If the June 12 export-control shutdown of Claude Fable 5 and Mythos 5 left you needing a working replacement, the most interesting answer arrived just 48 hours later. Chinese lab Zhipu AI — operating internationally as Z.ai — released GLM-5.2 on June 13, 2026: an open-weight, MIT-licensed model with a usable 1-million-token context window, priced at roughly a tenth of what comparable frontier access costs from Anthropic or OpenAI. The timing was not subtle, and Z.ai framed the release around the idea that frontier intelligence should be freely available.
What GLM-5.2 actually is
GLM-5.2 is a Mixture-of-Experts model with around 744 billion total parameters, of which roughly 40 billion activate for any given token — which keeps inference costs far lower than a dense model of the same size. It is built coding-first and agent-first: Zhipu positions it for long-horizon, multi-step software work rather than general chat.
- Context window: 1,000,000 tokens (model ID
glm-5.2[1m]) — five times the ~200K window of GLM-5.1, with output up to 131,072 tokens per response. - Two reasoning modes: “High” for faster everyday work and “Max” for complex multi-step coding. Zhipu recommends Max as the default for serious coding.
- License: MIT, no regional restrictions — self-host, fine-tune, and use commercially. Weights are published on Hugging Face under
zai-org/GLM-5.2. - Day-one agent support: works out of the box with eight agentic coding environments including Claude Code, Cline, OpenCode, and Roo Code.
How it compares to Opus 4.8 and GPT-5.5
Zhipu shipped GLM-5.2 with no benchmarks at launch, then published an official table on June 17 — figures that are vendor self-reported and still awaiting fully independent confirmation. On the numbers available so far, it lands as the strongest open-weight coding model on the market, close to (but not ahead of) the closed-source leaders on the work that matters most to developers:
- SWE-bench Pro: GLM-5.2 at 62.1 — ahead of GPT-5.5 (~58.6) and its own predecessor (58.4), trailing Claude Opus 4.8 (~69).
- Terminal-Bench 2.1: 81.0, within a few points of Opus 4.8.
- Artificial Analysis Intelligence Index v4.1: 51 — the highest of any open-weight model to date, an +11-point jump over GLM-5.1.
Fast.ai co-founder Jeremy Howard described it as at least as good as Opus 4.8 and GPT-5.5 for his own use, with the main gap being its lack of vision support. Treat that as a strong practitioner signal rather than a settled benchmark — multimodal support is unconfirmed, so plan around it as a text-and-code model for now.
The real headline: price
Where GLM-5.2 separates itself is cost. Via providers such as OpenRouter, metered API access runs around $1.40 per million input tokens and $4.40 per million output — against roughly $5/$30 for GPT-5.5 and $5/$25 for Claude Opus 4.8. That is a multiple cheaper on the work where token volume actually adds up: long, agentic coding sessions. The flat GLM Coding Plan starts at roughly $18/month for the entry tier.
Should you switch from Fable 5?
For high-volume coding, repo-scale refactoring, and long-context work, GLM-5.2 is the most compelling open replacement available right now — and the only one in this tier you can fully self-host to sidestep both the export-control issue and the question of where your data is processed. For the hardest abstract reasoning, it still trails Claude Opus 4.8, so the smart pattern is routing: GLM-5.2 for the bulk coding and long-context jobs, a frontier model kept in the loop for the reasoning-critical work. One caveat worth flagging — the standout benchmark numbers are vendor self-reported, and the hosted API is in China, which matters if data residency is a constraint for you.
Related reading
- Claude Fable 5 Banned: US Export Control Explained — the full breakdown of the shutdown and restoration odds
- Best Claude Fable 5 Alternatives While It’s Banned — how GLM-5.2, GPT-5.5, and Opus 4.8 stack up by use case