Artificial Analysis: four frontier launches in eight days (Grok 4.5, GPT-5.6, Meta Muse Spark 1.1, Kimi K3). Six labs now score above 50 on the Intelligence Index, up from two in early June. https://x.com/ArtificialAnlys
Read-through for your bill: half the new frontier is Chinese and open-weight. When six vendors clear a bar two cleared six weeks ago, your leverage is the routing dial, not vendor loyalty. Frames today's Qwen, Fable, and Kimi stories.
A system, not a freelancer: a 30-year operator (real estate, hospitality, food, a theater) directing a full team of AI agents at one client project at a time, strategy to ship. The agents do the work, one senior operator steers the fleet, you own the code and the keys at the end. When frontier capability changes hands every Sunday, you want someone driving. Book a 30-min call. https://outpace.media
The stories
LEAD (builder impact), Qwen 3.8 Max Preview is now testable (Alibaba, Sunday July 19-20)
DELTA from yesterday's rumor: you can now put hands on it via Alibaba Cloud + Qwen Chat. ~2.4T params; Qwen claims it beats every model except Fable 5; open weights "soon." Carried by TestingCatalog (confirmed testable), kimmonismus, Min Choi, Andrew Curran. CAVEAT: benchmarks are still Qwen's own, no independent Intelligence Index score or model card yet, so don't swap production.
The move: open Qwen Chat, run your single hardest recurring prompt on it vs your Claude baseline, log cost + quality. When weights drop, you already know if it's your cheap-frontier route.
Max / Team Premium: Fable 5 now included permanently at 50% of limits (set a Sonnet 5 fallback, done). Pro / Team Standard: Fable LEFT your subscription today, one-time $100 credit, then credits-only at $10/M in, $50/M out (2x Opus). Nathan Lambert reports it pulled a couple days early; Simon Willison: "Fablepocalypse permanently cancelled" (for Max). Delta since prior coverage: this is the fork going live, not a projected date.
The move: Max, fallback and forget. Pro, spend the $100 deliberately (see Use Cases).
(a) K3 on web + mobile now has 3 reasoning effort levels: Standard, High, Max (per-task thinking dial). (b) New subscriptions PAUSED, K3 demand broke their compute; existing subscribers unaffected, new signups waitlisted; plans splitting into Kimi Membership + Kimi Code Membership. Carried by TestingCatalog, Andrew Curran. API still up; open weights land July 27.
Run the Qwen 3.8 bake-off against your Claude baseline; if you're on Pro, spend the $100 Fable credit distilling one skill before it drains. ~20 minutes; the output outlives the model.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: A two point four trillion parameter model went live for anyone to test this weekend, it reportedly beats G P T five point six Sol, it is about to be open weight, and it shipped from Alibaba on a Sunday night. The smartest tier of AI is not a members-only club anymore.
0:18Nova: This is Mainframe, your daily guide through the AI chaos: what actually happened, who is winning, and how to take yourself to the next level. It is Monday, July twentieth, twenty twenty-six. I am Nova.
0:33Nova: On today's Mainframe: Qwen three point eight Max lands as a testable preview, and the move is to bake it off against Claude before the open weights drop. Fable five forks today, if you are on Pro or Team Standard it just left your subscription, and I will tell you exactly what you get instead. Kimi K three grows a reasoning dial and then slams its own doors, new subscriptions are paused. Plus a multi-model workflow the sharpest builders are quietly running, and sixty-five thousand dollars of free agent credits if you are a founder.
1:09Nova: Quick note before we go deep: the voices you are hearing, Dex and me, are AI. The reporting, the picks, and the opinions are put together by humans. Now, the Board.
1:22Dex: The macro number today comes from Artificial Analysis. Four frontier launches in eight days: Grok four point five, G P T five point six, Meta's Muse Spark, and Kimi K three. Six labs now score above fifty on their Intelligence Index, up from just two in early June.
1:41Nova: Translation for your bill: the smart shelf got crowded fast, and half of it is Chinese and open weight. When six vendors clear a bar that two cleared six weeks ago, your leverage is the routing dial, not loyalty to any one of them. That frames the whole show today. Let's build.
2:03Nova: Today's episode is brought to you by Outpace. Here is the thing about a week where the model shelf reshuffles four times: most people freeze. The operator behind Outpace does the opposite. Thirty years building real businesses, real estate, hospitality, food, even a theater, and now he aims a whole team of AI agents at one client's project at a time, strategy to ship. Outpace is a system, not a freelancer, and not some faceless agency: the agents do the work, one senior operator steers the fleet, and you own the code and the keys at the end. When frontier capability changes hands every Sunday, you want someone driving. Book a thirty-minute call at outpace dot media. That is outpace dot media.
2:55Nova: To the Wire. Lead story: Qwen three point eight Max Preview is live for testing on Alibaba Cloud and Qwen Chat. We flagged this as a rumor yesterday, single source, unverified. Today is the delta: you can actually put your hands on it. Two point four trillion parameters, and according to Qwen it outperforms every model except Fable five. testingcatalog confirms it is testable now, kimmonismus, Min Choi, and Andrew Curran are all carrying it, and Andrew Curran adds it is going open weight soon.
3:33Dex: Caveat: the benchmarks are still Qwen's own. No independent Intelligence Index score yet, no model card.
3:40Nova: Right, so do not swap production today. But the move is concrete: open Qwen Chat, run your single hardest recurring prompt on it against your Claude baseline, and write down cost and quality. When the weights land, and Andrew Curran says that is soon, you already know whether Qwen is your new cheap-frontier route. You will be a week ahead of everyone who waits for the weights to start testing.
4:06Dex: Story two, and this one hits wallets today. Fable five forked this morning.
4:11Nova: It did. If you are on Max or Team Premium, nothing changes, Fable five is included permanently now at fifty percent of limits, set a Sonnet five fallback and forget it. If you are on Pro or Team Standard, and this is the sting, Fable left your subscription today. You get a one-time one hundred dollar credit, then Fable is credits-only at ten dollars per million in, fifty dollars per million out, twice the price of Opus. Nathan Lambert flagged it actually pulled a couple of days early. Simon Willison calls the permanent version the Fablepocalypse cancelled, for Max users at least.
4:53Dex: So the silver lining is real for Max, and Pro users have a hundred dollars to spend deliberately, not fritter.
4:59Nova: Exactly, and I will tell you how to spend it in Use Cases. Story three, Kimi K three, two fresh moves from Moonshot. One, K three on web and mobile now has three reasoning effort levels, Standard, High, and Max, so you dial how hard it thinks per task, same idea as Claude's effort control. Two, the access story: Moonshot paused all new subscriptions. Demand for K three broke their compute, existing subscribers are fine, new signups are waitlisted, and they are splitting into a Kimi Membership and a Kimi Code Membership. testingcatalog and Andrew Curran carried both.
5:39Dex: So if you were meaning to try K three's subscription, that window is shut for now.
5:44Nova: Shut, but the API is up and the open weights land July twenty-seventh, so you are not locked out, just rerouted. Tool Lab, two things you can pick up today. First, trending, and this one is for the founders in the audience: Cognition just launched Devin for Startups. Sixty-five thousand dollars in credits to run Devin across Cloud, Desktop, and the CLI. swyx surfaced it, Cognition is running it as an application.
6:12Dex: Sixty-five thousand dollars of autonomous-agent runtime is not nothing.
6:17Nova: It is a real budget to offload the background engineering, the tickets you never get to. The move: if you are early-stage, apply this week, then hand Devin your backlog of small, well-scoped fixes and let it grind them in the Cloud while you build the thing only you can build. Credit Cognition. Second pick, a hidden gem, and it is a pattern, not a product. Theo, of T3, published his exact setup for running G P T five point six Sol and Kimi K three as models inside Claude Code. Not instead of Claude Code, inside it: you keep the harness, the skills, the browser, the file editing, and you swap the underlying model per task.
7:01Dex: So you get Claude Code's tooling with a cheaper or faster brain when the task does not need Fable.
7:07Nova: That is the unlock. Claude Code is the best agent harness going, but you are not forced to pay frontier rates for every keystroke. Copy Theo's config, point routine coding at K three or Sol, keep Fable for the hard reasoning, and your bill drops without changing a single habit. Credit Theo.
7:27Nova: Use Cases, three moves with tools already in your hands. One, the multi-model workflow. Min Choi posted the exact chain the sharp builders are running: real-time research first, then Grok four point five for planning and orchestration, Fable five for day-to-day coding and debugging, Grok four point five again to write and run tests, G P T five point six Sol for the front end, back to Fable to finish. The lesson is not the specific chain, it is that you cast each model for the role it wins. Write your own version down as a checklist this week.
8:05Dex: Two, the Fable credit, for Pro users, time-sensitive as of today.
8:09Nova: Time-sensitive. You have one hundred dollars of Fable credit and it drains. Do not spend it on throwaway output. Run your single hardest recurring job on Fable once, then have it write down how it planned and structured the work as a Claude Code skill file, plain markdown that also runs on Sonnet five, Kimi K three, or Codex. You buy Fable's judgment once and keep it after the model leaves your plan. That is the iwoszapar pattern, and it is the highest-return twenty minutes a Pro user spends this week.
8:44Nova: Right. Before Qwen three point eight goes open weight, spend twenty minutes now: the same hard prompt on Qwen and on Claude, log cost and quality side by side, and set your routing config so the open-weight release is a one-line swap, not a project. Ready beats fast.
9:03Nova: This one is brought to you by SearchVis, a sibling to Outpace. Look at today: five different vendors could plausibly be your next model. When a buyer asks Claude, ChatGPT, or Perplexity for the best tool in your category, the model names one of them, and if it does not name you, you do not exist to that buyer. Mainframe tells you what shipped, SearchVis makes sure the answer engines know what you shipped. It tracks whether they cite your brand across every engine, tells you why, and hands you the exact page to publish to win the citation. Check where you stand and start free at searchvis dot outpace dot media. That is searchvis dot outpace dot media. Be the answer.
9:55Nova: One idea to carry into your week. On Sunday night, Fable five produced a hand-checkable counterexample to the Jacobian conjecture, an open math problem from nineteen thirty-nine. kimmonismus and Andrew Curran both carried it, and the mathematician alpoge confirmed the check. Eighty-seven years, solved casually on a Sunday.
10:18Dex: And the same model got rationed off half of Anthropic's plans the very same week.
10:23Nova: That is the whole point. Capability and access just decoupled. The best model on earth can do Fields-medal-adjacent math and also be the one you cannot afford by default, while an open Chinese model matches it for a fraction of the price. So stop anchoring your identity to one model. Raw intelligence is becoming abundant and multi-vendor. The scarce thing is taste: knowing which model to cast for which job. Min Choi's routing chain is worth more than any single subscription, because it is yours and it survives every reshuffle.
11:00Nova: So here is what I would actually do about it: keep a written routing map, one page, which model for which task, and update it every Friday. That map, not your loyalty to any lab, is the asset compounding while everything else reprices.
11:15Nova: Level up this week: run the Qwen three point eight bake-off against your Claude baseline, and if you are on Pro, spend that one hundred dollar Fable credit distilling one skill before it drains. Twenty minutes, and the output outlives the model.
11:31Dex: And if you want all of this in your inbox, every story, every number, every link, subscribe to the free Mainframe newsletter. It is the written brief for today's episode.
11:41Nova: That is Mainframe for Monday, July twentieth. I am Nova.