Breaking

Claude Sonnet 5 Rated Below GLM 5.2 in Shared Coding Tests

June 30, 2026 at 19:50 EDT

Foundation Models
Software Dev & Coding
Open Source

A hands-on evaluation of Anthropic's mid-tier "Claude Sonnet 5," launched around June 30, 2026, has sparked debate among developers after it was rated "worse than GLM 5.2 across the board" on coding-oriented custom tests. The evaluation used several benchmarks, including a custom agentic/coding test known as the "Monica's apt test." The accompanying comparison video showed Sonnet 5 improving over the prior Sonnet 4.6 while falling short of GLM 5.2, an open-weight model from Zhipu AI (Z.ai), on several items.

Open-weight vs Proprietary · Coding LLM Showdown

Open-weight GLM-5.2 outscores Claude Sonnet 5 — at a fraction of the price

In the informal "Monica's apt test" and several other evals, Anthropic's new Sonnet 5 was judged clearly worse than China's MIT-licensed GLM-5.2 — which costs far less and can be self-hosted.

3.3×

Cheaper output tokens — GLM-5.2 from ~$3 vs Sonnet 5's ~$10 per 1M

39%

GLM-5.2 F1 on a cyber benchmark vs Sonnet's 32% (IDOR detection)

~1M

Context window for both — plus function calling & structured output

OUTPUT TOKEN PRICE — per 1M tokens

Taller column = more expensive. Sonnet 5 costs over 3× as much.

~$3

GLM-5.2

open-weight

~$10

Claude Sonnet 5

proprietary

THE "MONICA'S APT" TEST — recreate a Friends apartment in threejs

A single spatial-reasoning + code-generation challenge: build the floor plan accurately in 3D. The reviewer's verdict was clear.

GLM-5.2 — superior

Better layout fidelity and furniture arrangement

Sonnet 5 — underwhelming

Flagged for door placement and chair orientation errors

The case for open-weight

MIT-licensed weights allow self-hosting and fine-tuning; slots into Cursor or Claude Code as an Opus alternative — strong on coding, agentic and security tasks for far less.

The caveats

Sonnet 5 still leads narrowly on some benchmarks; one spatial test isn't the whole picture, and results hinge heavily on how the tool harness is built.

Continue reading

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