Is GLM-5.3 better than Kimi K3 and DeepSeek V4 for coding?

On long-horizon agentic coding, GLM-5.3 (released August 14, 2026) posts the strongest claims — Terminal-Bench 3.0 jumped from 4.6 to 28.3 and Z.ai reports leads on CyberGym and AutomationBench. But its open weights are staged, not yet released, while Kimi K3 (weights out July 27) and DeepSeek V4/0731 (MIT license, August 2026) are self-hostable today.

Which is cheapest: GLM-5.3, Kimi K3, or DeepSeek V4?

DeepSeek V4 Pro is cheapest at about $0.435/$0.87 per million tokens off-peak (2x during peak hours). GLM-5.3 is $1.40/$4.40 with $0.26 cached input. Kimi K3 is the most expensive at a flat $3/$15. For subscription coding, the GLM Coding Plan starts at $18/month.

Which open-weight model has the best license?

DeepSeek V4/0731 — its weights shipped under the MIT license in August 2026, the most permissive of the three. Kimi K3's weights have been out since July 27, 2026. GLM-5.3's weights are staged and were not available at its August 14 launch.

Quick Answer

GLM-5.3 vs Kimi K3 vs DeepSeek V4: Coding (Aug 2026)

Published: August 15, 2026

The Short Answer

As of August 15, 2026: GLM-5.3 (Z.ai, Aug 14) makes the strongest agentic-coding claims of China’s open-line trio, but it launched API-first with weights staged. DeepSeek V4/0731 is the cheapest and has the cleanest license (MIT). Kimi K3 is the pricier flat-rate all-rounder with weights available since July 27. If you need self-hosting today, GLM-5.3 isn’t yet an option — that’s the deciding fact this week.

Quick Comparison

	GLM-5.3	Kimi K3	DeepSeek V4 Pro
Vendor	Z.ai	Moonshot	DeepSeek
API price (per MTok)	$1.40 / $4.40	$3 / $15 flat	$0.435 / $0.87 off-peak (2× peak)
Cached input	$0.26	—	~$0.043
Open weights	Staged, not out	Yes (Jul 27, 2026)	Yes, MIT (Aug 2026)
Architecture	~743B base (GLM-5.2 base, post-trained)	Frontier MoE	284B MoE, ~13B active
Coding claim	Terminal-Bench 3.0: 28.3 (from 4.6); leads CyberGym, AutomationBench	Strong agentic coding since July	Strong coding per $, best value
Quirks	Mandatory thinking (low/high/max)	Flat pricing, no peak windows	Peak pricing 1-4 & 6-10 UTC

GLM-5.3 — Biggest Claims, No Weights Yet

Z.ai’s GLM-5.3 (August 14, 2026) runs the same ~743B base as GLM-5.2 with all gains from scaled post-training. The Terminal-Bench 3.0 jump from 4.6 to 28.3 is the headline — long-horizon terminal agents are exactly where cheap models used to fall apart. At $1.40/$4.40 it’s priced well below Western frontier coding models. The catch: weights are staged, thinking is mandatory, and until weights land you’re on Z.ai’s API or the $18/month GLM Coding Plan. Details: What is GLM-5.3?

Kimi K3 — Flat-Rate, Weights Available

Kimi K3 ($3/$15 flat, weights released July 27, 2026) skips peak/off-peak games — the same price at any hour, which makes cost forecasting trivial. It has been the open-weight agentic benchmark leader for much of the summer. It’s now the most expensive of the three per token, so its case rests on consistency and its established tooling ecosystem.

DeepSeek V4 — Best Value, Best License

DeepSeek V4/0731 is the price floor: $0.435/$0.87 off-peak (2× during peak, 1-4 and 6-10 UTC) with cache hits near $0.043. In August 2026 DeepSeek shipped the weights under MIT — 284B total, ~13B active — making it the most permissively-licensed frontier open model of the year. For high-volume agent loops where cost per task dominates, V4 remains the default pick.

Which Should You Pick?

Best agentic-coding ceiling right now: GLM-5.3 (accept API-only for now).
Self-host today with the cleanest license: DeepSeek V4/0731 (MIT).
Flat, predictable billing: Kimi K3.
Cheapest per task: DeepSeek V4, especially off-peak with cache hits.

Last verified: August 15, 2026.