for APIMaster · Explore the model marketplace · Test model identity

FAQ

Is GLM-5.3-FlashX a new model? No. It is the high-speed serving tier of GLM-5.3-Flash. Same model, same context window, same multimodal input — a faster inference configuration and a different price.

What is the GLM-5.3-FlashX model ID? glm-5.3-flashx. The standard tier is glm-5.3-flash.

How fast is GLM-5.3-FlashX? Zhipu publishes a maximum of 200 tokens/s. That is a peak figure; measure time to first token and sustained throughput on your own prompts.

How much does GLM-5.3-FlashX cost? Zhipu lists $0.37 per 1M input tokens, $1.25 per 1M output tokens, and $0.075 per 1M cached input tokens — about 2.5× the price of GLM-5.3-Flash on input and output.

What is the context window? 1M tokens, with up to 128,000 output tokens, the same as GLM-5.3-Flash.

Can GLM-5.3-FlashX accept images and video? Yes. It accepts video, image, text, and file input and returns text. Images are sent as an image_url content block, by URL or Base64 data URL.

Is GLM-5.3-FlashX in the GLM Coding Plan? Not at launch. GLM-5.3-Flash is fully available on the plan with 3× the GLM-5.3 quota, and off-peak calls including all weekend consume 50% of the standard points.

Can I turn thinking off to save tokens? No. thinking.type only supports enabled. Zhipu recommends keeping the thinking history (clear_thinking: false) for multi-turn work.

Are the benchmark numbers independent? No. They are published by Zhipu. Treat them as directional and re-run your own evaluation before committing production traffic.

Sources and further reading

• Z.ai — GLM-5.3-Flash/FlashX model documentation — model IDs, parameters, capabilities, recommended settings, benchmarks and serving details
• Z.ai — Pricing — official per-1M-token list prices
• Z.ai — GLM-5.3-Flash technical blog — architecture, efficiency comparisons and evaluation tables
• Hugging Face — zai-org/GLM-5.3-Flash — open weights for the base model
• PANews — Zhipu launches GLM-5.3-FlashX — launch report and Ox-Alpha background

All sources checked on September 18, 2026. Prices and availability change; verify current terms before committing a large workload.