GLM-5.2

Spec

Attribute	Value
Developer	Z.ai (Z.ai / Zhipu AI)
Released	2026-06-13 (subscribers) / 2026-06-16 (open weights + API)
Announced	unknown
Context window	1,000,000 tokens
Pricing	$1.40/M input · $4.40/M output — Z.ai's own endpoint on the OpenRouter catalogue, read 2026-08-01. Earlier revisions of this row quoted ≈$0.62–$0.77, which was not a price change: OpenRouter serves this model from 33 providers between $0.72 and $2.31 on input, and /models reports whichever one is routing at that instant. Four alerts in a day named $0.97, $1.19, $1.12 and $0.76 — Alibaba, SiliconFlow, an intermediate route, and StreamLake. Reseller quotes are real prices but they are not this model's price; quote the first-party figure, or quote a reseller with its name attached.
License	MIT (unrestricted, "no regional limits")
Availability	GLM Coding Plan (highest subscription tier), standalone API, open weights, NVIDIA NIM hosted
Total parameters	744B
Active parameters	~40B per token (Mixture-of-Experts)
Max output tokens	131,072
Reasoning modes	high, max effort levels

Release Date

• June 13, 2026: Available to Z.ai GLM Coding Plan subscribers
• June 16, 2026: Open weights published + standalone API launched

⚡ Note: Captured 27 days late (July 13, 2026 ingest). GLM-5.2's release period overlapped with the Grok V9-Medium (June 16) launch and was not surfaced in Tier-1 source polling during that week.

Architecture: IndexShare

GLM-5.2 introduces IndexShare — an architectural optimization to sparse attention that reduces per-token compute by 2.9× at the full 1-million-token context length compared to standard sparse attention. This is the primary reason the 1M-context capability is practical (not just theoretical).

Benchmarks

Benchmark	GLM-5.2	Notes
Code Arena	#2 globally	Trails Claude Opus 4.8 by ~1pp on 3 long-horizon coding evals
Long-horizon coding (vs GPT-5.5)	Beats GPT-5.5	At ~1/6th API cost
LMArena (2026-09-06)	6.23% ±0.77%, rank #10	(Max). The board's own percentage with a confidence interval, not Elo
2026-09-06 — first appearance on LMArena's visible leaderboard. GLM 5.2 (Max)
enters at #10, displacing Claude Opus 4.7. **No GLM row appears in any prior
top-10 snapshot this repo holds** (2026-08-09, 08-16, 08-23, 08-30 all have zero),
so this is a first, not a rise. The ±0.77% interval is the second-narrowest in
the top ten, most rows sitting at ±1.5% to ±2.1% — a statement about vote volume,
not about quality. The capture sees **only the top 10 the page renders
server-side**, so it says nothing about where GLM-5.3 or
GLM-5.3-Flash rank, and nothing read explains why the older model is the
one that surfaced (source).

• Described by Interconnects.ai as "the step change for open agents"
• Data risk note: TechTimes (June 17) — the Z.ai hosted API routes queries through Zhipu AI's China-based infrastructure. For data-sensitive workloads, self-hosting the open weights is recommended.

Use Cases

• Long-horizon agentic coding tasks (primary target — Code Arena #2)
• Cost-effective alternative to Opus 4.8 for organizations that can accept the licensing/data-routing tradeoffs
• Self-hosted deployment for organizations that need frontier-class open weights without data-routing risk

Compared To

Model	Code Arena / SWE-Pro	License	Cost
GLM-5.2	#2 (long-horizon coding)	MIT	~1/6th GPT-5.5 API cost
Claude Opus 4.8	#1 (long-horizon)	Closed	$15/$75 per Mtok
GPT-5.5	Below GLM-5.2 (coding)	Closed	Standard OpenAI API
Mistral Large 3	TBD (July 2026 GA)	Apache 2.0	TBD
Llama 4	Below GLM-5.2	Meta custom	Free (self-hosted)

Sources

• VentureBeat — open-weights GLM-5.2 beats GPT-5.5 on coding
• AI Weekly
• TechTimes — data risk warning
• Interconnects.ai
• Z.ai Docs
• (source)