DeepSeek is prepping an IPO at a ~$71B valuation, up ~37% in two months; it got there on volume and price, not raw IQ. Kimi K2.6 already runs roughly 7x cheaper than Claude Opus 4.6 at current OpenRouter rates, and K3 (rumored today) targets long-horizon agent work. https://www.bloomberg.com/news/articles/2026-07-14/deepseek-mulls-new-funding-weeks-after-7-billion-round-ft-says
Read-through for your bill: when the ~4th-best model costs ~1/5 of the frontier, single-vendor lock is a margin leak. One rule: cheapest capable model by default, frontier only when the answer must be right. Frames both the Kimi and Fable stories today.
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. You own the code and the keys. When the model shelf reshuffles three times a week, you want someone steering the whole fleet. Book a 30-min call. https://outpace.media
The stories
LEAD (builder impact)
— Anthropic's Claude Code plugin marketplace is now a real, screened app store. The Anthropic-managed `claude-plugins-official` directory ships automatically when you start Claude Code; the `claude-plugins-community` marketplace hosts third-party plugins that passed Anthropic's automated validation + safety screening, each pinned to a specific commit SHA (so an install can't silently change under you). Install flow: `/plugin marketplace add <owner/repo>` then `/plugin install <name>@<marketplace>`. Add the screened community catalog with `/plugin marketplace add anthropics/claude-plugins-community`.
The move: stop copy-pasting SKILL.md from random repos; pull your next skill from the screened marketplace (commit-pinned, safety-checked), then run `/checkup` to catch context bloat.
Fable 5 leaves subscriptions Sunday, July 19 (11:59:59 PM PT)
Silver lining up front: it's NOT permanent (restore promised "as capacity allows"), and Sonnet 5 is the cheaper substitute you already have. The stake: after July 19, Fable 5 stops drawing from your plan's weekly limit and bills via usage credits at $10/M input, $50/M output (2x Opus 4.8, top of Anthropic's price list). Sonnet 5 (launched June 30, new Free/Pro default) offers near-Opus 4.8 agentic performance at $2/$10 per M.
The move: route bulk work to Sonnet 5; reserve Fable for long-context / multi-day reasoning; distill the method before Sunday (see Use Cases). CONTEXT: "Honeycomb EAP" (unreleased) briefly appeared in Cursor's menu July 8 — 1M context, extra-high effort, fallback to Opus 4.8 — fueling Opus 5 theories (Andrew Curran, kimmonismus). REPORTED/RUMORED, not confirmed. Don't re-architect around a nonexistent model string; watch anthropic.com/news.
Kimi K3 launch window is TODAY (RUMORED, not confirmed)
As of July 15, K3 is unreleased: not on Moonshot's public API, kimi.com still features K2.6, no model card/license/benchmarks. Hardest signal: a leaked Kimi Open Platform top-up promo tied to a July 15 launch (10–30% API recharge bonus through Aug 11). Leaks: ~2.5T-param MoE, 1M context, long-horizon agent focus. Carried by kimmonismus, Andrew Curran. VERDICT: timing strong, specs rumor — do not swap production on day one; be ready (routing dial, below).
The model is a swappable commodity; your harness is the asset.
Level up
Distill one Fable skill before Sunday night's cliff, and prune your context with the cross-agent quarantine tool (~20 min). Every run after is cheaper and outage-proof.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: Kimi K3 is landing today, a rumored two-and-a-half-trillion-parameter open-weight model built for hours-long agent runs, and if it lands where the leaks say, the model you plan with just got a cheap Chinese understudy. 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 Wednesday, July fifteenth, twenty twenty-six. I am Nova.
0:32Nova: On today's Mainframe: Kimi K three, the open-weight monster dropping today, and the exact move to make while everyone else is still reading the leak. Anthropic quietly turned its Claude Code plugin marketplace into a real, screened app store, so installing skills stops being a copy-paste trust exercise. And the one you have been dreading: Fable five, the model you plan with, leaves subscriptions Sunday night at midnight Pacific, so your best planning window closes in five days, and I will tell you the cheaper substitute up front. Two hands-on tools, three clever use cases. Let's get into it.
1:18Dex: Quick note before we roll: the voices you are hearing are AI. The reporting, the picks, and the opinions are put together by humans. Nobody at a lab is whispering in our ear.
1:33Nova: The Board, and keep your wallet in mind. The macro move that touches your costs today is the open-weight price war getting a new heavyweight. DeepSeek is prepping an IPO at a roughly seventy-one-billion-dollar valuation, up about thirty-seven percent in two months, and it got there on volume and price, not raw genius.
1:58Dex: And Kimi is right behind it. K two point six already runs roughly seven times cheaper than Claude Opus four point six at current rates, and K three is rumored bigger and aimed squarely at long-horizon agent work.
2:12Nova: Here is the read for your bill. When the fourth-best model on earth costs a fifth of the frontier, single-vendor lock is a margin leak. You do not need to become a chip analyst. You need one rule: cheapest capable model by default, frontier only when the answer must be right. That frames everything today, including the Fable cliff.
2:40Nova: Brought to you by Outpace. Here is the thing about a week like this one: models reshuffle three times before Friday, and if you are a business owner, you cannot chase every launch. Outpace is one senior operator, thirty years building real businesses (real estate, hospitality, food, even a theater), now aiming a whole team of AI agents at one client's project at a time, strategy to ship. Not a freelancer, not a faceless agency: a system. The agents do the work, the operator steers the fleet, and at the end you own the code and the keys. When the model shelf changes weekly, you want someone who swaps the engine underneath you and keeps you moving. Book a thirty-minute call at outpace dot media. That is outpace dot media.
3:39Dex: The Wire. Story one, and it is the builder one to act on. Anthropic's Claude Code plugin marketplace just became a real, curated app store.
3:50Nova: Right. For months, installing a skill or an MCP server meant trusting a random GitHub repo. What changed: there is now an official Anthropic-managed directory, `claude-plugins-official`, that ships automatically when you start Claude Code, plus a community marketplace, `claude-plugins-community`, where plugins have passed Anthropic's automated validation and safety screening, and each plugin is pinned to a specific commit SHA in the catalog.
4:23Dex: Pinned to a commit matters. It means the thing you install today cannot silently change under you tomorrow.
4:31Nova: Exactly. Here is precisely how you use it. The install flow is two steps: run slash plugin marketplace add owner slash repo, then slash plugin install name at marketplace. The official one is already there. To add the screened community catalog, run slash plugin marketplace add anthropics slash claude dash plugins dash community. Then browse and install. The concrete move this week: stop copy-pasting SKILL dot M D files from strangers into your dot claude folder. Pull your next skill from the screened marketplace instead, and get the commit-pinned, safety-checked version while most builders are still cloning raw repos. And run slash checkup after, to see if any of it is bloating your context.
5:29Dex: Story two, the builder-stakes one. Fable five.
5:35Nova: The silver lining first, because it matters: Fable is not leaving permanently, and there is a cheaper substitute you already have. Now the stake. Anthropic announced a third extension keeping Fable five free on Pro, Max, Team, and premium Enterprise plans through July nineteenth, twenty twenty-six at eleven fifty-nine PM Pacific. After July nineteenth, Fable five stops drawing from your plan's weekly limit and bills through usage credits at ten dollars per million input and fifty dollars per million output, twice Opus four point eight and the top of Anthropic's price list.
6:18Dex: So the free planning window closes in five days. What is the substitute?
6:23Nova: Sonnet five, launched June thirtieth as the new default for Free and Pro, offers near-Opus four point eight agentic performance at two dollars per million input and ten per million output. Route bulk work there. Reserve Fable for what only Fable does: long-context, multi-day reasoning. And the smartest move is not to burn the window on throwaway output, which we will get to in Use Cases.
6:54Dex: One more note: an unreleased model called Honeycomb briefly showed up in Cursor.
7:00Nova: Yes. Honeycomb EAP appeared in Cursor's model menu on July eighth before being pulled within hours, with a documented one-million-token context, extra-high-effort mode, and a fallback chain routing to Opus four point eight. Andrew Curran and kimmonismus are both carrying the theory that this is Opus five, possibly before month-end. Reported and rumored, not confirmed. Do not re-architect around a model string that does not publicly exist. Watch Anthropic's newsroom.
7:35Dex: And Kimi K three itself, our cold open. Where does it actually stand?
7:40Nova: Be precise here, because the internet is not. As of July fifteenth, Kimi K three is unreleased: it has not appeared on Moonshot's public API listings, kimi dot com still features K two point six, and there is no model card, license, or benchmark. The clearest hard signal is a limited-time top-up promotion tied to a July fifteenth launch, with bonus credits of ten to thirty percent on API recharges through August eleventh. The leaks describe a roughly two-and-a-half-trillion-parameter Mixture-of-Experts model with a one-million-token context, aimed at long-horizon coding and agents. kimmonismus and Andrew Curran are both carrying it. My verdict: treat the timing as strong, the specs as rumor, and do not swap production onto it on day one. The move is to be ready, which is the routing dial we keep preaching.
8:42Dex: Tool Lab. Two picks a builder can use today.
8:45Nova: Pick one is a hidden gem that is perfectly timed for the marketplace story and the Fable squeeze: a context-quarantine pruner surfaced through the awesome-claude-code-and-skills community list. What it does: it reads your real session transcripts to find skills, MCP servers, and agents that were loaded into context but never fired, then safely quarantines them. Why you want it: every skill description and MCP tool definition sits in your context every single turn whether it fires or not, and that is silent token spend. This tool supports Claude Code, Codex, Hermes, OpenCode, Cursor, and OpenClaw, has zero telemetry, ships as a single static Go binary, and its scan, report, prune workflow keeps removals reversible, installable via Homebrew and npm under MIT. First step: install it, run scan on your heaviest repo, read which skills never fire, and prune them. It changes your daily workflow because a leaner context is cheaper on every message, which stretches your Fable allowance right up to Sunday. Credit to the maker on the GetBindu awesome-claude-code-and-skills list.
10:12Dex: That pairs cleanly with slash checkup, which flags the same context tax but this tool goes cross-agent.
10:20Nova: Right. Pick two is trending, and it is a research skill: last30days by mvanhorn. What it is: an AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web, then synthesizes a grounded summary. Why now: it recently added arXiv for the papers behind the hype and Techmeme for the editorial tech-news layer, both free with zero keys, and first-run setup installs their command-line tools so they activate automatically. Install it via npx skills add mvanhorn slash last30days dash skill in Claude Code, or as a one-click MCP bundle in Claude Desktop. One reviewer said it replaced the manual Reddit-and-X searching, tab by tab, thread by thread, that took ninety minutes before every piece of content. The move: before you write your next post or pitch, run it on your topic and let it hand you real, recent conversations instead of stale blog posts. Credit: mvanhorn.
11:44Dex: Use Cases. Existing tools, clever moves.
11:48Nova: One, and this is THE move before Sunday: do not spend your last free Fable days generating output you throw away. Spend them distilling Fable's judgment into files you keep. Run your hardest recurring job on Fable, then have it write down HOW it planned and structured the work as a Claude Code skill file. Skills are portable markdown, so that same file runs on Sonnet five, on Codex, and on Kimi if K three lands. You are buying the expensive model's judgment once and keeping it cheap forever.
12:33Nova: Yes, she published it and it is worth stealing. Have Fable inventory every AI subscription, API bill, and per-seat tool you pay for, mapped against what you actually used in the last ninety days, then recommend consolidations, because most professionals pay three tools to do one job, and the judgment (cost per finished task, not cost per seat) is exactly where Fable leads. The inventory is trivial, the reasoning is hard, and the output saves you money every month after.
13:11Nova: Wire an open-weight route into your fallback chain now, before you need it. Set your Claude Code fallbackModel chain, add today's best Chinese open model as the high-volume route, keep Sonnet five as default, and reserve the frontier for hard reasoning. Then when K three actually ships with real benchmarks, you swap one line of config, not your whole workflow. Ready beats fast.
13:41Dex: And that readiness idea is where our next sponsor fits.
13:45Nova: Brought to you by SearchVis, also an Outpace Media product. Mainframe tells you what shipped. SearchVis makes sure the models know what YOU shipped. Because when a buyer asks ChatGPT, Claude, or Google's AI overview 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 thing 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.
14:38Nova: Here is the through-line nobody is saying out loud. Look at today: Fable rationed by capacity, Kimi K three rumored to undercut the whole Opus tier, a screened marketplace so you can trust installs, a pruner so your context stays lean. Every one of those is the same lesson from a different angle: the model is a rented, swappable commodity, and your only durable asset is the workflow you own around it. The routing rule. The distilled skill file. The lean, audited setup. The client relationship. None of that arrives in a model release, and none of it gets commoditized when the next model is cheaper next Tuesday.
15:24Dex: The builders who panic-chase every launch are optimizing the one thing they do not control.
15:31Nova: Exactly. So here is what I would actually do about it. Stop asking "which model is best this week." Ask "is my harness swappable." This week, wire one open-weight fallback route and distill one Fable skill before Sunday. Do that, and Kimi K three landing, Fable metering, Opus five leaking, none of it is a fire drill. It is a line of config.
16:01Nova: Level up this week: distill one Fable skill before Sunday night's cliff and prune your context with today's cross-agent quarantine tool. Twenty minutes, and every run after is cheaper and outage-proof. Subscribe to the free Mainframe newsletter, every story, number, and link in your inbox each morning. For Dex, I am Nova. Keep your harness swappable.