One-Line Summary

China’s Moonshot puts a near‑frontier, open‑weight model on the market while Google wires AI into everyday apps — a week to measure AI by value delivered, not tokens.

Big Tech

Google links AI Mode to your apps

Google’s conversational Search experience, AI Mode, now lets you securely connect select third‑party apps so you can take actions directly from Search instead of copying results between tabs. Google says you can link and interact with your go‑to services right in AI Mode. ¹

At launch, examples include adding grocery items to Instacart from a planning prompt, pulling Canva templates, and saving a playlist to YouTube Music — with the feature rolling out in the U.S. and more partners to come. This shifts AI Mode from answers to task completion. ²

Separately, Google is renaming NotebookLM to Gemini Notebook and starting to roll out a secure cloud computer in each notebook, enabling it to write and execute code for deeper analysis. It’s available today for Google AI Ultra users and certain Workspace business customers, with Pro users on the web to follow in the coming weeks; notebooks also sync with the Gemini app and are coming to AI Mode in Search. ³

Industry & Biz

Moonshot’s Kimi K3 puts open weights near the frontier

Moonshot AI, the Beijing startup behind the Kimi assistant, unveils Kimi K3 — a 2.8‑trillion‑parameter, 1‑million‑token‑context model that it positions on par with top U.S. labs’ systems except Anthropic’s Claude Fable 5 and OpenAI’s GPT‑5.6. Crucially, K3 is open‑weight, meaning customers can download and customize it. Bloomberg also notes Moonshot is pricing at roughly Anthropic Sonnet levels in China. ⁴

Early signals show strength on practical work: Arena.ai’s blind front‑end coding leaderboard ranked K3 ahead of Fable 5 and GPT‑5.6 Sol, and Moonshot says it will release K3’s weights on July 27 so third parties can verify claims. That combination — near‑frontier results and open weights — directly lowers switching costs for teams already building on U.S. tools. ⁵

Pricing undercuts premium U.S. models: Moonshot’s API lists $0.30 per million cached input tokens, $3 without cache, and $15 per million output tokens — regardless of context length. For comparison, SiliconANGLE notes Fable 5 is $1 per million input and $50 per million output, while GPT‑5.6 Sol is $0.50 input and $30 output. That gap invites teams to move heavier workloads to K3 where quality holds. ⁶

Why it matters for budgets: OpenAI is telling CFOs to track “useful intelligence per dollar” — the full cost per successful task — not just per‑token price. In its July 17 scorecard, OpenAI says GPT‑5.6 Sol set a new state of the art on a coding‑agent index while using 54% fewer output tokens than another leading model, and on DeepSWE v1.1 it reached 72.7% versus Fable 5’s 69.9% at 36.2% lower estimated API cost. If K3 can deliver comparable outcomes at lower prices, the value equation shifts further. ⁷

K3’s debut jolts markets and raises policy stakes

CNBC reports K3 still trails Fable 5 and GPT‑5.6 Sol on overall performance by Moonshot’s own accounting, but it surpasses GPT‑5.5 and Opus 4.8 on some coding and agent benchmarks, with Bank of America highlighting gains despite China’s compute constraints. Moonshot’s backers include Alibaba and Tencent, and the company has rapidly climbed China’s model ranks since 2023. ⁸

The reveal coincides with a selloff in Chinese model stocks, with Z.ai down 28% and MiniMax down 16% in Hong Kong trading, and comes as U.S. lawmakers weigh curbs on domestic use of Chinese AI models. For enterprise buyers, the short‑term takeaway is competitive pressure on price; regulatory review remains a gating factor for deployments handling sensitive data. ⁸

New Tools

Capital One open-sources VulnHunter to find exploitable code paths

Capital One released VulnHunter, an open‑source, agentic AI tool that scans source code starting from real attacker entry points (APIs, messages, file uploads), reasons forward to see if an exploit survives defenses, and then tries to disprove its own finding with a built‑in “falsification engine.” The project, Apache 2.0‑licensed and currently using Anthropic’s Claude Opus 4.8 in a Claude Code environment, ships proposed fixes alongside evidence so developers spend less time triaging false positives. ⁹

For a highly regulated bank still marked by a 2019 breach and an $80 million fine, shipping a public tool that emphasizes exploitability and false‑positive reduction signals a push to make AI security practical at scale — and gives engineering teams something concrete to pilot on non‑critical repos. ⁹

Community Pulse

Hacker News (55↑) — Cautious optimism about VulnHunter’s promise, with calls to judge Capital One’s specific agentic tools and concerns about false confidence and transparency. ¹⁰

"True, but this post in about Capital One. It doesn’t take a genius or system architect to figure out AI use in banking both internally and for consumers is a huge benefit especially for capital one. Maybe we should critique the actual agentic products they are creating rather than critique the entire AI spend." — Hacker News ¹⁰

What This Means for You

As budgets tighten, measure AI by outcomes per dollar. Borrow OpenAI’s framing — cost per successful task and how often outputs are “ready to use” — to compare your current stack with K3 or any alternative. This keeps pilots honest when a cheaper model requires more retries, or a pricier one finishes in one pass with fewer edits. ⁷

For marketing, ops, and e‑commerce teams, Google’s AI Mode integrations reduce tab‑hopping. Connect Instacart to move lists into carts or pull Canva templates from a Search conversation, and decide if this shortens planning cycles or handoffs with design. ¹

If you own a product or engineering backlog, consider a limited security exercise with VulnHunter on a non‑sensitive repo. The “attacker‑first” pathing and falsification step are designed to cut noise; compare the findings with your current scanner’s alerts before expanding scope. ⁹

Leaders should watch two signals before any major vendor shift toward K3: the July 27 weight release and independent benchmark replications that mirror your workflows (e.g., repo‑scale coding or web UI tasks). Those will clarify whether K3’s early lead on select tests is durable for your use case. ⁵

Action Items

1. Link one app to Google’s AI Mode: In the U.S., connect Instacart or Canva and complete a small task end‑to‑end (e.g., party list to cart, flyer template to draft) to gauge real time saved.
2. Build a one‑page AI value scorecard: Pick one workflow, define “done,” and track total cost per successful task and review time this week to baseline your current model.
3. Pilot VulnHunter on a safe repo: Ask your engineering or security team to run it on a non‑critical codebase and compare exploitability‑backed findings with your existing scanner’s alerts.
4. Test Kimi K3 for a single task: If accessible in your region, run a front‑end mock or long‑document analysis and compare output quality, retries, and cost against your current model using the same rubric.