Breaking

GLM-5.2 Matches Sonnet 5 on Benchmarks at a Fraction of the Price

July 9, 2026 at 13:21 EDT

Foundation Models
Open Source
Software Dev & Coding

Zhipu AI's (Z.ai) open-weight model GLM-5.2 posts benchmarks nearly on par with Anthropic's Claude Sonnet 5 while costing far less per input and output token and shipping under an MIT license that allows self-hosting. Since the two models arrived in quick succession in mid-June 2026, developers have kept comparing their closeness on coding and agentic tasks against the gap in cost.

June 2026 · Open Weights vs Proprietary

GLM-5.2 Matches Claude Sonnet 5 — At a Fraction of the Price

Zhipu AI's MIT-licensed, self-hostable model runs neck-and-neck with Anthropic's proprietary flagship on coding and agentic benchmarks — while charging a fifth of the token cost.

~1/5

the output-token price of Sonnet 5

~80%

of the performance, as framed by users

token context — matched by both models

Output cost per 1M tokens

Representative upper-range levels drawn to scale — same unit, USD.

~$4.40

GLM-5.2

~$15

Claude Sonnet 5

Input tokens tell the same story: ~$0.75–1.40 vs ~$2–3.

Benchmarks: a razor-thin split

The advantage flips depending on the task.

SWE-bench Pro — Sonnet 5 edges ahead

GLM-5.2

62.1%

Sonnet 5

63.2%

Terminal-Bench 2.1 — GLM-5.2 leads

GLM-5.2

81.0%

Why pick GLM-5.2

MIT license — self-host & commercial use
~1/5 the token cost at high volume
Runs on vLLM / SGLang, mixed local + cloud
Beat Claude in a cyber benchmark

Why pick Sonnet 5

Stronger single-shot reliability
Better visual judgment & complex agents
Mature ecosystem (Claude Code, computer use)
GLM can produce more bugs in practice

With benchmarks nearly tied, the basis for choosing shifts from raw performance toward a combined judgment of price, license, and ecosystem.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.