💪 Anthropic closed Fable 5 and Mythos 5 to foreign users. I found a replacement - and honestly, my stack only won from it
What GLM-5.2 is:
An open model from Z.ai (formerly Zhipu AI) with ~750B parameters, MoE architecture and a 1M-token context window. Released June 16 under the MIT license with open weights on HuggingFace - no regional restrictions and no risk of it getting "switched off" by a regulator's decision. Leader among open-weights models on the Artificial Analysis Intelligence Index, 2nd place on the Code Arena WebDev leaderboard (after Fable 5). Kilo Code benchmark on a planning task: Fable 9.1 vs GLM 9.0 - at 10x lower cost
So I'm not just talking - I looked at the alphaXiv case: a service for automatically reproducing scientific experiments. It also lost Fable and switched to GLM-5.2. In the demo, the agent independently worked on an RL experiment from a fresh paper on two servers with 8xH100 each, within the SkyRL framework
What the GLM-5.2 agent did without an operator:
💡 Fixed the environment on its own - spotted a dependency conflict, picked a strategy and applied the fix without any hints
💡 Compared two RL training modes - fully asynchronous and collocated-synchronous, on competitive programming tasks
💡 Put together the final report - autonomously analyzed throughput and reward stability for both setups
There's one honest limitation: no vision. Where Fable reads a chart as an image - GLM has to parse the numbers through numpy code. For most of my tasks it's not critical, but I'm not going to keep quiet about it
My stack right now - three models, three jobs:
⚡️ Claude Opus 4.8 - complex reasoning, architecture decisions, long context chains
⚡️ Codex 5.5 - autonomous coding, tasks I launch and let go
⚡️ GLM-5.2 - agent pipelines and tasks that need a model that doesn't depend on the mood of American regulators
It's exactly this combo that took my development to another level. Three providers, different strengths, and - most importantly - if one stops being available, the system doesn't go down
🎯 For six months I thought one top closed provider was enough and there was no need to complicate the stack. Turns out, no. An open model with an MIT license isn't a backup option. It's insurance against decisions made without you
How this principle lives in my system now
The specific models have changed since then, but the rule stayed: no part of the system should depend on a single provider. My personal agent Jes runs on OpenAI Codex and a second model. If the main one fails or hits a limit, the backup takes over by itself, without me. Jes has been running for 134 days and handles 10 scheduled jobs, and it has survived model changes in that time without stopping
For development I have 8 Claude Code agents working every day, with different models for different roles: Sonnet 5 where speed and volume matter, and Opus 5.5 where depth is needed. It's the same logic as the stack in this note: each model gets its own job, not one model for everything
What this gives a business
For clients I set systems up on their own server and their own AI subscription. That's how assistants run for five people, and a clinic owner has 6 agents, with a voice entrance to them running on his own ChatGPT subscription. If tomorrow some model gets closed for their country or goes up in price, one setting changes, while memory, rules and the schedule stay in place
Checking your own system is simple: imagine your main model disappears tomorrow morning. If the answer is "everything stops", you need a backup. It's cheaper than it seems, and much cheaper than a day of downtime. Test the backup ahead of time, not on the day of an outage: switch to it once a month and see whether it handles your usual tasks
🔥 Plug in GLM with my referral - link