GLM 5.1 Thinks Strategically, Data-Center Revolt Intensifies, When Helpful LLMs Turn Unhelpful, Humanoid Robots Get to Work
The Batch Analytics DeepLearning.AI ● Covered by 2 sources
Someone in the US actually fired shots at a data-center construction site over grid strain and noise. GLM 5.1 can now grind on hard tasks for eight straight hours without giving up.
Z.ai just shipped GLM-5.1, an open-weights model built less for snappy one-shot answers and more for staying power. Instead of stopping once it hits a token budget or decides further reasoning won't help, it loops through planning, execution, and self-evaluation for as long as eight hours, sometimes firing off thousands of tool calls before calling a task done. That's a meaningfully different design goal than most chatbots chase, and it shows: on SWE-Bench Pro, a benchmark built from real GitHub engineering problems, GLM-5.1 edged out GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro. It also topped Z.ai's own cybersecurity-reasoning tests, partly because rivals like Gemini and GPT-5.4 declined to run certain tasks on safety grounds.
The model isn't dominant everywhere. On pure reasoning and math benchmarks like GPQA Diamond and AIME 2026, it trails the big closed labs by a real margin. And Z.ai raised prices sharply this round — API costs jumped roughly 40 percent, coding-plan subscriptions basically doubled — even though it's still cheaper per token than Claude Opus 4.6. The bet, clearly, is that persistence on long agentic coding tasks is worth more to developers than shaving a few points off a math test.
Meanwhile, humanoid robots are quietly punching into paid factory work. Agility Robotics' Digit is now moving 25-pound parts bins at a Schaeffler auto-parts plant in South Carolina, running two four-hour shifts a day behind a plexiglass barrier because it can't yet detect humans nearby. Schaeffler bumped the person who used to do that job into a supervisory role and plans to add hundreds more Digits across its US and European plants by 2030. At $10 to $25 an hour against a $20 entry-level wage, the economics are close enough to matter, even if only around 200 humanoids are actually working in factories worldwide right now.
And back on the ground, literally, the backlash against data centers is getting louder and, in at least two cases, violent. Roughly $64 billion in projects got blocked or delayed by local opposition in under a year, and Maine's legislature just passed a moratorium on new facilities over 20 megawatts until 2027 — potentially the first statewide ban, with a dozen other states drafting similar bills. Complaints range from noisy cooling systems to spiking electricity prices to sheer size. None of this slows model training directly, but it's a real constraint on where the compute powering models like GLM-5.1 actually gets built.
My take
The data-center pushback is the most important story of the three, and everyone building AI seems to be underestimating it. You can ship the smartest model on the planet, but if neighbors are shooting at construction crews and states are passing moratoriums, compute capacity becomes a political bottleneck, not just an engineering one. Open-weights labs like Z.ai racing to squeeze more autonomy per dollar are, ironically, the ones best positioned to weather that — smaller, cheaper deployments need less land, less power, and less local goodwill than another hyperscale campus.
Read more about this at: The Batch
