🇨🇳 China just caught up to Claude

Kimi K3 beat Opus 4.8 at 40% less cost. Plus: the MLB considers AI in dugouts cheating and the White House controls AI access.

Author

Anthony Chebat
July 18, 2026

Sponsored by

My fellow AI explorers

Two things happened this week that quietly moved the AI industry's center of gravity, and neither of them was a shiny new product launch.

In today’s edition:

The best voice models, now across all channels

Most CX platforms do not own the voice. They orchestrate a workflow, then call a third party for speech and transcription. Every hop adds latency, cost, and another vendor to manage.

ElevenAgents is the opposite. They make the voice models the market builds on, and ElevenAgents puts full orchestration on top. Voice, transcription, text-based chat, and reasoning run in one vertically integrated pipeline, so responses come back in <400 milliseconds and sound human, not synthetic.

Plus, you keep full control. Plug in any LLM, integrate tools, webhooks, and MCP servers, and ground responses in your knowledge base. Get an agent live in minutes, then A/B test with Experiments, enforce Guardrails, and version every change.

The payoff: more human conversations, lower latency, and far less time stitching infrastructure together. You build on the models you already trust. Pricing is transparent and flat at $0.08 per minute.

Chinese AI

Kimi Just Ate a Slice of Anthropic's Lunch

Moonshot AI, a Beijing-based lab backed by Alibaba, released Kimi K3 this week, and it landed like a small earthquake in Silicon Valley.

  • The benchmarks: Kimi K3 beat Anthropic's Opus 4.8 and OpenAI's GPT-5.5 in front-end coding tests, and only trailed the very top tier, Claude Fable 5 and GPT-5.6 Sol, on overall capability.

  • The price: Moonshot says it costs roughly 40% less to run than Opus 4.8.

  • The catch: Unlike the premium US labs, Moonshot is releasing Kimi as a fully open-weight model on July 27. It means that anyone, including foreign governments, can download it and run it on their own infrastructure.

  • The pattern: Kimi K3 lands just weeks after z.AI's GLM-5.2 and Meituan's LongCat 2.0, the latter reportedly trained entirely on domestic Chinese chips.

Here's the asterisk nobody's saying out loud: this release came days after US export controls briefly cut off access to Anthropic's own Fable and Mythos models. While Washington was busy deciding who's allowed to touch the best American AI, China was busy giving its best AI away for free. One AI startup founder told CNBC he'd already switched 100% of his company's traffic to a cheaper Chinese model and watched his costs "crash to the ground."

Cost is becoming the story just as much as capability. If your product roadmap assumes American labs will always sit untouchable at the top, this is the week to double check that assumption.

🔮 Prediction: Expect at least one major US enterprise AI buyer to publicly announce a shift to a Chinese open-weight model in the next 60 days, purely on cost grounds. The "China is behind" narrative doesn't survive 2026 in its current form.

Reply and tell me: would you ever run your company's AI stack on an open-weight Chinese model? I want to know where the line is for you.

Politics

The White House Just Became AI's New Bouncer

For the past few years, Anthropic and OpenAI have decided who gets early access to their most powerful models. That's changing, and fast, according to a new CNBC report.

  • What's new: The Trump administration is asserting more control over which companies and agencies get access to frontier models, a decision that used to sit entirely with the labs themselves.

  • The mechanism: Anthropic ran its own vetting through Project Glasswing for its Mythos cybersecurity model. OpenAI had a similar setup called Daybreak. The White House's new "Gold Eagle" clearinghouse could effectively sit on top of both.

  • The denial: A White House official told CNBC it doesn't "approve" AI releases, and that engagement with labs is "voluntary," pointing to Trump's recent executive order as the framework.

  • The receipts: Anthropic's Fable 5 and Mythos 5 were briefly suspended last month over export control concerns before access was restored. OpenAI was separately asked to gate GPT-5.6 to "trusted partners" before its public release.

The asterisk here is the gap between what's being said and what's happening. Officially, nothing has changed; releases are still "entirely up to the companies." Unofficially, two of the most powerful private companies on Earth just had their release schedules interrupted by a government directive, twice. And this happens in the same month a rival Chinese lab released its best model completely free and open.

If you're building a product roadmap around a frontier model, you now have a third party in the room who isn't your vendor and isn't your customer.

🔮 Prediction: By year-end, expect a formal, named review process (not a "voluntary" one) for frontier model releases, and expect at least one lab to push back publicly on the timeline it creates.

Hit reply and tell me if you think government gatekeeping on AI access makes the US safer or just slower.

30-Second AI Play

Turn Any Topic Into a Vox-Style Mini Documentary

Ever wanted your own explainer video, the kind with the punchy narration, the dramatic zoom, the text stamps, but without touching a timeline or a video editor? One creator just walked through the exact workflow, and it's simpler than it looks.

  1. Build a style guide. Feed Claude a batch of Vox-style videos (NotebookLM works well here) and ask it to produce a written breakdown of the editing style, plus a reusable "style prompt" and "animation prompt."

  2. Write the script first. Ask Claude to draft a 30-second script broken into three 10-second chunks. Review it, tighten it, and make sure the opening line has a hook.

  3. Generate the first image. Use an image model (GPT Image 2 works well) to create the opening frame, no text or overlays yet, since those get added separately.

  4. Chain your scenes. Have Claude pull the last frame of each generated clip, analyze it, and use that to plan the prompt for the next scene, so the video flows instead of jump-cutting.

  5. Stitch and ship. Once all clips are generated, stitch them together with ffmpeg for one finished video, all for around a dollar in generation costs.

💡 Pro tip: Don't skip the style guide step. It's the difference between a video that looks AI-generated and one that actually looks like it belongs on a real channel.

Advertise to 180k engineers and CTOs choosing what tools their companies build with.

Other Relevant AI News!

📉 AI stocks got hit hard this week as chipmakers slid and investors questioned whether the AI rally has outrun its fundamentals, all while oil climbed on the expanding Iran conflict.

💡 A new light-powered chip out of Monash University can generate, steer, and read information using light instead of electricity, a potential step toward AI computing that needs far less power than today's data centers.

🎬 Netflix confirmed it paid $587 million in cash for Ben Affleck's stealth AI startup InterPositive, whose tools have already touched roughly 300 Netflix titles in post-production this year.

⚾ MLB is cutting off dugout iPads from custom AI tools after teams reportedly used them for pitch calling and substitution recommendations, a job that used to belong to managers.

Golden Nuggets

  • 🇨🇳 Kimi K3 just proved cost, not just capability, decides who wins the AI race, and it's giving China a real opening.

  • 🏛️ The White House is no longer just watching frontier AI releases. It's steering them.

  • 🎬 A Vox-style AI documentary now costs about a dollar and one afternoon to make.

Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)

Until our next AI rendezvous,

Anthony | Founder of Uncover AI