Weekly Intel - 2026-07-12

The theme this week is control: who owns the technology, who gets to fix it, who gets to build on it. Apple’s trade-secrets suit against OpenAI, the right-to-repair win against John Deere, and the push to make more chips on U.S. soil all come back to who holds the keys.

AI & Software

GPT-5.6 OpenAI released GPT-5.6 on July 9 as a three-tier model family (Sol, Terra, and Luna), letting organizations pick a cost-versus-capability tier per task, with Terra priced at roughly half of the prior GPT-5.5 generation. The release adds an “ultra” mode that coordinates several agents in parallel, four by default, to break down complex work.

ChatGPT Work OpenAI also launched ChatGPT Work, an agent built on GPT-5.6 that connects to a company’s apps, breaks a project into steps, and produces finished deliverables (slides, docs, spreadsheets, and small web apps) on its own over hours. It runs on the same Codex technology OpenAI already sells to developers, now pointed at non-engineering work, and it competes directly with Anthropic’s Claude Cowork. It is available first on Pro, Enterprise, and Edu plans.

Grok 4.5 xAI released Grok 4.5, a model trained with Cursor on real developer session data and aimed at coding and agentic tasks. It scored 62% on the provider’s DeepSWE 1.0 benchmark, ahead of Anthropic’s Opus 4.8 there, though it trails Opus on the neutral DeepSWE 1.1 test. On efficiency it went further: on SWE-Bench Pro it resolved tasks with roughly a quarter of the output tokens Opus used.

GPT-Live OpenAI introduced GPT-Live, a voice model that listens and speaks at the same time, uses backchannel cues like “mhmm,” and stays quiet while a user is thinking. For harder questions it hands off to GPT-5.5 in the background while keeping the conversation going, splitting a fast conversational layer from a slower reasoning one. It ships in two sizes and replaces Advanced Voice Mode as the ChatGPT default.

Legal & Regulation

Apple sues OpenAI over trade secrets Apple sued OpenAI in federal court on July 10, alleging that former Apple employees took confidential information about its unreleased hardware and handed it to the AI company as OpenAI builds its own AI devices. The complaint names two former employees and OpenAI as the beneficiary, and Apple says more than 400 of its former staff now work there.

John Deere settles on right to repair The FTC and five state attorneys general reached a settlement on July 8 requiring Deere & Co. to give farmers and independent shops the same repair software and tools it provides its own dealers, with FTC oversight of compliance for ten years. It is Deere’s second right-to-repair agreement this year, following a $99 million class-action settlement in April.

Tech Industry

Apple’s $30B Broadcom chip deal Apple committed to a multiyear deal with Broadcom worth more than $30 billion to produce over 15 billion U.S.-made chips through 2031, covering custom silicon and wireless components. Broadcom is putting $1.5 billion into expanding its facility in Fort Collins, Colorado, to make radio-frequency and connectivity parts for future Apple products. It is Apple’s largest single commitment under its American manufacturing program and part of a broader $600 billion U.S. pledge.

Privacy & Governance

EU extends Chat Control to 2028 The EU Parliament let Chat Control 1.0, the interim rule permitting suspicionless scanning of private messages, stay in force until April 2028. A majority of members actually opposed it (314 against, 276 in favor), but the motion to reject fell short of the 361-vote absolute majority a second-reading procedure required, so the regulation stood despite the vote against it. The amendment excluding end-to-end encrypted messages does not change much in practice, since providers were not scanning those under the prior rule, and talks on a permanent Chat Control 2.0 resume in September.

AI Industry Moves

GLM 5.2 and the coming AI margin collapse A widely-shared analysis by Martin Alderson argues the pressure on AI economics comes from inference cost rather than training cost. Training is a fixed, one-time expense; inference scales with every query, and frontier labs currently run roughly 90% gross margins on it. His case is that GLM 5.2, an open-weights model reaching comparable quality at 15 to 20% of the price, is what starts compressing those margins, more than the DeepSeek training-cost panic ever did.

That’s what I’m watching. What caught your attention this week?

-Eric