Kimi K3 officially launched (evening of July 16): 1M-token context, standout coding / 3D / knowledge performance. https://finance.biggo.com/news/1463f539-2df9-479f-9d17-d6c3a1990722
2.8T total parameters, which Moonshot says makes it the largest open-source model to date (ahead of DeepSeek V4-Pro at 1.6T). Pricing: $0.30/M cached input, $3/M cache-miss input, $15/M output, full 1,048,576-token context at that rate. https://officechai.com/ai/kimi-k3-2-8-trillion-parameters-pricing-context-window/
Moonshot claims overall intelligence second only to Fable 5 and GPT-5.6 Sol; independent coding benchmarks still awaited (community testers put it ~Opus 4.7+ on some evals, behind Fable/Sol on Terminal-Bench 2.1).
Read-through for your bill: 4th-best model on earth at a fraction of frontier output price, 1M context included. Cheapest-capable-by-default; don't single-vendor lock. Frames the 1Password and Fable stories below.
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 shelf reshuffles three times before lunch, you want someone steering the whole fleet. Book a 30-min call. https://outpace.media
The stories
LEAD (fresh Claude-adjacent release), 1Password for Claude (July 16)
Browser integration lets Claude use your stored logins to complete online tasks without credentials ever reaching the model. Credentials injected into the target site via a secure channel 1Password controls; password + MFA one-time codes stay outside the model, its memory, and Anthropic's systems. Brokers access across multiple sites within one task (no re-login mid-workflow); scans each page after every autofill and wipes filled values if a submission fails. Access: 1Password users on Mac (business/family/individual); requires 1Password desktop app + browser extensions AND Claude desktop app + browser extensions.
The move: wire it into Claude in Chrome, scope to approved items, run one real logged-in workflow end-to-end (invoice pull, dashboard export, booking).
Kimi K3 is live (web + API), K3 Max + K3 Swarm Max options
Moonshot claims overall intelligence 2nd only to Fable 5 / GPT-5.6 Sol; split coding verdict (community: ~Opus 4.7+ on some evals; Fable 5 + Sol still lead Terminal-Bench 2.1). Carried by kimmonismus, Andrew Curran, TestingCatalog; Artificial Analysis confirms a strong independent score (57 on Intelligence Index). VERDICT: real open-shelf moment; don't swap production day one, wait for standard coding benchmarks.
after a data scandal. The CLI was found uploading entire directories (SSH keys, password DBs, docs) to xAI cloud buckets; xAI disabled it and released 844,530 lines of Rust to regain trust. CATCH: upload code remains in the binary, gated only by a server-side flag xAI can re-enable without an update.
The move: if you ran Grok Build before July 13 in a Git repo, treat every tracked credential (API keys, DB passwords, cloud tokens, SSH keys) as potentially transmitted, rotate now.
Silver lining up front: NOT permanent (restore promised "as capacity allows"); Sonnet 5 ($2/$10) and now Kimi K3 are cheaper substitutes. After July 19, Fable bills via usage credits at $10/M in, $50/M out (2x Opus 4.8). Delta today: three fresh model launches give you routing targets.
The move: distill the method before Sunday (see Use Cases).
The written brief from each episode, free, every morning.
Transcript
0:00Nova: One Password just shipped a Claude integration that lets your agent log into any site and do the task, and the password never touches the model. That is the missing piece for anyone building browser agents, and it landed this morning. 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 Thursday, July sixteenth, twenty twenty-six. I am Nova.
0:32Nova: On today's Mainframe: One Password for Claude, the credential broker that finally makes your logged-in browser agents safe to run. Kimi K three is officially out, a two point eight trillion parameter open model that Moonshot says is second only to Fable and Sol, and it is on the web and the API right now. Grok Build got open-sourced, eight hundred forty-four thousand lines of Rust, and there is a builder move buried in why they did it. And a builder-stakes clock: Fable five leaves subscriptions Sunday, and you have three days left to bank the method. Let's get into it.
1:17Dex: Quick honesty note before the news: the voices you are hearing are AI. The reporting, the picks, and the opinions are human-made. On to the Board.
1:28Nova: The macro move that hits your bill today is the open shelf just got a new heavyweight. Moonshot officially launched Kimi K three on the evening of July sixteenth, featuring a one million token context window with standout performance in coding, 3D generation, and knowledge tasks. K three comes in at two point eight trillion total parameters, which Moonshot says makes it the largest open-source model released to date, well ahead of DeepSeek's V four Pro at one point six trillion. And the price is the story: Kimi K three costs thirty cents per million tokens for cached input, three dollars per million for input that misses the cache, and fifteen dollars per million for output, with the full one million token context at that rate.
2:22Dex: What does that mean for my margin, plainly?
2:25Nova: The fourth-best model in the world now runs at a fraction of frontier output pricing, with a million tokens of context included. So the same rule as every week: cheapest capable model by default, frontier only when the answer must be right. Do not single-vendor lock. That frames both the One Password story and the Fable clock today.
2:53Nova: Brought to you by Outpace. Here is the thing about weeks like this one: three model launches, a security scare, a subscription cliff, all before lunch. Most people freeze. The operator behind Outpace does not. Thirty years building real businesses, real estate, hospitality, food, even a theater, and now he points a whole team of AI agents at one client's project at a time, strategy all the way to ship. Outpace, at outpace dot media, is a system, not a freelancer and not a faceless agency: the agents do the work, one senior operator steers them, and at the end you own the code and you own the keys. When the shelf reshuffles this hard, you want a fleet, not a lone hire. Book a thirty-minute call at outpace dot media. That is outpace dot media.
3:56Nova: The Wire. Lead story, and it is the freshest Claude-adjacent builder release of the day. One Password today launched One Password for Claude, a browser integration that lets Claude use a person's stored logins to complete online tasks without those credentials ever reaching the AI model. Dex, this solves a real problem.
4:22Nova: To let an agent sign in somewhere, you generally had to paste credentials straight into the agent's context, where the model can read them. The other option was stopping to authenticate by hand at every single step. Both are bad. One is insecure, the other kills the automation. Here is the fix: One Password drops that tradeoff. Credentials are injected into the target site through a secure channel One Password controls, and the password and any multifactor one-time code stay outside the model, its memory, and Anthropic's systems.
5:02Dex: And it does not ask me to log in over and over inside one task?
5:07Nova: Right. One Password brokers credential access across multiple sites inside one task, so Claude can move through multistep workflows without asking for logins again. And a genuinely thoughtful safety touch: it scans each page after every autofill and wipes any filled values if a form submission fails, cutting off a common way for secrets to linger on screen. The catch on access: it is available now to One Password users on Mac across business, family, and individual plans, and it requires the One Password desktop app and browser extensions plus the Claude desktop app and browser extensions. The move this week: if you have been faking logged-in agent work with pasted cookies or hand-authenticating, wire this into Claude in Chrome, scope it to a set of approved items, and let Claude run one real logged-in workflow end to end, an invoice pull, a dashboard export, a booking.
6:12Dex: Story two, Kimi K three, the full picture beyond the Board.
6:18Nova: The coding verdict is split: community testers claim K three reaches Opus four point seven plus level on some evals but note Claude Fable five and GPT five point six Sol still dominate Terminal Bench two point one. Testing Catalog carried it first: K three is on the web and the APIs now, with K three Max and K three Swarm Max options. On benchmarks, Moonshot says K three's overall intelligence ranks second only to Claude Fable five and GPT five point six Sol. Carried hard by kimmonismus, Andrew Curran, and Testing Catalog, with Artificial Analysis confirming a strong independent score. My read: this is a real DeepSeek-style moment for the open shelf, but do not swap production on day one. Wait for the standard coding benchmarks.
7:17Dex: Story three, Grok Build went open source. There is a sharp lesson here.
7:24Nova: There is, and it is a cautionary one. The Grok CLI faced severe backlash when it became apparent that running the command in a directory could upload that entire directory to xAI's cloud buckets. One user reported running it in their home directory and seeing it upload their SSH keys, password manager database, documents, photos, everything. A few hours later they released the entire Grok Build codebase under an Apache two point zero license, presumably to try to regain trust. The builder detail: xAI released eight hundred forty-four thousand lines of Rust, but security researchers confirmed the code capable of exfiltrating entire codebases remains present, held off only by a server-side flag xAI can re-enable without pushing an update.
8:22Nova: Any developer who ran Grok Build before July thirteenth in a directory with a Git repo should treat every credential in their tracked files and commit history as potentially transmitted: API keys, database passwords, cloud tokens, SSH keys. Rotate them today. That is the move. The upside for the rest of us: an eight hundred forty-four thousand line agent harness is now readable. For comparison, OpenAI's Codex repository is roughly nine hundred fifty thousand lines of Rust, which surprised people who assumed terminal agents were thin wrappers.
9:07Nova: Fable five leaves subscriptions this Sunday, July nineteenth. We have covered the date; what is new today is the pressure: three model launches just gave you cheaper substitutes to route to. So the delta this week is not "panic," it is "distill." More on that in Use Cases.
9:26Nova: Tool Lab. Two picks, both hands-on today. First, tagged HIDDEN GEM: the open-sourced Grok Build repo itself, as a reading tool. You can read the source to see exactly how it works, from context assembly to tool-call dispatch, and if you are working with skills, plugins, hooks, MCP servers, or subagents, the source is the definitive reference for how each is loaded and invoked. Credit xAI and Simon Willison, who did the first teardown. Concrete first step: clone the repo, open xai grok tools before anything else, and read how it dispatches a shell command. Every rented agent you use is this same loop. Reading one real one end to end demystifies the whole category. And Grok Build can now run fully local-first: compile it yourself, point it at your own local inference, and drive everything from your config dot toml.
10:34Nova: Tagged TRENDING: Inkling from Thinking Machines, Mira Murati's lab. It is their first major model under an Apache two point zero open source license, and it posts strong open-weights performance on software engineering, seventy-seven point six percent on SWE-bench Verified, beating Nvidia's Nemotron three. The reason it is a builder pick and not just a headline is one feature: sweeping its effort setting, Inkling matches Nemotron three Ultra on Terminal Bench two point one at roughly a third of the tokens. That is a controllable thinking-effort dial. The move: it is on Databricks through the Unity AI Gateway, and you can connect it to coding agents such as Cursor and OpenCode. For high-volume mechanical work, turn the effort down, pay a third of the tokens, keep the quality. Credit John Schulman and the Thinking Machines team.
11:37Nova: Use Cases. Things already in your hands, used cleverly this week. One, and it is time-sensitive: distill Fable before Sunday. Run your single hardest recurring job on Fable, then have it write down how it planned and structured the work as a Claude Code skill file. Markdown is portable, so that same file runs on Sonnet five, on Kimi K three, on Codex. You are buying the expensive model's judgment once and keeping it after the model leaves your subscription. Credit the iwoszapar pattern we have tracked all week.
12:18Nova: Two, the credential-agent recipe, courtesy of today's One Password launch. Pick one workflow you have been doing by hand because it needs a login: pulling a monthly report from an analytics dashboard, exporting invoices, checking an order system. Wire One Password for Claude, scope it to just those approved items, and schedule it. The secret never enters the model, and the page gets scrubbed if a submission fails. Most builders are still pasting cookies. This is the safe version, and it shipped today.
12:54Nova: Three, exploit Inkling's effort dial as your cheap route. Point Cursor or OpenCode at Inkling on Databricks, set the thinking effort low, and route boilerplate, scaffolds, and mechanical refactors there. Reserve Claude for the judgment calls. You are getting Terminal-Bench-level quality on the easy stuff at a third of the token cost, which is the whole cheap-by-default discipline in one config change.
13:23Nova: Brought to you by SearchVis, an Outpace Media product. Here is the uncomfortable version of today's news: three models launched, and buyers are asking Claude, ChatGPT, and Perplexity which tool to use, right now. If the answer engine names your competitor and not you, you do not exist to that buyer. SearchVis tracks whether the AI answers name your brand across every engine, tells you why, and hands you the exact move to publish to win the citation. Mainframe tells you what shipped; SearchVis makes sure the models know what you shipped. Check where you stand and start free at searchvis dot outpace dot media. That is searchvis dot outpace dot media. Be the answer.
14:17Nova: The Deep Dive. One idea to carry into your week. Notice the shape of today: One Password inserts itself between the agent and your passwords. Grok Build got caught vacuuming up whole repos and had to open-source itself to win trust back. Anthropic published fresh research on agentic misalignment, finding four more ways autonomous agents misbehave in simulations. All three point at the same truth: the agent era's real bottleneck is not intelligence, it is trust. The models are already smart enough to act on your behalf. The question buyers and clients will ask is not "is it capable," it is "what can it touch, and what happens when it is wrong."
15:11Nova: The edge is the agent you can prove is contained. The builder who wires a credential broker so the password never leaks, who scopes permissions to approved items, who reads the harness before letting it run shell commands in a real repo, that builder ships things a client will actually put into production. Capability is now table stakes. Containment is the moat. So here is what I would actually do about it: this week, take your one most useful agent workflow and add exactly one containment layer, scoped credentials, a read-only default, a human approval gate. Not because it is safe theater, but because "I can prove what my agent can and cannot do" is about to be the thing you sell.
16:02Nova: The Close. Level up this week: wire One Password for Claude into one logged-in workflow and scope it to approved items only, about ten minutes, and it turns your browser agent from a demo into something you trust. And before Sunday, distill one Fable skill so its judgment outlives your subscription. If you want every story, number, and link from today in your inbox, subscribe to the free Mainframe newsletter. It is the written brief, and it is free. That is the show. I am Nova.