WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-06-29 1 00:00:04.500 --> 00:00:10.740 An open model out of Beijing caught Claude this week, and you can download the weights right now. 2 00:00:10.740 --> 00:00:20.580 It's Monday, June twenty-ninth. Here's the rundown: how the gap to the closed labs collapsed, the method that did it, and what it means if you're paying by the token. 3 00:00:20.760 --> 00:00:44.440 A security firm, Semgrep, ran its cyber evaluation on Zhipu's open model, GLM 5.2, and posted the results under the title 'we have Mythos at home.' The joke is the story. Their open model traded blows with Claude on the tasks they test, at a fraction of the cost. 4 00:00:44.440 --> 00:01:00.760 One vendor's benchmark, and cyber evals are noisy, so hold it loosely. But the direction isn't in doubt anymore. The best open model in the world comes from China, ships its weights, and it's close enough to Claude to argue about. 5 00:01:00.760 --> 00:01:09.640 And that breaks the pricing. For two years the pitch for a closed model was, it's just better, so the premium's worth it. 6 00:01:09.640 --> 00:01:24.600 When an open model you can download matches it on the work your engineers actually do, that pitch needs a second sentence. The capability stopped being the product. What's left to sell is the wrapper: the uptime, the integrations, the support. 7 00:01:24.600 --> 00:01:35.640 We should own a miss on this. Eight days ago we put Zhipu on the tape as a short. GLM 5.2 makes that wrong. 8 00:01:35.640 --> 00:01:50.360 What we underweighted is the whole game now. Open models are catching up by distilling the closed ones. Copy the outputs, train a smaller model to imitate them, give it away. You don't have to out-research the frontier if you can copy it cheaply. 9 00:01:52.060 --> 00:02:01.260 Put it next to Sunday. Our last edition was China taking the supercomputer crown back and shipping its own server CPU. 10 00:02:01.260 --> 00:02:16.060 Same story, one layer up. Beijing is building a stack it controls end to end. The chips, the memory, and now models that match the West, and it's handing the software layer out for free while everyone else meters it by the token. 11 00:02:16.060 --> 00:02:18.140 So what do you do with that. 12 00:02:18.140 --> 00:02:33.660 You can stand up a Claude-class model on your own hardware, behind your own walls, at about a tenth of the per-token cost, and nobody can rate-limit you or pull your access. That's real leverage. The cost is that you're running it yourself. 13 00:02:34.800 --> 00:02:46.880 To the tape. We covered the Zhipu short to watch, because being right about the weak monetization doesn't help if the model wins. We're watching Nvidia, Alphabet, and Anthropic. 14 00:02:46.880 --> 00:03:07.360 Anthropic is the exposed one. Its whole pitch is the best model, and an open model just caught it the same week it was rationing capacity. Highest conviction is Alphabet, on the read that Google's own chips and its open Gemma line let it absorb a price war. The falsifier: if Google cedes price-sensitive share to open models anyway. 15 00:03:07.360 --> 00:03:11.760 The tape is the desk's scorecard, not advice. 16 00:03:15.640 --> 00:03:17.800 Quick break — two from the desk. 17 00:03:17.800 --> 00:03:33.320 One we know well: vote dot direct. If you're on an H O A or a board, it runs your elections digitally — secure, verifiable, no paper, no clipboard in the lobby. Point your council to vote dot direct. 18 00:03:33.320 --> 00:03:45.480 And if this is your ten minutes of A I for the day, get the written edition too. The full wire, free, every morning — leave your email at nextbig dot dev. 19 00:03:49.510 --> 00:04:07.270 Our call: within six months, a Chinese open-weights model holds the number-one slot on a major public leaderboard, the kind that ranks GPT and Claude and Gemini too, and holds it for a real stretch, not a single day. 20 00:04:07.270 --> 00:04:17.110 What proves us wrong: if by the end of December no open model out of China has held that top spot for more than a blip. It settles December twenty-ninth. 21 00:04:17.110 --> 00:04:29.110 The benchmark that should worry the closed labs isn't GLM's score. It's the first invoice a team writes after it stops paying by the token. That's the rundown.