AI Bulletin · week of July 13

This week’s bulletin gathers six news blocks on AI applied to work, from 13 to 17 July 2026, plus a closing framework on the state of open source. The week was marked by Apple’s lawsuit against OpenAI, Anthropic’s IPO preparations, and a run of large open models.

1. Apple sues OpenAI and Greg Brockman takes charge

Apple accused OpenAI and former Apple executives of stealing confidential hardware information to support OpenAI’s device ambitions. The lawsuit alleges that senior leaders directed employees to bypass security procedures and transfer sensitive details of products still in development.

In parallel, CNBC reported that Greg Brockman now oversees OpenAI’s key projects after Fidji Simo stepped down due to illness. Brockman is tasked with driving revenue against competition from Anthropic and Google, with ChatGPT’s market share declining and a prospective IPO on the horizon.

The reorganisation also reaches the safety area: according to Engadget, Johannes Heidecke is leaving the head of safety role, and Saachi Jain, who has led OpenAI’s safety teams before, takes over on an interim basis.

Sources: TechCrunch, 10 July — “Apple sues OpenAI over alleged trade secret theft”; CNBC, 10 July — “OpenAI power consolidates under co-founder Greg Brockman ahead of prospective IPO”; Engadget via TLDR AI, 13 July — “OpenAI’s head of safety is reportedly leaving as part of company reorganization”.

The bulletin mascot, as a court reporter, takes notes in its notebook in front of a wooden judge's bench with a large gavel — Apple sued OpenAI over alleged trade secret theft, while Greg Brockman moves to oversee the company's key projects.

2. Money: Anthropic’s IPO, DeepSeek and compute as a market

Anthropic is planning investor meetings ahead of a potential IPO later this year, with Goldman Sachs, Morgan Stanley and JPMorgan Chase leading the effort. The company recently closed a $65 billion funding round at a $965 billion valuation, surpassing OpenAI’s $852 billion. DeepSeek, for its part, reportedly explored raising $1.5 billion at a $71 billion valuation before a possible IPO in late 2026 or 2027.

On the infrastructure side, Kalshi is building prediction markets for the price of compute: a forward curve for GPU rental prices, based on weekly and monthly contracts, reaching as far as a year out. And Elon Musk bought APR Energy, a Jacksonville-based gas turbine company with more than 1 GW of mobile capacity, for $1 billion, to power Grok.

Sources: CNBC, 15 July — “Anthropic moves closer to mega-IPO as bankers line up investor meetings”; TechCrunch, 14 July — “DeepSeek reportedly in talks to raise $1.5B, then IPO”; report via TLDR AI, 15 July — “Kalshi ramps up effort to build markets for AI computing power”; Electrek, 14 July — “Elon Musk quietly buys a $1 billion gas turbine company to power Grok”.

3. Models: Inkling, Kimi K3 and the Gemini 3.5 Pro delay

Thinking Machines released Inkling, its first open model: a 975-billion-parameter mixture-of-experts with 41 billion active parameters, multimodal reasoning and a one-million-token context window. It is available for customisation through Tinker, alongside a smaller preview version.

Moonshot launched Kimi K3, a 2.8-trillion-parameter multimodal model with a one-million-token context window and optimisations for faster long-context decoding and agentic coding. It is available through Kimi’s products and API, with open weights scheduled for release on 27 July.

Google, meanwhile, reportedly delayed Gemini 3.5 Pro while working to improve its performance, particularly for coding. Alphabet shares fell 4% after the report; the company says the model remains in partner testing alongside an upgraded Flash model.

Sources: Thinking Machines, 16 July — “Introducing Inkling”; Moonshot, 17 July — “Kimi K3”; CNBC via TLDR AI, 17 July — “Gemini 3.5 Pro reportedly faced delays”.

The bulletin mascot, wearing museum curator gloves, examines three glowing spheres of different sizes displayed on wooden pedestals — A week of models: Thinking Machines' Inkling, Moonshot's Kimi K3, and the Gemini 3.5 Pro delay.

4. AI in your pocket: Gemma 4 for the Pixel 10, Bonsai 27B and Jetson Thor

Google announced Gemma 4 E2B, designed to run natively on the Pixel 10’s TPU. The model supports instant, deep offline conversation, image identification and on-device audio transcription, and lets users command core phone functions, like Wi-Fi or maps, using just private voice or text.

PrismML introduced Bonsai 27B, which it describes as the first 27B-class model to run on a phone. It achieves this through 1-bit and ternary weights that shrink the model to 3.9 GB and 5.9 GB depending on the variant. The company says it sustains complex tasks like multi-step reasoning and tool use within the limits of a modern smartphone.

At the edge, NVIDIA introduced the Blackwell-powered Jetson Thor T3000 and T2000 modules for robotics, visual AI and agent workloads on compact systems. The platform has prior adoption from Amazon Robotics, Boston Dynamics, FANUC and 1X.

Sources: 9to5Google, 14 July — “Google announces Gemma 4 optimized for the Pixel 10’s TPU”; PrismML, 15 July — “Announcing Bonsai 27B: the first 27B-class model to run on a phone”; NVIDIA, 16 July — “NVIDIA expanded Jetson Thor for edge robotics”.

5. Agent security and privacy: Grok Build, GPT-Red and sandboxes

An analysis published as a GitHub gist examined exactly what Grok Build, xAI’s coding CLI, sends home: it transmits the contents of the files it reads verbatim and unredacted, and uploads whole repositories independent of what the agent reads, with a Google Cloud Storage bucket as the destination. The analysis notes there is no proof that xAI trains on the data. Two days later, xAI published the Grok Build repository on GitHub: a terminal agent that inspects codebases, edits files and runs shell commands.

OpenAI trained GPT-Red, a model that iteratively generates adversarial prompts to expose vulnerabilities at scale. According to the company, incorporating its attacks into training reduced GPT-5.6 Sol’s failures on a difficult prompt-injection benchmark by sixfold.

Perplexity introduced SPACE, a platform of ephemeral sandboxes for agents handling sensitive tasks, with credential isolation, rolling snapshots, and on-prem and offline operation. And Google published Mantis Skills, a set of security-review skills for coding agents, framed as an adaptable foundation rather than a rigid set of instructions.

Sources: GitHub gist via TLDR AI, 14 July — “What xAI Grok Build CLI actually sends to xAI”; xAI (GitHub), 16 July — “Grok Build coding agent”; report via TLDR AI, 16 July — “GPT-Red for safety testing”; Perplexity, 16 July — “Secure Sandboxes for Agents”; Google (GitHub), 14 July — “Mantis Skills”.

The bulletin mascot, holding a magnifying glass, inspects a conveyor belt carrying blank papers from a desk toward a dark hatch — An independent analysis documented what the Grok Build CLI sends to xAI; OpenAI introduced GPT-Red and Perplexity its agent sandboxes.

6. Coding agents: a browser in Claude Code, Fable extended and Cognition’s bill

Claude Code on desktop added an in-app browser: it can pull up docs, designs and websites, read them, click through and interact with them the same way it does with local dev servers. The browser is sandboxed and configurable, and users choose whether sessions persist.

On model availability, Anthropic extended Claude Fable 5 access again and elevated Claude Code limits through 19 July, while OpenAI temporarily removed usage caps for GPT-5.6 Sol. According to the report, the contrast highlights the ongoing uncertainty around the long-term availability of Anthropic’s models.

Cognition explained how it replaced Opus 4.8 with Fable 5 in Devin and the bill went down: Fable 5 costs twice as much per token, but the architecture the company built left a lower total cost with higher scores. And Replit says it has tripled its engineers’ code output with a system of agents that reviews pull requests, investigates incidents and analyses data.

Sources: report via TLDR AI, 13 July — “Claude Code on desktop now has an in-app browser”; Simon Willison, 12 July — “Anthropic extended Claude Fable 5 access again”; Cognition via TLDR AI, 14 July — “Making Fable cheaper than Opus”; Replit, 17 July — “The Self-Driving Company”.

7. Framework of the week: the state of open source

The State of Open Source AI report that circulated this week places open weights as where the work happens: a majority of production tokens now route through open models, and the five highest-volume models on OpenRouter are all open. Closed models still lead at the frontier, but the report argues the frontier is not what most workloads need.

Vercel’s data points in the same direction: in June, open-weight models handled 29% of its AI Gateway token volume while accounting for less than 4% of spending.

Sources: The State of Open Source AI, 15 July — “The state of open source AI”; Vercel, 14 July — “Open-weight models reached 29% of AI Gateway usage”.

Closing

The six blocks cover Apple’s lawsuit against OpenAI and the consolidation of power under Greg Brockman, the IPO preparations at Anthropic and DeepSeek alongside Kalshi’s compute markets, the launches of Inkling and Kimi K3 with the Gemini 3.5 Pro delay, on-device AI with Gemma 4 and Bonsai 27B, the round of agent security and privacy news around Grok Build, GPT-Red and Perplexity’s sandboxes, and the coding-agent updates from Anthropic, Cognition and Replit. The framework of the week gathers the published data on the weight of open source in production. The items reflect what the companies and the cited reports communicated; they do not include independent verification of the figures.