Kimi K3 vs GLM-5.2 vs DeepSeek V4: Price and Agent Routing

Kimi K3, GLM-5.2, and DeepSeek V4 all offer roughly one-million-token context, but they occupy different cost and capability positions. K3 is the premium multimodal, very-long-output route. GLM-5.2 is a lower-cost long-horizon engineering model. DeepSeek V4 sets the price floor for high-volume text agents.

Price Comparison

International API prices per million tokens:

Model	Input	Cached input	Output	Context	Maximum output
Kimi K3	$3.00	$0.30	$15.00	1M	Up to 1,048,576
GLM-5.2	$1.40	$0.26	$4.40	1M	128K
DeepSeek V4 Pro	$0.435	$0.003625	$0.87	1M	384K
DeepSeek V4 Flash	$0.14	$0.0028	$0.28	1M	128K

K3 costs more because it targets a broader frontier-agent role, not the budget routing tier. Whether the premium pays off depends on task completion, not parameter count.

Example Coding-Agent Bill

For 2M input and 500K output tokens, before cache hits:

Model	Cost per task
Kimi K3	$13.50
GLM-5.2	$5.00
DeepSeek V4 Pro	$1.305
DeepSeek V4 Flash	$0.42

K3 is 2.7 times the GLM-5.2 example cost and more than 10 times the DeepSeek V4 Pro cost. It needs a measurable advantage in multimodal understanding, long output, tool execution, or reduced retries to justify that gap.

Capability Differences

Kimi K3

K3 accepts text, image, and video, supports function calling and strict structured output, and can return exceptionally long completions. Its automatic prefix cache has no separate cache ID or TTL. The API currently always reasons at max.

Moonshot reports 2.8T total parameters with 16 of 896 experts active per token. Treat that architecture as context, not a quality guarantee.

GLM-5.2

GLM-5.2 is positioned for long-horizon coding and agent work with a 1M context window and 128K maximum output. At $1.40/$4.40, it is the middle option here: far cheaper than K3 output, but more expensive than DeepSeek V4.

DeepSeek V4

V4 Flash is the cheap scout and high-throughput route. V4 Pro costs more but is the better comparison for harder coding and agent steps. Their extremely low cache-hit rates make them attractive when a workflow repeats stable repository context.

The current canonical DeepSeek entries are text-first. Choose K3 when native image or video input is a core requirement rather than an occasional pre-processing step.

Recommended Multi-Model Routing

Agent step	Starting route
File search, classification, log summary	DeepSeek V4 Flash
Normal code patch or tool loop	DeepSeek V4 Pro or GLM-5.2
Very long engineering execution	GLM-5.2
Image/video-grounded task or output above 384K	Kimi K3
Difficult failure after cheaper routes	Kimi K3, then proprietary frontier fallback

Do not route solely by benchmark rank. Capture first-pass success, valid tool calls, wall-clock time, retries, cache hit rate, total output tokens, and human repair minutes.

Bottom Line

K3 is not the value winner on raw token price. It is the capability expansion: native visual input, dynamic tools, and a completion ceiling that the other models do not match. GLM-5.2 is the balanced long-horizon route, while DeepSeek V4 is the economic default for text-heavy traffic.

Read the K3 API guide and the broader Chinese AI model pricing comparison, then test your workload in the AI Pricing Calculator.

Official sources checked July 19, 2026: Kimi K3 launch, K3 API quickstart, Kimi global pricing, Z.AI pricing, GLM-5.2 docs, and DeepSeek pricing.