Which is the cheapest multimodal model in August 2026?

GLM-5.3-Flash at $0.15/$0.50 per MTok is the cheapest of the three, costing about $0.007 for a 30K-input/5K-output task. DeepSeek V4 Flash Vision Exp costs $0.0099 off-peak and $0.0198 peak. Gemini 3.7 Flash costs about $0.041 at its introductory rate, roughly six times GLM-5.3-Flash.

Which cheap multimodal model supports video input?

GLM-5.3-Flash and Gemini 3.7 Flash both accept video. DeepSeek V4 Flash Vision Exp is image-only as of August 2026. If your pipeline needs video understanding, DeepSeek is not a candidate regardless of its price advantage.

Which of these models has open weights?

Only GLM-5.3-Flash, released under the MIT license on August 26, 2026. DeepSeek V4 Flash Vision Exp is API-only with no published weights, and Gemini 3.7 Flash is a closed Google model. Self-hosting GLM-5.3-Flash needs roughly 192GB of GPU memory at 4-bit precision.

Should I pick Gemini 3.7 Flash despite the higher price?

Pick it when you need Google Cloud integration, enterprise support terms, or a GA endpoint with stability guarantees. DeepSeek's vision model is experimental and GLM-5.3-Flash comes from a smaller vendor. Also note Gemini 3.7 Flash's $0.75/$3.75 is introductory through December 31, 2026 and reverts to $1.50/$7.50 on January 1, 2027.

Quick Answer

DeepSeek V4 Flash Vision vs GLM-5.3-Flash vs Gemini Flash

Published: August 29, 2026

The Short Answer

Three cheap multimodal models landed within two weeks of each other in August 2026. They are not interchangeable.

GLM-5.3-Flash ($0.15/$0.50) — cheapest, takes video, MIT-licensed weights. Released August 26, 2026.
DeepSeek V4 Flash Vision Exp ($0.22/$0.66 off-peak) — image only, experimental, no weights. Released August 21, 2026.
Gemini 3.7 Flash ($0.75/$3.75) — most expensive, GA and enterprise-supported, video capable. Released August 13, 2026.

If you only read one line: GLM-5.3-Flash wins on price and capability, Gemini 3.7 Flash wins on reliability, and DeepSeek wins only if you are already on DeepSeek.

Head-to-Head

	GLM-5.3-Flash	DeepSeek V4 Flash Vision	Gemini 3.7 Flash
Vendor	Z.ai	DeepSeek	Google
Released	Aug 26, 2026	Aug 21, 2026	Aug 13, 2026
Status	Released	Experimental	GA
Price in/out	$0.15 / $0.50	$0.22 / $0.66 off-peak	$0.75 / $3.75
Peak penalty	none	2× ($0.44/$1.32)	none
Cached input	$0.03	$0.007 off-peak	—
Image input	✅	✅	✅
Video input	✅	❌	✅
Context	1M	1M	1M
Max output	—	384K	65,536
Open weights	✅ MIT	❌	❌
Architecture	320B-A18B MoE	V4 Flash line	Closed

Last verified: August 29, 2026.

Real Cost on a Reference Task

Everyone quotes headline rates. Here is a 30,000-input / 5,000-output task, which is a realistic single document-extraction or agent-step call:

Model	Task cost	vs cheapest
GLM-5.3-Flash (launch discount)	$0.0035	0.5×
GLM-5.3-Flash (standard)	$0.0070	1×
DeepSeek V4 Flash Vision (off-peak)	$0.0099	1.4×
DeepSeek V4 Flash Vision (peak)	$0.0198	2.8×
Gemini 3.7 Flash (introductory)	$0.0413	5.9×
Gemini 3.7 Flash (from Jan 1, 2027)	$0.0825	11.8×

At a million calls a month, that spread is $7,000 versus $82,500. This is the rare comparison where price genuinely dominates the decision — a ten-fold gap is not something better ergonomics recover.

Two expiry dates to diary. GLM-5.3-Flash’s 50% launch discount ends September 9, 2026 (24:00 UTC+8) — plan at $0.15/$0.50. Gemini 3.7 Flash’s introductory rate ends December 31, 2026, doubling on January 1, 2027.

The Peak/Off-Peak Problem

DeepSeek is the only one of the three with time-of-day pricing, and it is easy to get wrong.

Since August 16, 2026, DeepSeek V4 charges peak rates from 01:00-04:00 and 06:00-10:00 UTC — seven hours — and off-peak for the other seventeen, at exactly half. Off-peak is $0.22/$0.66; peak is $0.44/$1.32.

For a US or European workload, most traffic is off-peak and $0.22/$0.66 is the honest number. For an Asia-Pacific business-hours workload, a large share lands in peak, and DeepSeek’s effective cost approaches $0.0198 per task — nearly triple GLM-5.3-Flash. Any comparison that quotes a single flat DeepSeek price is wrong in one direction or the other.

Capability Differences That Actually Matter

Video. GLM-5.3-Flash and Gemini 3.7 Flash accept video; DeepSeek’s vision model does not. For media pipelines, meeting analysis or screen-recording agents, this eliminates DeepSeek outright.

Max output. DeepSeek’s 384K maximum output is far beyond Gemini 3.7 Flash’s 65,536. If you generate long structured extractions in a single call — full document transcriptions, large JSON payloads — Gemini forces chunking that DeepSeek does not.

Weights. Only GLM-5.3-Flash ships weights, under MIT. That matters for data-residency and air-gapped deployments, and it is an insurance policy: if the API changes, you still have the model. Practically it needs about 192GB of GPU memory at 4-bit, so this is a server-class option, not a laptop one. Below roughly 14 million tokens per hour of sustained load, the API remains cheaper than renting the GPUs.

Stability. Gemini 3.7 Flash is GA with Google Cloud support terms. GLM-5.3-Flash is a normal release from a smaller vendor. DeepSeek’s is explicitly experimental — the model ID, behaviour and availability carry no stability guarantee.

Which to Pick

Pick GLM-5.3-Flash for almost every cost-sensitive multimodal workload. It is the cheapest, takes video, has the longest feature list, and the MIT weights mean you are never trapped. The trade is vendor scale and a benchmark profile that sits well below frontier — this is a workhorse, not a reasoning model.

Pick DeepSeek V4 Flash Vision if you already run DeepSeek V4 Flash for text and want image support without changing providers, and your traffic sits off-peak. The text-parity claim means one model handles both modalities in the same pipeline. Do not build anything you cannot re-point in an afternoon.

Pick Gemini 3.7 Flash when procurement, support terms or Google Cloud integration outweigh a 6-12× cost difference. That is a legitimate reason in regulated enterprises and a poor one in a startup. Budget at the 2027 rate of $1.50/$7.50, not the promo.

The router answer applies here too. Send bulk volume to GLM-5.3-Flash, escalate the ambiguous 5-10% to a frontier multimodal model, and the blended cost stays near the cheap tier while accuracy tracks the expensive one.