Which open-weight model is cheapest: Kimi K3, GLM-5.2, or DeepSeek V4 Pro?

DeepSeek V4 Pro is by far the cheapest at roughly $0.43 input / $0.87 output per MTok. GLM-5.2 runs around $4.40 output, and Kimi K3 is the premium option at $3 input / $15 output — roughly 17x DeepSeek V4 Pro's output cost.

Is Kimi K3 worth the higher price over DeepSeek V4?

Kimi K3 adds native vision, always-on reasoning, and a 1M-token context as the first open 3T-class model, and it edges DeepSeek on some intelligence indexes. But at ~17x the output cost, DeepSeek V4 Pro wins on pure cost-per-task unless you specifically need vision or K3's peak reasoning.

Which open-weight model should I deploy in 2026?

For cost-sensitive, text-only workloads, deploy DeepSeek V4 Pro. For multimodal or top open-model intelligence, Kimi K3. GLM-5.2 (744B, MIT) is a strong middle option that's cheaper to serve than K3 while staying capable.

Quick Answer

Kimi K3 vs GLM-5.2 vs DeepSeek V4 Pro: Cost (July 2026)

Published: July 30, 2026

The Short Answer

Among open trillion-scale MoE models in July 2026, DeepSeek V4 Pro is the cheapest by a wide margin, GLM-5.2 is the mid-priced middle ground, and Kimi K3 is the premium option — the first open 3T-class model with native vision, but ~17x DeepSeek’s output cost.

Head-to-Head

	Kimi K3	GLM-5.2	DeepSeek V4 Pro
Maker	Moonshot AI	Zhipu	DeepSeek
Params (total)	~2.8T MoE	744B	1.6T (49B active)
API in/out ($/MTok)	$3 / $15	~ / ~$4.40 out	$0.43 / $0.87
Context	1M	large	1M
Multimodal	Yes (native vision)	Text-focused	Text-only
License	Modified MIT	MIT	MIT-style
Blended cost/task	~$0.94	mid	~$0.04

Where Each Wins

Cheapest to serve → DeepSeek V4 Pro. ~$0.43/$0.87 per MTok and roughly $0.04 per task on blended cost — an order of magnitude below Kimi K3’s ~$0.94. Text-only, 1M context, permissive license. The default open pick for cost-sensitive production.
Best value middle → GLM-5.2. 744B, MIT-licensed, ~$4.40 output. Cheaper to serve than K3 while staying capable — a sensible pick if DeepSeek feels too minimal but K3 is overkill.
Top open intelligence + multimodal → Kimi K3. The first open 3T-class model, with native vision, always-on reasoning, and a 1M context. It edges DeepSeek on some intelligence indexes — but at $3/$15 it’s a premium-tier open model, ~17x DeepSeek V4 Pro’s output cost.

The Cost Gap Is Huge

On output tokens alone, Kimi K3 ($15) is roughly 17x DeepSeek V4 Pro ($0.87). Unless you specifically need vision or K3’s peak reasoning, that gap is very hard to justify at scale. GLM-5.2 sits in between and is the pragmatic compromise for teams that want more than DeepSeek without K3’s bill.

How To Decide

Pure cost-per-task, text-only → DeepSeek V4 Pro.
Need vision or top open intelligence → Kimi K3.
Want a capable, cheaper-than-K3 middle → GLM-5.2.
Self-hosting? All three are open-weight; K3’s 3T scale needs the most hardware, DeepSeek V4 Pro the least among these for a given throughput.

Sources

MarkTechPost — Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2: marktechpost.com
BenchLM — DeepSeek V4 Pro vs Kimi K3: benchlm.ai/compare/deepseek-v4-pro-vs-kimi-3
DeepSeek — API pricing: api-docs.deepseek.com

Last verified: July 30, 2026