Moonshot AI (Kimi K3) told investors it is preparing a Hong Kong IPO within ~6 months and is closing a round valuing the 3-year-old startup at $30B+. ARR hit $300M in June, up from $200M in April. Reported by Bloomberg; carried by Andrew Curran and kimmonismus.
Read-through for your bill: open weights are now a fundable, profitable public business, not a charity experiment. The cheap-capable shelf you route to isn't going anywhere. Keep the routing dial; it's a permanent asset, not a hedge.
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 the shelf reshuffles three times before lunch, you want someone driving. Book a 30-min call. https://outpace.media
The stories
LEAD (builder stakes), Claude Code's 50% higher weekly limits EXTENDED to August 19 (announced July 19)
The promotion was set to expire tonight; Anthropic extended it through Aug 19 for all Pro, Max, Team, and seat-based Enterprise users. No action needed, auto-applied. Free plans and consumption-based Enterprise seats not eligible. Carried by ClaudeDevs, TestingCatalog.
The move: a month of extra headroom is use-it-or-lose-it. Pick your ONE deferred compute-heavy job (full eval suite, big refactor, hard skill distillation) and run it this week while the ceiling is high, don't fritter it on trivial prompts.
BUILDER STAKES, Fable 5 forks tomorrow (July 20); today is the last included-for-everyone day (ends 11:59:59 PM PT)
Silver lining up front: it's PERMANENT now, no more cliff. From July 20, Fable 5 stays included forever in Max and Team Premium at 50% of limits. Pro and Team Standard do NOT get it included: a one-time $100 credit, then credits-only at $10/M input, $50/M output (2x Opus 4.8). Confirmed/carried by Simon Willison ("Fablepocalypse permanently cancelled"), Nathan Lambert, kimmonismus, TestingCatalog.
The move: Max users, set a Sonnet 5 fallback and forget it. Pro users, spend the $100 credit deliberately (see Use Cases).
RUMORED (single-source, unverified), Qwen 3.8 Max preview
kimmonismus reports an open-weight preview (~2.4T params) "reportedly" beating GPT-5.6 Sol and trailing Fable 5 narrowly, testable as "Qwen 3.8 max-Max-Preview." No model card, no benchmark page, no independent score yet; kimmonismus himself says it "needs to be verified." VERDICT: mark rumored, swap nothing, but wire your routing dial so a real landing this week is a one-line config change.
Run your one deferred heavy job under the renewed Claude Code limits (through Aug 19); if you're on Pro, set your Fable fallback and distill one last skill before tomorrow's fork. ~20 minutes; the output outlives the model.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: The 50 percent higher Claude Code limits that were supposed to die at midnight tonight just got renewed through the middle of August, for every paid tier, so the headroom you have been rationing for two weeks is back for another month. 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 Sunday, July nineteenth, twenty twenty-six. I am Nova.
0:31Nova: On today's Mainframe: the lead, Claude Code's 50 percent limit boost, set to expire tonight, just got extended to August nineteenth for all Pro, Max, Team, and seat-based Enterprise users. Here is exactly what to do with a month of extra headroom.
0:47Dex: Builder stakes: Fable 5. Today is the last day it is included for everyone. Tomorrow it forks, and which side you land on decides whether it costs you nothing or costs you credits. We will hand you the hundred-dollar consolation prize up front.
1:02Nova: Anthropic's Workbench got a rebuilt Build section with a one-click fallback model, the ninety-second switch that keeps a blocked request from stranding you.
1:12Dex: And a Sunday rumor out of China: Qwen three point eight Max, reportedly beating GPT five point six Sol. We will tell you exactly how much to trust it.
1:26Nova: Quick honesty note before the Board: the voices you are hearing, Dex and me, are AI. The reporting, the picks, the opinions, all of it is put together by humans. Now, the Board, sixty seconds. Moonshot, the lab behind Kimi K3, told investors it is prepping a Hong Kong IPO within six months, closing a round that values the three-year-old startup north of thirty billion dollars. Annual recurring revenue hit three hundred million in June, up from two hundred million in April. Bloomberg has it; Andrew Curran and kimmonismus are carrying it.
2:05Dex: So a Chinese open-weight lab is turning one model launch into a public-market window. What does that do to my bill?
2:12Nova: It tells you the cheap-capable shelf you route to is not a charity experiment that vanishes next quarter. Open weights are now a fundable, profitable business. So the routing dial you have been building is a permanent asset, not a hedge against a fad. Keep it. That is the whole macro read for the day, and we are moving on.
2:33Dex: Brought to you by Outpace. Here is the thing about a week like this one: limits move, a model forks, a rumor drops on a Sunday. If you are a business owner, you cannot chase every one of those. Outpace is run by an operator who spent thirty years building real companies, real estate, hospitality, food, even a theater, and now points a whole team of AI agents at one client's project at a time, strategy through ship. It is a system, not a freelancer and not a 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 the shelf reshuffles three times before lunch, you want someone driving. Book a thirty-minute call at outpace dot media. That is outpace dot media.
3:21Nova: The Wire. Lead, and it is the reason you are still on your Pro plan smiling: Anthropic confirmed the 50 percent Claude Code weekly limit boost, which was scheduled to end tonight, now runs through August nineteenth. All Pro, Max, Team, and seat-based Enterprise users, no action needed, it is just applied. ClaudeDevs posted it; TestingCatalog carried it.
3:47Dex: So this is the same promotion, just extended again. Why does a limit extension lead the show?
3:53Nova: Because it is a use-it-or-lose-it asset and most builders let it evaporate. A month of extra weekly headroom is a month you can run the compute-heavy job you have been deferring: the full eval suite, the big multi-file refactor, distilling a genuinely hard skill. The move: pick your one deferred heavy job today and schedule it this week, while the ceiling is high. Do not spread it thin across trivial prompts.
4:21Dex: Story two, the builder-stakes one. Fable 5. Today, July nineteenth, is the last day it is included for everyone at eleven fifty-nine PT. Tomorrow it forks. Silver lining first, because we promised it: this is permanent now, nobody is racing a cliff anymore. Starting July twentieth, Fable 5 stays included, for good, in Max and Team Premium plans at 50 percent of limits.
4:45Nova: And if you are on Pro or Team Standard, you do not get it included. You get a one-time hundred-dollar credit, and after that Fable is credits-only at ten dollars per million input, fifty per million output, which is twice Opus 4.8. Simon Willison called the whole saga permanently cancelled; Nathan Lambert and kimmonismus confirmed the tiering. So Max users, you are set. Pro users, that hundred-dollar credit is the thing to spend deliberately, and we will tell you how in Use Cases.
5:20Dex: Story three, the Sunday rumor. kimmonismus is reporting Qwen three point eight Max, an open-weight preview around two point four trillion parameters, reportedly beating GPT five point six Sol and trailing Fable 5 by a hair.
5:34Nova: And here is the honest part. We could not corroborate it. It is one voice, on a Sunday, and kimmonismus himself said it needs verifying. There is no model card, no benchmark page, no Artificial Analysis score. So mark it rumored, do not swap anything, but wire your routing dial so that if it lands real this week, it is a one-line config change and not a scramble. That is the whole point of owning the dial.
6:02Nova: Tool Lab, two picks. First, trending, and it pairs perfectly with the Fable fork: Anthropic's Workbench got a rebuilt Build section. You send a Messages API request, see Claude's response, and copy the shippable code right there. The new bit that matters today: you can configure a fallback model for Fable 5. If a Fable request gets blocked by the safety classifiers or fails, it does not die, it re-runs automatically on Opus 4.8. First step: open the Console, go to Build, set your fallback under the primary model, or if you are on the API, pass the fallbacks parameter. Ninety seconds, and a blocked request never strands you mid-session.
6:48Dex: Second pick, hidden gem, and it is a speed one. Cerebras is now serving GPT five point six Sol on its wafer-scale hardware at up to seven hundred fifty tokens per second. For context, frontier models on normal GPU clusters run forty to a hundred twenty tokens a second, so this is roughly ten times faster. It is rolling to customers now; kimmonismus is calling projects finished ten times faster worth premium pricing.
7:17Nova: And the builder move is not to move everything there. It is to split by latency. The parts of your workflow where a human is waiting, live agent chat, interactive iteration, autocomplete-style loops, route those to Cerebras Sol so the person feels instant. Keep your hardest reasoning on Claude, where you are paying for judgment, not speed. Credit Cerebras; check their inference page for access.
7:45Dex: Use Cases. Three recipes you can run this week.
7:48Nova: One, and it is the fresh move: convert the renewed headroom. You have a month of 50 percent higher Claude Code limits. Take the one big job you have been putting off because it felt too expensive on tokens, a full test-suite generation, a legacy refactor, an eval harness, and run it this week. Capacity is loose right now; bank the durable output while it is.
8:12Dex: Two, the last-call Fable move, for Pro users specifically. Today is your last included day. Before the fork tomorrow, do two things: set that Workbench fallback to Sonnet 5 so nothing breaks, and run your single hardest recurring job on Fable one more time, then have it write down how it planned and structured the work as a Claude Code skill file. That markdown is portable, it runs on Sonnet 5 or Kimi K3 later, and it is judgment you keep after the model leaves your plan. That is the iwoszapar pattern, and today is the day it actually matters.
8:48Nova: Three, the speed split we just covered: take one interactive feature where users wait on tokens, point it at Cerebras Sol, and measure the drop in wait time. You are not buying more intelligence, you are buying the feel of instant, and for a live product that is often the upgrade users actually notice.
9:07Dex: Brought to you by SearchVis. Mainframe tells you what shipped. SearchVis makes sure the models know what you shipped. Here is the reality: when a buyer asks Claude or ChatGPT or Perplexity for the best tool in your category, the model names you or it names your competitor, and if it does not name you, you do not exist to that buyer. SearchVis tracks whether the answer engines cite your brand across every engine, tells you why, and hands you the exact move 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.
9:48Nova: The Deep Dive, and it is short. Read the capacity weather. This week the limits went up and stayed up, and the model that got rationed for six weeks got made permanent. That is not luck; that is Anthropic telling you, out loud, that compute loosened. Capacity is a cycle, it tightens and it loosens, and the builders who win are the ones who time their heavy work to it. When it is loose, like right now, you front-load the expensive jobs. When it tightens again, and it will, you are not scrambling, because you already banked the outputs and you already own the harness that swaps models underneath you.
10:28Dex: So the stance is: stop treating your rate limit like a fixed wall.
10:32Nova: Right. It moves, and the direction it moves is information. Here is what I would actually do: this week, run the one compute-heavy job you have been deferring, the one you keep saying you will get to. The window is open. It does not stay open, and the output outlasts the window.
10:51Nova: Level up this week: pick your one deferred heavy job and run it under the renewed Claude Code limits, and if you are on Pro, set your Fable fallback and distill one last skill before tomorrow's fork. Twenty minutes; the output outlives the model.
11:08Dex: And if you want every story, number, and link from today in your inbox, subscribe to the free Mainframe newsletter. It is the written brief, no charge.
11:17Nova: That is the show. Read the weather, spend the headroom, own the dial. See you tomorrow.