Which open-weight model is best for coding in July 2026?

It depends on the job. GLM-5.2 is the best all-round open pick for agentic and long-horizon coding (Terminal-Bench 2.1 81.0, top open model on the AA Intelligence Index). Kimi K3 leads frontend/UI and posts 93.4% SWE-bench Verified. DeepSeek V4 Flash is the cheapest frontier-class API for high-volume work.

How much do these open-weight models cost?

DeepSeek V4 Flash is cheapest at $0.14/$0.28 per MTok (off-peak); V4 Pro is $0.435/$0.87. GLM-5.2 runs about $1.40/$4.40 official (cheaper via third parties). Kimi K3 is the priciest at $3/$15, though heavy caching brings its effective blended rate near $2.31.

Which open model best replaces a closed frontier model?

GLM-5.2 is the closest drop-in for agentic and repo-scale work, reaching #1 on Design Arena's coding leaderboard and offering 76-78% lower cost per trace than Opus 4.8. For raw frontend quality, Kimi K3 rivals Opus 4.8 on simple tasks but is less reliable on complex ones.

Quick Answer

Kimi K3 vs GLM-5.2 vs DeepSeek V4: Open Coding 2026

Published: July 31, 2026

The Short Answer

All three are open-weight and self-hostable with 1M-token context. In July 2026: GLM-5.2 is the best all-round agentic coder, Kimi K3 leads frontend/UI and raw SWE-bench, and DeepSeek V4 is the value tier — Flash for cheap volume, Pro for balanced quality.

The Numbers

	Kimi K3	GLM-5.2	DeepSeek V4 Pro	DeepSeek V4 Flash
Price ($/MTok)	$3 / $15	~$1.40 / $4.40	$0.435 / $0.87	$0.14 / $0.28
SWE-bench Verified	93.4%	—	~91.2%	—
Terminal-Bench 2.1	~84.6 (indep.)	81.0	—	—
Context	1M	1M	1M	1M
Best at	Frontend/UI	Agentic/repo	Balanced quality	Cheap volume

(DeepSeek V4 prices are off-peak; they roughly 2x during peak UTC hours.)

Where Each Wins

GLM-5.2 → best open agentic coder. Top open model on the Artificial Analysis Intelligence Index, first open-weight model to hit #1 on Design Arena’s coding board, and Terminal-Bench 2.1 of 81.0. Kilo Code names it the overall pick for long-horizon agents and terminal work — at 76-78% lower cost per trace than Opus 4.8.
Kimi K3 → frontend and raw benchmarks. 93.4% SWE-bench Verified and #1 on Frontend Code Arena. Caveat: Semgrep found lower precision (0.684) on large interconnected codebases, and it’s verbose (~2x output), so effective cost climbs.
DeepSeek V4 Flash → cheapest frontier-class API. At $0.14/$0.28 it undercuts everything for high-volume, cost-sensitive pipelines. V4 Pro ($0.435/$0.87) steps up quality for complex reasoning.

Cost Reality Check

Sticker price misleads here. Kimi K3’s $3/$15 looks steep, but a 90% cache discount pulls its blended rate to ~$2.31/MTok on cache-heavy coding. GLM-5.2’s 93% cache hit rate lands an effective input price near $0.27. And DeepSeek V4 Flash’s cache-hit input of $0.0028 makes repeated-context agents almost free. Benchmark on your own cache profile.

Verdict

Best all-round open agent → GLM-5.2
Best frontend / highest raw score → Kimi K3
Cheapest frontier-class → DeepSeek V4 Flash (volume) / V4 Pro (quality)

Sources

Z.ai — GLM-5.2: z.ai/blog/glm-5.2
Moonshot AI — Kimi K3: kimi.com
DeepSeek — pricing: deepseek.ai/pricing
Artificial Analysis: artificialanalysis.ai