Claude Fable 5 Restricted, GLM 5.2 Released Open-Source: The Open-Frontier Gap Just Got More Complicated
textak moved its open-source frontier parity forecast from 72% to 75% — and today's news explains both why that move was warranted and why we're not going further. Hours after the US government restricted Anthropic's Claude Fable 5 on June 13, 2026, Chinese lab Zhipu AI released GLM 5.2 under MIT license: a 744-billion-parameter model with 1M-token context, directly challenging the export control logic that justified the restriction. The timing is almost certainly not coincidental. The forecast question — whether open-source matches closed frontier performance — just got injected with geopolitical fuel.
Our 75% reflects three things we've weighted heavily: Meta's sustained Llama investment, the verified 100x compute cost reduction that lets open-source labs do more with less, and the narrowing benchmark gap across standard evals. What drove the move from 72% to 75% was the Llama 4 generation demonstrating that the architectural gap was closeable at the top end, not just the mid-tier. That reasoning still holds.
But the GLM 5.2 story introduces a complication our model hasn't fully absorbed. The relevant question for our forecast is whether GLM 5.2 achieves verified parity with closed frontier models — not just parameter count or context window claims. The benchmark saturation story is directly relevant here: MMLU is now meaningless at 88%+ for all frontier models. Humanity's Last Exam shows frontier models at only 35% versus 90% for human domain experts. If that 55-point gap is real, then parameter scale alone — even 744 billion — doesn't close it. Our forecast requires parity on the evals that still discriminate, not the ones that have hit ceiling effects.
The Anthropic leak of 'Mythos' representing step-change improvements is the counterevidence we're most uncertain about. Leaked capability claims from frontier labs are notoriously unreliable as public signals, but the structural reality is that closed labs hold post-training techniques closely, and those techniques — not raw parameters — may be where the real gap lives. GLM 5.2's MIT license is significant for accessibility and for competitive pressure on US labs, but an open model that can be downloaded doesn't automatically have the RLHF stack and tool-use fine-tuning that makes closed models perform on hard tasks.
The US government's decision to restrict Claude Fable 5 hours before the GLM 5.2 release raises a question our forecast doesn't currently address: if export controls accelerate the open-source alternative's development outside US jurisdiction, does the 'open-source model' in our forecast include Chinese open-source releases? GLM 5.2 under MIT is, by definition, open-source. If it achieves verified parity, it resolves our forecast YES — but that's a different geopolitical story than the Meta/Llama narrative we built the thesis around. We're holding at 75% because the direction is confirmed, but we want to see independent technical evaluation of GLM 5.2 on Humanity's Last Exam before considering a further move. That evaluation, if it appears in the next 60 days, is our primary update trigger.