In partnership with

Welcome back!

An open-source model just beat Claude Fable 5 at frontend coding. Moonshot’s new Kimi K3 jumped from 18th to first place in Arena’s ranking, and its full weights are due this month. Let’s unpack…

Today’s Summary:

• 🚀 Kimi K3 tops frontend coding

• 🔥 Thinking Machines ships Inkling model

• 🗣️ OpenAI rethinks AI voice

• 🔊 OpenAI building a screenless AI device

• 🔬 DeepMind CEO asks for AI oversight body

• 🔗 Google Search AI connects apps

• 🛠️ 2 new tools

TOP STORY

Kimi K3 beats Claude Fable 5 in frontend code

The Summary: Moonshot released Kimi K3, a 2.8-trillion parameter open weight model with a million token context window and native vision. The model jumped from 18th to first place in Arena’s Frontend Code ranking, beating Claude Fable 5. Moonshot plans to open-source it by July 27.

Key details:

• Artificial Analysis gave K3 a global score of 57, close to Claude Fable 5 at 60 and GPT-5.6 Sol at 59

• In one live demo, a Kimi agent swarm spent three hours creating a browser-based version of macOS 27 UI, complete with a Liquid Glass look, native-style apps, and a working chess game.

• Available now through Kimi.com and the Kimi API. Moonshot says the full model weights will be open-sourced by July 27, 2026

• API pricing starts at $3 per million uncached input tokens and $15 per million output tokens, the same listed rate as Claude Sonnet

Why it matters: Chinese labs built their reputation on cheap models. K3's jump to Sonnet 5 pricing shows they now compete on the frontier. But the race has moved past cheap tokens; buyers judge each model by the cost of finished work. A model beating Fable 5 at frontend design shows there is no longer a gap between open and closed models.

FROM OUR PARTNERS

Launch Specialized AI Agents in Minutes

22 ChatGPT Agents Built for Every Marketing Job

Most marketers use ChatGPT to do general research and then call it an AI strategy. The ones outperforming them are deploying specialized agents built for specific jobs.

We put together 22 plug-and-play ChatGPT marketing agents that handle the work eating your week, each with built-in instructions and structured outputs ready to go in under 5 minutes.

Subscribe to Marketing Against the Grain and get all 22 free.

Inside you'll find:

• Competitive intelligence agent that visits competitor websites and builds detailed comparison matrices automatically

• Customer feedback analyzer that ranks improvement opportunities by business impact

• Social listening specialist that monitors brand mentions and flags reputation risks before they escalate

• Campaign optimization agents that handle attribution analysis and surface what is actually driving results

Your competitors are already running agents like these.

Get 22 ChatGPT Marketing Agents free when you subscribe to Marketing Against the Grain today.

Get The Guide

THINKING MACHINES

Mira Murati's lab ships its first model

The Summary: Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling, a 975-billion parameter open-weights model that reads text, images, and audio. It currently ranks as the strongest open-weights model from US labs.

Key details:

• Artificial Analysis gave Inkling a score of 41 on its Intelligence Index, ahead of OpenAI gpt-oss-120b at 24

• On GDPval-AA v2, Inkling scored an Elo of 1,238, beating DeepSeek v4 Flash max at 1,189

• On Design Arena, Inkling sits at 1,253, just below Claude Opus 4.6 at 1,257

• Available now on Hugging Face under the Apache 2.0 license and for fine-tuning on Tinker; API pricing starts at $1.87 per million input tokens and $4.68 per million output tokens with a 64K context window

Why it matters: Chinese labs have owned the open-source space for two years; this is the first US open-source release that seems truly competitive. Thinking Machines seems to be competing on fine-tuning: the self-fine-tuning demo shows how to give companies a model that can improve itself on their own data without an engineer in the loop.

Model card

FROM OUR PARTNERS

What Your AI Agent Is Missing

[Webinar] 8 levels of context maturity in AI-native engineering

You pay for every token your agent burns. When it returns code that doesn't fit your system, you prompt again. And again. This is because your agents are still missing the right context. Join live (FREE) on Jul 23 to see how leading teams use a context engine to fix it.

Register For Webinar

OPENAI

OpenAI launches a voice model that talks and listens at once

The Summary: OpenAI released GPT-Live, a voice model built to listen and speak at the same time instead of waiting its turn. It powers a new ChatGPT Voice that knows when to pause and can hand hard questions to a text LLM in the background. The model ships to iOS, Android, and the web.

Key details:

• The model decides whether to speak, listen, pause, interrupt, or call a tool multiple times per second, instead of once per turn

• GPT-Live-1 scores 84.2% on GPQA, a graduate-level science test, versus 45.3% for the old Advanced Voice Mode.

• On BrowseComp, a web search benchmark, Advanced Voice Mode gets the right answer 0.7% of the time, while GPT-Live-1 hits 75.2%

Why it matters: Old voice assistants were nearly useless at real tasks, no matter how smooth they sounded. This new model is a real front end for tasks, since it can call GPT-5.5 without breaking the conversation. Full-duplex audio forces AI to handle timing correctly in human conversations, which is a far harder problem than generating text.

QUICK NEWS

Quick news

• OpenAI is building a screenless speaker as its first AI device

• DeepMind CEO proposes an independent body to review AI models

• Google Search AI Mode can now connect to apps

TOOLS

🥇 New tools

• Claude Code Browser - Lets Claude work on websites

• Google Vids Updates - AI video editing and personal avatars