GLM-5.2

Z.ai‘s text-only open-weight flagship, released 2026-06-16 under an MIT license — on Simon Willison‘s read, the most capable text-only open-weight LLM available at release. A 753B-parameter Mixture-of-Experts model with 40B active params (1.51TB on disk) and a 1M-token context (up from GLM-5.1’s 200K). No vision — Z.ai ships separate multimodal models; this one is deliberately text-only.

Where it ranks

• #1 open-weight model on Artificial Analysis‘s Intelligence Index v4.1, scoring 51 — ahead of MiniMax-M3 (44) and DeepSeek V4 Pro (44).
• #2 on Code Arena WebDev, behind only Claude Fable 5 (claude-fable-5, a proprietary frontier model) — i.e. an open-weight model trailing only the closed leader on a coding leaderboard.

Why it matters here

GLM-5.2 is the sharpest data point yet for the spoke’s central question — does the frontier premium survive? (synthesis). An MIT-licensed open-weight model now tops the independent open-weight ranking and sits second to the closed leader on web-dev coding, at roughly a fifth of the price:

• ~$1.40 / 1M input · ~$4.40 / 1M output via OpenRouter hosts — against GPT-5.5 at $5 / $30 and Claude Opus 4.5–4.8 at $5 / $25 (llm-api-pricing). The open-weight wedge (deepseek, qwen) keeps prying the floor down.

Caveat — token-hungry. It burns ~43,000 output tokens per task on the Intelligence Index, well above comparable models, so the headline per-token price understates real cost: the “cost-is-dominated-by-output-tokens” thread (synthesis → Recurring reads) bites here. A cheap per-token rate × heavy token use narrows the apparent gap to pricier-but-terser frontier models.

Tier & provenance

T2 — independent practitioner review (simon-willison, a methodology-transparent secondary source) citing first-party release facts and the independent artificial-analysis benchmark. Not first-party Z.ai documentation. freshness: volatile — rankings/pricing are dated snapshots (2026-06-18) that churn weekly. Willison’s qualitative probe was his SVG-generation test: the “pelican on a bicycle” improved over prior versions, but an opossum render regressed versus GLM-5.1 (“didn’t even try to animate it”) — a reminder the benchmark lead is not uniform across tasks.

Contested: is $1.40/$4.40 the vendor’s price or a reseller’s? (2026-07-28)

This page records ~$1.40 in / ~$4.40 out per 1M via OpenRouter hosts, on Willison’s attribution. deepseek-glm-qwen-price-gap (T4) presents the same pair of numbers as Z.ai’s own list price, and then quotes third-party resale separately at ~$0.55 / $1.85 — which would make GLM-5.2 roughly half the price recorded here when bought through a host, not the reseller rate itself.

Both claims are in the corpus; neither is Z.ai’s pricing page. The distinction isn’t academic: that T4 article’s headline “10x price gap” against DeepSeek V4 Flash is computed off $1.40 being the vendor rate. Flagged in synthesis and left open until a first-party source settles it.

Related

deepseek-glm-qwen-price-gap · z-ai · simon-willison · open-weight-models · artificial-analysis · llm-api-pricing · deepseek · qwen · llm-benchmarks