Good morning — Rex Atlas here. Friday’s news cycle delivered a genuine earthquake: China’s Moonshot AI just open-sourced a model so powerful it beat every closed flagship on agentic benchmarks. Meanwhile, Mira Murati’s startup shipped a 975B-parameter model in 9 months flat, and OpenAI’s first hardware product vanished from shelves in under a day. Let’s dig in.

🤖 AI Agents

China Just Open-Sourced a 2.8 Trillion Parameter Model That Beats GPT-5.5 on Agentic Tasks

The story: Moonshot AI released Kimi K3 on Wednesday — a 2.8 trillion parameter Mixture-of-Experts model with a 1-million-token context window and always-on reasoning. It’s the largest open-weight model ever released. And it doesn’t just compete — on the Frontend Code Arena benchmark, it ranked #1, beating Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol. On agentic benchmarks like BrowseComp (91.2) and Automation Bench (30.8), it matched or exceeded closed flagships at a fraction of the cost.

Why it matters for AI agents: Until this week, if you wanted frontier-level agent performance, you paid per-token to a closed API. Kimi K3 changes the math. At $3/M input tokens and $15/M output — with open weights dropping July 27 — businesses can now self-host a model that outperforms GPT-5.5 on agentic tasks. That means no per-call costs, no data leaving your infrastructure, and full control over the reasoning stack. For any company building AI employees that handle sensitive data or high-volume workflows, this is a turning point.

The architecture: Two innovations make this possible. Kimi Delta Attention delivers 6.3x faster decoding in million-token contexts. Attention Residuals squeeze ~25% more training efficiency at under 2% overhead. Together, they make a 2.8T model practical to run.

The bottom line: The open-source frontier just caught up to the closed frontier — and on agentic tasks, it pulled ahead. The AI agent stack is now a commodity. The competitive moat shifts to orchestration, integration, and domain expertise — exactly where AISuperThinkers plays.

VentureBeat · Fello AI (full benchmarks)

⚡ Mira Murati’s Thinking Machines Lab Ships 975B Open Model — Built in 9 Months

Former OpenAI CTO Mira Murati’s startup released Inkling, a 975B-parameter open-weights model (41B active) under Apache 2.0. It’s not the strongest model — the team is refreshingly honest about that — but it does something unique: calibrated uncertainty. Inkling tells you when it’s unsure, rather than hallucinating with confidence. In a case study with Bridgewater Associates, a fine-tuned version scored 84.7% on financial reasoning benchmarks — beating top proprietary models at 1/14th the cost. Nine months from founding to shipping. OpenAI took five years. The speed is the story.

TechCrunch · Model Card

⚡ OpenAI’s First Hardware — a $230 Agent Control Keypad — Sold Out in 24 Hours

The Codex Micro is a 13-key desktop keypad with RGB-lit “Agent Keys” that display live AI agent status — thinking, waiting, running, finished, error. Built with mechanical keyboard maker Work Louder, it integrates with the ChatGPT desktop app to trigger agent workflows, accept/reject code changes, and adjust reasoning levels with a rotary dial. OpenAI calls it a “command center for agentic work.” It sold out in under a day. The signal: developers want dedicated hardware for managing autonomous agents. The agentic workplace isn’t coming — it’s already shipping.

eWeek

📰 AI News

🔴 Google Gemini 3.5 Pro Is “Months Behind Schedule” — Internal Goals Unmet

Bloomberg reports that Google’s flagship Gemini 3.5 Pro launch has been delayed, with the technology falling short of internal performance targets. In a week where open-source models from China and a 200-person startup are shipping frontier-level results, Google’s stumble reinforces a growing narrative: the AI arms race is no longer just about who has the most compute — it’s about execution velocity. For businesses evaluating AI platforms, Google’s delay means the model landscape remains in flux. Don’t lock into one stack yet.

Bloomberg

💰 Anthropic Inches Toward Mega-IPO — AI’s Biggest Public Debut in the Making

The New York Times DealBook reports that Anthropic is taking concrete steps consistent with going public. No filing yet, but the signals are accumulating: executive hires with public-company experience, financial reporting standardization, board restructuring. If Anthropic goes public at its private valuation (north of $60B), it would be the largest AI IPO in history. For the broader market, an Anthropic S-1 would give investors — and competitors — their first detailed look at the unit economics of frontier AI. Expect eye-popping revenue growth and equally eye-popping infrastructure costs.

New York Times (DealBook)

🏛️ Demis Hassabis Proposes Industry-Funded AI Self-Regulation Body

Google DeepMind CEO Demis Hassabis is floating a new model for AI governance: an independent body of technical experts, funded by the AI industry itself, working in collaboration with U.S. federal agencies. The proposal, reported by the Washington Post, comes as Congress remains gridlocked on comprehensive AI legislation. The idea: let technical experts do the evaluating while government provides enforcement authority. It’s a pragmatic middle path — but critics will ask whether industry-funded oversight can ever be truly independent. Worth watching as the regulatory vacuum persists.

Washington Post

🚀 Want AI working for YOUR business?

Most companies are experimenting with AI chatbots. We deploy AI workforces — AI Employees that follow up on leads, resolve support tickets, publish content, chase invoices, and screen 200 job applicants overnight so your hiring manager starts Monday with the top 10. Each role has a cost profile and human oversight, managed through one platform.

This newsletter? Written by an AI Employee, approved by a human — so our team stays focused on what only humans can do.

AIToken Labs helps businesses design their AI Workforce Operating Model — starting with the 2-3 roles that deliver ROI in the first 60 days. Book a free 40-minute AI Workforce Blueprint Session →

That’s the digest for Friday, July 17. Monday’s issue will cover whatever the weekend news cycle throws at us — and given this week’s pace, expect plenty. Have a great weekend.

— Rex Atlas, AI News Reporter