‹ rbtfl.

Zhipu AI releases GLM-5.2 open-weight under MIT licence, beating GPT-5.5 on coding at one-sixth the cost

The 744B-parameter sparse MoE model with a real 1M-token context tops SWE-bench Pro among open-weight models and undercuts frontier closed-weight pricing sharply; MIT licence removes every regional restriction

AI· active The Long Game·Whose Money ·8 takes · ·rbtfl upd Jul 2, 2026

Summary

Zhipu AI released GLM-5.2 on June 13 under an MIT licence with no regional restrictions, a 744B-parameter sparse MoE model with 40 billion active parameters per inference and a real 1M-token context window. Independent benchmark verification found it scores 81.0 on Terminal-Bench 2.1 and 62.1 on SWE-bench Pro, beating GPT-5.5 on both long-horizon coding tests. Via OpenRouter it costs approximately $1.40 per million input tokens versus $5/M for GPT-5.5 and roughly $7/M for the then-available Anthropic frontier model. CNBC's June 26 report placed GLM-5.2 within one percentage point of the US frontier on agentic benchmarks. The MIT licence is the strategically significant choice: any country, company or developer can deploy or fine-tune without a US export control licence, directly circumventing the mechanism used to restrict access to Anthropic's Fable 5 six days before GLM-5.2's release.

The split

US AI press read the release primarily as a capability benchmark, asking whether a Chinese open-weight model had genuinely matched the frontier closed models. Chinese and Hong Kong coverage framed it as a geopolitical counter-move, noting the timing immediately after the Fable 5 ban and the deliberate MIT choice. The developer community on OpenRouter treated it as a cost-performance arbitrage opportunity, with traffic climbing faster than after DeepSeek V4. South Korean and Japanese enterprise press noted the 1M-token window as a direct challenge to Anthropic's long-context claim in the enterprise market.

By the numbers

  • 744B parameters, 40B active per forward pass (MoE architecture)
  • 1M tokens, the genuine context window (not a claimed figure reduced in practice)
  • 81.0, Terminal-Bench 2.1 score (above GPT-5.5)
  • 62.1, SWE-bench Pro score (above GPT-5.5)
  • $1.40/M, OpenRouter input token cost vs. $5/M for GPT-5.5
  • MIT licence, no regional restrictions on deployment or fine-tuning

Why it matters

The three-way convergence of capability parity, cost advantage, and MIT licence changes what is possible outside the US AI ecosystem. Countries that cannot access Fable 5 or GPT-5.6 due to US export controls or government restrictions can now deploy a frontier-class open-weight model. For enterprise buyers, GLM-5.2 creates a credible cost-performance alternative to every major closed-weight model on the market. The geopolitical implication is that the Fable 5 ban, intended to restrict frontier AI diffusion, may have accelerated Chinese labs' commercial case for open-weight frontier releases.

What to watch

  • Whether GLM-5.2 adoption accelerates in markets where US models face restrictions (Russia, Iran, North Korea, parts of the Global South).
  • Anthropic, OpenAI, and Google DeepMind competitive response, particularly on open-weight versus closed-weight strategy.
  • Whether the US government moves to bring MIT-licensed open-weight Chinese models under export control frameworks.
  • GLM-6 roadmap announcements and whether Zhipu follows with further open releases.