What is GLM-5.2 and when was it released?

GLM-5.2 is Z.ai's new flagship open-weight model, released to coding-plan subscribers on June 13, 2026 and made fully open-weight under an MIT license on June 16, 2026. It is a 753B-parameter Mixture-of-Experts model with 40B active parameters, 1 million token context window, and is text-only (no vision input). Z.ai positions it as the leading model for long-horizon software engineering tasks. On the Artificial Analysis Intelligence Index v4.1, GLM-5.2 scored 51, the highest open-weight score recorded, ahead of MiniMax-M3 (44), DeepSeek V4 Pro max (44), and Kimi K2.6 (43).

How does GLM-5.2 compare to Claude Fable 5 on benchmarks?

Claude Fable 5 still leads on closed-frontier intelligence with 64.9 on the Artificial Analysis Intelligence Index v4.1, roughly 14 points ahead of GLM-5.2 at 51. Fable 5 also tops Code Arena WebDev, with GLM-5.2 ranked second despite GLM-5.2 lacking image input. Where GLM-5.2 closes the gap is cost and licensing: GLM-5.2 is MIT-licensed and self-hostable; Fable 5 is closed and priced at $10/$50 per million tokens. For agentic coding workflows where the marginal task quality matters less than total cost across thousands of long-horizon runs, GLM-5.2 is competitive. For the hardest 10% of tasks — long agentic loops, vision-heavy work, multi-step reasoning with safety considerations — Fable 5 still wins.

How does GLM-5.2 compare to GPT-5.5 on cost and coding?

GLM-5.2 is roughly 4-7x cheaper than GPT-5.5 at the model layer. GPT-5.5 is priced at approximately $5/$30 per million input/output tokens. GLM-5.2 is available on OpenRouter from nine providers at roughly $1.40/$4.40 per million tokens. Z.ai's own reported coding benchmarks show GLM-5.2 outperforming GPT-5.5 on long-horizon coding tasks. The honest read: for everyday coding workflows, GLM-5.2 is the better economic choice; for tasks that require ChatGPT-side tooling, ecosystem maturity, or GPT-5.5's stronger reasoning on novel non-coding problems, GPT-5.5 still wins. Also: GLM-5.2 is token-hungry, using around 43k output tokens per Intelligence Index task versus 26k for GLM-5.1, so factor that into total-cost-of-ownership math.

Should I switch from Claude Fable 5 to GLM-5.2 today?

For most teams, no — not as a full switch. The cleanest pattern is to keep Claude Fable 5 (or Opus 4.8) for the hardest 10-20% of tasks and route the long-tail bulk of agentic coding to GLM-5.2 via OpenRouter or self-hosted. With Fable 5's free Pro/Max access ending June 22, 2026, and credit-based pricing kicking in June 23, this routing pattern becomes economically urgent. The risk: GLM-5.2 lacks vision input, which closes some workflows (UI screenshots, scientific figures, multi-modal agentic tasks). For those, you still need Fable 5, Opus 4.8, or GPT-5.5.

Quick Answer

GLM-5.2 vs Claude Fable 5 vs GPT-5.5: June 2026 Showdown

Published: June 19, 2026

GLM-5.2 vs Claude Fable 5 vs GPT-5.5: June 2026 Showdown

Z.ai released GLM-5.2 to coding-plan subscribers on June 13, 2026 and opened full weights under MIT on June 16, 2026. It is now the #1 open-weight model on the Artificial Analysis Intelligence Index at 51 points. Here is how it stacks up against Claude Fable 5 (the new closed-frontier leader) and GPT-5.5 (OpenAI’s current flagship), and how to think about routing between them.

Last verified: June 19, 2026.

TL;DR

GLM-5.2 is real. 753B-parameter MoE, 40B active, 1M context, MIT-licensed, released June 16, 2026.
Intelligence Index: Fable 5 leads at 64.9; GPT-5.5 at ~60; GLM-5.2 at 51 (top open-weight).
Cost: GLM-5.2 at $1.40/$4.40 per M tokens is roughly 4-7x cheaper than GPT-5.5 and Fable 5.
Code Arena WebDev: Fable 5 is #1, GLM-5.2 is #2 — impressive given GLM-5.2 has no vision input.
Best pattern: Keep Fable 5 / Opus 4.8 for the hardest 10-20%, route the long tail to GLM-5.2.

What GLM-5.2 actually is

GLM-5.2 is Z.ai’s seventh-generation flagship. The headline specs:

753B total parameters, 40B active (Mixture-of-Experts)
1.51 TB model weights
1 million token context window (up from 200K in GLM-5.1)
Text input only (no vision)
MIT license — fully open weights
Released to coding-plan subscribers June 13, full weights June 16, 2026
Available via OpenRouter from 9 providers

Z.ai’s pitch is “built for long-horizon tasks” — multi-step agentic coding workflows where the model needs to plan, edit, test, iterate over thousands of tokens without losing track. The 1M context window matters here.

Direct comparison

Feature	Claude Fable 5	GPT-5.5	GLM-5.2
Release	June 9, 2026	April 23, 2026	June 16, 2026 (open)
Lab	Anthropic	OpenAI	Z.ai
License	Closed	Closed	MIT (open weights)
Active parameters	Not disclosed	Not disclosed	40B (of 753B MoE)
Context window	1M	256K	1M
Vision input	Yes	Yes	No
AA Intelligence Index v4.1	64.9	~60	51
SWE-Bench Pro	80.3%	58.6%	Not officially reported
Code Arena WebDev rank	#1	—	#2
Input price per M	$10.00	$5.00	~$1.40 (OpenRouter)
Output price per M	$50.00	$30.00	~$4.40 (OpenRouter)
Self-hostable	No	No	Yes
Sovereign deployment	No	No	Yes

When Claude Fable 5 wins

Hardest 10-20% of coding tasks. Fable 5 leads SWE-Bench Pro by 11+ points over the next frontier model.
Vision-heavy work. Screenshot-to-code, scientific figure extraction, multi-modal agentic loops.
Long agentic runs where quality compounds. Fable 5’s GDPval-AA Elo is 1932 — a significant jump over Opus 4.8.
You are inside the free Pro/Max window through June 22, 2026. After June 23 it becomes credit-based.

When GPT-5.5 wins

You are in the ChatGPT / OpenAI ecosystem. Codex, GPT Store, Sora integration, Apple’s iOS Siri-AI default.
Reasoning on novel non-coding problems. GPT-5.5 still tops some scientific reasoning evals.
Voice and multimodal latency-sensitive flows. OpenAI’s Realtime API is more mature.

When GLM-5.2 wins

Cost dominates. Roughly 4-7x cheaper than both Fable 5 and GPT-5.5 at the model layer.
You can self-host or need sovereign deployment. MIT license, weights on Hugging Face.
Long-horizon agentic coding bulk. 1M context, optimized for multi-step engineering loops.
You are routing in an OpenAI-compatible harness. GLM-5.2 is on OpenRouter from 9 providers (Together, Hyperbolic, DeepInfra, Fireworks, and others) with drop-in OpenAI-compatible APIs.

The token-hungry caveat

Artificial Analysis flagged that GLM-5.2 uses roughly 43k output tokens per Intelligence Index task, up from 26k for GLM-5.1 and above MiniMax-M3 (24k), Kimi K2.6 (35k), and DeepSeek V4 Pro max (37k). Cheaper per token, but more tokens consumed per task. For total cost of ownership, the 4-7x cost advantage shrinks but does not disappear.

How to route between them (June 19, 2026 playbook)

Default to Claude Fable 5 through June 22 while the free Pro/Max window is open.
From June 23, route the bulk of agentic coding work to GLM-5.2 via OpenRouter or self-hosted. Keep Fable 5 / Opus 4.8 for the hardest 10-20% (hard SWE tasks, vision, agentic depth).
Use GPT-5.5 for ChatGPT-side flows where ecosystem matters more than capability.
Watch for GPT-5.6, which Polymarket assigns 83% probability of releasing between June 22-28, 2026, and which will likely reset OpenAI’s pricing competitiveness against GLM-5.2.

The honest read

GLM-5.2 is the strongest open-weight model in the world as of June 19, 2026. It is not the smartest model in the world — that title belongs to Claude Fable 5. But for teams that need open weights, sovereign deployment, or aggressive cost optimization on the long tail of agentic coding work, GLM-5.2 changes the routing math. Most production teams in late June 2026 will end up using two or three of these — Fable 5 for the hardest tasks, GLM-5.2 (or GPT-5.6 when it ships) for the bulk, and Opus 4.8 as a safety fallback.

The race is no longer “closed beats open.” It is “how cheaply can you serve the long tail without sacrificing the hardest 10%.”