Latest briefing
July 19, 2026 · 5 stories (site) · 5 stories (base)
On July 19, 2026, the AI agent ecosystem consolidates a post-WAIC Shanghai week: Moonshot AI releases Kimi K3, the largest open-weight model ever distributed, already compatible with agentic coding tools; Thinking Machines (Mira Murati, John Schulman, Barret Zoph) ships Inkling under Apache 2.0, the first open multimodal model from a US lab built from the ground up to drive agents; OpenAI launches ChatGPT Work, a desktop agent that consolidates Operator, Deep Research and Workspace Agents into a single offer; Prime Intellect closes a billion-dollar Series A for its enterprise agentic training platform; and Anthropic responds with Sonnet 5 and Claude Cowork now available on web and mobile.
🔥 Top story
01
China releases the largest open-weight model in the world — already compatible with agentic coding tools
Behind every AI assistant sits a model, a giant file of numbers trained to read and write language. Moonshot AI, a Chinese start-up backed by Alibaba, has just put Kimi K3 online: 2,800 billion parameters, a context window of one million tokens (enough to swallow a large software project in one go), and out-of-the-box compatibility with the main agentic coding tools — Codex, Claude Code, Cline, RooCode — via an API in the OpenAI format. The full weights land on July 27. The model also understands images, so screenshots go to the agent with no pre-processing. For an agent builder, this is a serious candidate for an open-weight stack with controlled costs — the Chinese competition pushes this one-million-token frontier, where American models often lag. The practical catch: at 2,800 billion parameters, the electricity bill is not the same as for a small desktop model. Worth testing on real long-running tasks before flipping production over, but the signal is clear: China distributes capable open-weight models while American closed models get more expensive.
02
An open and free model that understands text, images and audio — signed by former OpenAI hands
Thinking Machines Lab — founded by Mira Murati, John Schulman and Barret Zoph, all ex-OpenAI — has just published Inkling under Apache 2.0, a genuine open-source licence with no commercial restriction. The model weighs 975 billion parameters but only activates 41 billion per pass (a mixture-of-experts design, where a small slice of neurons handles any given question). Its context window hits one million tokens, and it natively understands text, images and audio. The bet is explicit: Inkling is not the best on every benchmark, but it is open, modifiable, and shipped with documentation explaining how to adapt it to a profession. That gives a company the foundation to build its own tailored agent, without paying licence fees on the underlying model. The flip side: the compact version still demands around 600 GB of video memory — about half a dozen high-end graphics cards. It will not run on a regular SME server; it needs a cloud provider or a lighter variant. In short: openness is back as a selling point, but hosting remains a cost line worth measuring.
03
OpenAI launches a desktop agent that bundles everything together — pushing back against Claude and Copilot
An employee preparing a presentation or a summary brief spends most of their time juggling email, spreadsheets, internal documents and a browser. OpenAI announced on July 9 ChatGPT Work, an agent that pulls together that multi-application context and produces on demand documents, spreadsheets, presentations and even websites. Describe what you want, the agent pulls the right data from the authorised tools, and delivers a result ready to review. It runs on GPT-5.6, on web and mobile, for Pro, Enterprise and Edu subscriptions. The release folds three separate products shipped in recent months — Operator, Deep Research and Workspace Agents — into a single integrated offer. For a project manager: fewer tools to learn, and an agent that respects each user permissions (a salesperson only sees what they are entitled to). For the market, it is a direct answer to Anthropic Claude Cowork and Microsoft Copilot Agents — the desktop-agent battle is on. Pre-IPO nugget: Bank of America has already committed $520 million in credit to OpenAI to support the effort.
04
$130 million to train AI agents on a company core business tasks
Prime Intellect, a San Francisco start-up, announced on July 16 a $130 million Series A, led by Radical Ventures with Nvidia Ventures, Intel Capital, Dell Technologies Capital and Iconiq. The valuation hits one billion dollars. The platform delivers a full stack: compute to train a model, a reinforcement-learning tool (the method that teaches an agent by trial and error on real cases), and evaluation tools to measure whether the agent truly does the job it is given. The bet is simple: let a company become its own AI lab on its own business tasks, without depending on a generalist provider. A roll-call of marquee angel investors is on the cap table — Srinivas (Perplexity), Levie (Box), Weinberg (Harvey), Wang (Cognition), Foody (Mercor) — a clear sign of the network this play sits in. In short: worth watching closely, exactly the kind of platform to test on a monitoring or analysis workflow before writing the full spec.
05
Anthropic fights back: Sonnet 5 launched, Cowork now on web and mobile — for every day agentic workflows
Anthropic released on June 30 Sonnet 5, billed as the most agentic Sonnet yet — in plain language, a model shaped to drive chains of actions across external tools. Since July 7, Claude Cowork, the interface for those workflows, has moved from desktop to web and mobile, with a gradual rollout from the Max plan. A monthly recap feature entered beta on July 9. The positioning is clear: Anthropic goes for controlled costs on recurring agentic workflows, in direct competition with OpenAI ChatGPT Work and Microsoft Copilot Agents. For a user, the concrete difference shows over time: Sonnet 5 accepts longer tasks without losing the thread, Cowork opens up on everyday devices. In short, it is a signal to fold into the choice of agentic stack — Claude remains a serious option for workflows that must hold up over time without blowing the API budget. The catch: the subscription tier required to unlock the web and mobile version.
📡 To watch
China lays the groundwork for an alternative AI governance — 29 countries around Shanghai
The opening of WAIC 2026 in Shanghai on July 18 came with a striking geopolitical move: the creation of WAICO, an international AI governance body that brings together 29 countries. The read: China no longer waits for the United States to set the rules — it lays down its own, with its allies. To watch: who signs in the next 30 to 60 days, and what this bloc becomes.
DeepSeek V4 sets its new pricing — and pulls the plug on older models on July 24
DeepSeek announced for mid-July V4-Pro (1,600 billion parameters, 49 active) and V4-Flash (284 billion, 13 active), each with a one-million-token window and a peak/off-peak pricing grid that doubles the price at peak hours. The cache (memory that avoids reprocessing the same questions) cuts the cost by -80% on Flash and -92% on Pro. The old identifiers deepseek-chat and deepseek-reasoner will be deprecated on July 24 at 15:59 UTC — a mandatory migration for any agent still using them.
NVIDIA proposes a new unit of measurement — no longer the token, but success per dollar spent
In a manifesto published mid-July, NVIDIA argues that the central metric for agentic systems is no longer the cost per token but the success per dollar of compute spent, folding in the gains from post-training by reinforcement learning. The new Vera Rubin GPUs, combined with NeMo Gym and NeMo RL, are positioned as the infrastructure for that loop. In short: to fold into the FinOps thinking on quality-to-compute price ratio.
Bunkerhill Health raises for AI agents in healthcare — $55 million cumulative according to sources
Bunkerhill Health, a start-up offering agentic tools for American hospitals and health systems, announced mid-July a funding round led by Sequoia Capital with Felicis, Optum Ventures and Y Combinator. Valuation: $367 million. Caveat: the communicated $55 million may include cumulative funding (Seed + Series A + Series B); the exact round still needs confirmation. In short: a good example of a vertical agentic play (healthcare) with strong market potential, worth comparing with what France does on the same niche.
📊 Trend
On July 19, 2026, the AI agent ecosystem consolidates a week of structural shifts, marked by the release of Kimi K3 by Moonshot AI (the largest open-weight model ever published, already compatible with agentic coding tools), the arrival of Inkling at Thinking Machines (the first multimodal Apache 2.0 model from a US lab, built from the ground up to drive agents), the launch of ChatGPT Work by OpenAI (desktop agent that consolidates Operator, Deep Research and Workspace Agents), the billion-dollar Series A of Prime Intellect (full agentic training platform for enterprises), and Anthropic counter-punch with Sonnet 5 and Cowork on web and mobile. Four lessons emerge for those who build agents with the intent of generating revenue. (1) The desktop-agent battle is on — OpenAI (ChatGPT Work), Anthropic (Cowork) and Microsoft (Copilot Agents) all position on the same terrain, and the consolidation of existing products into a single offer is becoming the norm. (2) The open model is back as a commercial argument — Kimi K3 on the Chinese side, Inkling on the American side under Apache 2.0: two visions of openness, but a clear signal that lock-in is no longer the dominant position. (3) The one-million-token context window is becoming the new normal — Moonshot and Thinking Machines push it together, which changes the game for agents working on long projects. (4) Funding for enterprise agentic platforms is exploding — Prime Intellect at a billion, Bunkerhill in healthcare: verticals with strong market potential pull in ever bigger tickets.