How much does GLM-5.3-Flash cost?

Standard pricing is $0.15 per million input tokens, $0.03 for cached input, and $0.50 per million output tokens. Z.ai is running a 50% launch discount through September 9, 2026 at 24:00 UTC+8, which halves those rates to roughly $0.075 in and $0.25 out. A reference 30K-input/5K-output task costs about $0.007 at standard rates and $0.0035 discounted.

How is GLM-5.3-Flash different from GLM-5.3?

GLM-5.3, released August 14, 2026, is the ~743B post-trained coding flagship at $1.40/$4.40 per MTok. GLM-5.3-Flash, released twelve days later on August 26, is a separate 320B-A18B multimodal MoE priced roughly ten times cheaper at $0.15/$0.50, with image and video input that GLM-5.3 does not lead on. Flash is the volume workhorse; GLM-5.3 is the hard-task model.

Can I self-host GLM-5.3-Flash?

Yes. The weights are MIT-licensed, which is one of the most permissive licenses in the current open-weight field. Practically it needs around 192GB of GPU memory at 4-bit precision, so server-class hardware such as NVIDIA Hopper or newer, or an AMD MI300X. Below roughly 14 million tokens per hour of sustained load, the API is cheaper than renting the GPUs.

Quick Answer

What Is GLM-5.3-Flash? Z.ai's 320B MoE at $0.15/MTok

Q: What is GLM-5.3-Flash?

GLM-5.3-Flash is Z.ai's cheap multimodal model, released August 26, 2026. It is a natively multimodal mixture-of-experts model with 320 billion total parameters and 18 billion active per token, a 1 million-token context window, and image and video input. Standard API pricing is $0.15 per million input tokens and $0.50 per million output tokens, with weights released under the MIT license.

Published: August 27, 2026

The Short Answer

GLM-5.3-Flash is Z.ai’s cheap multimodal model, released August 26, 2026. It is a natively multimodal mixture-of-experts model with 320 billion total parameters and 18 billion active per token, carries a 1 million-token context window, accepts image and video input, and costs $0.15 per million input tokens and $0.50 per million output tokens. The weights ship under the MIT license.

The one-line framing: Z.ai just put a multimodal, million-token, self-hostable model into the same price bracket as the cheapest text-only APIs on the market.

Key Facts

	GLM-5.3-Flash
Released	August 26, 2026
Vendor	Z.ai (Zhipu AI)
Architecture	Natively multimodal MoE, 320B total / 18B active
Context	1,000,000 tokens
Input types	Text, image, video
API price	$0.15 in / $0.50 out per MTok
Cached input	$0.03 per MTok
Launch discount	50% off through Sept 9, 2026 (24:00 UTC+8)
Weights	MIT license
Coding Plan	Lite $18/mo · Pro $80/mo · Max $168/mo

Last verified: August 27, 2026.

What “320B-A18B” Actually Means

The model holds 320 billion parameters but routes each token through only 18 billion of them. That ratio — roughly 5.6% of the network active per token — is why the price can sit at $0.15/MTok while the model still behaves like something much larger.

You pay compute proportional to active parameters. You pay memory proportional to total parameters. That split is the entire economic argument for sparse MoE in 2026, and it is why “how big is it” stopped being a useful question about a year ago.

For self-hosting, the memory number is the one that bites: around 192GB of GPU memory at 4-bit precision. The 18B active figure does not help you there.

Pricing in Context

A reference task of 30,000 input tokens and 5,000 output tokens, priced against the cheap tier as of August 27, 2026:

Model	Input / Output per MTok	Task cost
GLM-5.3-Flash (discounted)	$0.075 / $0.25	$0.0035
GLM-5.3-Flash (standard)	$0.15 / $0.50	$0.007
DeepSeek V4 Flash (off-peak)	$0.22 / $0.66	$0.0099
GPT-5.6 Luna	$0.20 / $1.20	$0.012
DeepSeek V4 Flash (peak)	$0.44 / $1.32	$0.0198
Gemini 3.7 Flash	$0.75 / $3.75	$0.041
GLM-5.3	$1.40 / $4.40	$0.064

At standard rates GLM-5.3-Flash is the cheapest model in that set, and it is the only one there that takes video input.

Two caveats before you build a budget on it. First, the 50% discount expires September 9, 2026 — plan at $0.15/$0.50, not at the promo rate. Second, Z.ai’s own framing is that Flash beats GLM-5.2 “at approximately one-tenth the price,” which is a comparison against its predecessor, not against the current frontier.

How It Compares to GLM-5.3

These are easy to confuse and they are not the same product.

	GLM-5.3	GLM-5.3-Flash
Released	Aug 14, 2026	Aug 26, 2026
Params	~743B base (post-trained)	320B total / 18B active
Price	$1.40 / $4.40	$0.15 / $0.50
Multimodal	Text-focused coding	Native image + video
Weights at launch	Staged, not published	MIT, published
Best for	Hard long-horizon coding	High-volume, multimodal, cheap

Note the reversal on weights. GLM-5.3 shipped without open weights on day one, which was a break from Z.ai’s open-weight-first reputation. GLM-5.3-Flash shipped with MIT weights. If open weights are your requirement, Flash is the release that actually delivers.

Where It Lands on Benchmarks

Z.ai reports that GLM-5.3-Flash approaches Claude Opus 4.8 on internal coding benchmarks — treat vendor-internal comparisons with the usual scepticism. The independent number worth quoting is from Artificial Analysis: an Intelligence Index v4.1.1 score of 57 at roughly $0.045 per task on the discounted tier.

That is not frontier intelligence. Claude Opus 5 and GPT-5.6 Sol sit well above it. What it is, is a cost-per-point result that is hard to beat — which is exactly the tier this model targets.

Self-Host or API?

The MIT license makes self-hosting legally trivial. The economics are the constraint.

At roughly 192GB of GPU memory for 4-bit, you are looking at server-class hardware. On an MI300X at around $2.90/hour, that is roughly $2,088/month of continuous rental. Against $0.15/$0.50 API pricing, the crossover sits near 14 million tokens per hour of sustained load.

Below that, use the API. Above it — or if data residency, air-gapping, or regulatory constraints make the API a non-starter — self-host. Most teams asking this question are two orders of magnitude below the crossover and should just use the API.

Who Should Use It

Good fit: high-volume classification, extraction and summarisation; document and video pipelines where the multimodal input removes a separate vision model; long-context retrieval over the 1M window; teams that need MIT-licensed weights for compliance.

Poor fit: hard multi-file agentic coding and long-horizon planning, where GLM-5.3, Claude Opus 5 or GPT-5.6 Sol remain meaningfully stronger. Flash is a workhorse, not a frontier model.

The router pattern applies here. Send routine volume to GLM-5.3-Flash at $0.007 a task, escalate the genuinely hard 5-10% to a frontier model, and the blended cost lands far below running everything on the expensive tier.