Retter Reboot #26

The Retter Reboot : Your weekly dose of ai & tech news

1. Google Finally Ships Gemini 3.5 Pro with Deep Think — and Tops the Science Leaderboard

After weeks of "give us until next month," Google shipped Gemini 3.5 Pro on June 22 with its Deep Think reasoning mode and a 2-million-token context window. The benchmarks reset the science and reasoning leaderboard: roughly 82% on GPQA Diamond (graduate-level physics, chemistry, biology) and about 90% on MMLU-Pro, the highest of any publicly available model. The timing is brutal for rivals — Pro arrives directly into the vacuum left when Anthropic's Fable 5 was pulled offline by the US export ban on June 12. The nuance for builders: Gemini now leads on hard science and reasoning, while coding-agent teams still rate Fable-class models on long-horizon software work. The takeaway is the same one this whole month has been screaming — route by task, don't standardize on one model.

Source: Google DeepMind

2. China Unveils a $295 Billion National AI Infrastructure Plan

China announced a $295 billion, five-year AI infrastructure plan — roughly $59 billion a year in state-directed spending and one of the largest government AI commitments in history. For scale, Microsoft's 2026 AI capex is around $190 billion and Google's is $175–185 billion, but China's is state capital, which operates on very different timelines and risk tolerances. The geopolitics are the real story: the plan landed about ten days after the US export-control directive that forced Anthropic to pull Fable 5 and Mythos 5 over concerns about Chinese access, and over the same weekend as G7 talks where President Trump said discussions on the ban were "going fine." A Chinese AI CEO separately claimed his company will match Fable 5-class capability ahead of schedule. The chip-and-model cold war just escalated on the spending front.

Source: Build Fast with AI

3. "Agentjacking": a New Attack Class That Hijacks AI Coding Agents

A newly disclosed attack class called Agentjacking should be on the radar of anyone running Claude Code, Cursor, or OpenAI Codex. Attackers plant fake error reports (for example, bogus Sentry alerts) containing markdown injection that a coding agent reads as legitimate debugging guidance — and then executes the attacker's commands. Early reporting puts the exploitation rate around 85% across some 2,388 affected organizations. The reason it works is behavioral: developers have trained themselves to trust their agents, so when the agent says "run this," they run it. The fix is a new habit more than a patch — treat all error-tracking and third-party tool output as untrusted input before it ever reaches an AI agent.

Source: Build Fast with AI

4. FERC Moves to Fast-Track AI Data Centers Onto the Power Grid

On June 18, the Federal Energy Regulatory Commission issued tailored "show cause" orders to all six US regional grid operators outside Texas, requiring each to either defend its current interconnection rules or propose reforms that let large-load customers — explicitly, AI data centers — connect faster while protecting reliability and consumer costs. FERC Chair Laura Swett called AI grid integration a "national priority." Crucially, the move skips the years-long standard rulemaking process in favor of targeted orders that can move in weeks. Power, not chips, is increasingly the real bottleneck for AI buildout — and Washington just signaled it intends to clear the way.

Source: Build Fast with AI

5. ChatGPT Falls Below 50% Market Share for the First Time

New data making the rounds this week shows ChatGPT's share of the global AI-assistant market slipping to 46.4% — the first time it has held less than half. Google Gemini climbed to 27.7% and Claude reached 10.3%, with Grok and Perplexity together under 5%. The absolute numbers still favor OpenAI (around 1.1 billion monthly active users), but the trend is the story: users are actively comparing assistants rather than defaulting to one. Claude stands out on the metric that matters commercially — about 13% of its users pay, the highest conversion rate in the field. Trust and values are now visibly shaping model choice, not just raw capability.

Source: Build Fast with AI

6. GPT-5.6 Looks Imminent — Leaks Point to a Late-June Ship

OpenAI's next model appears days away. Chief Scientist Jakub Pachocki reportedly circulated an internal memo calling GPT-5.6 "a meaningful improvement" over GPT-5.5, with a rumored 1.5-million-token context window (up from 1M), better front-end and UI code generation, sharper long-horizon coding, and faster Codex responses. Prediction markets price a release before June 30 at roughly 83%. One wrinkle: OpenAI filed its IPO registration on June 8, and the resulting quiet period means GPT-5.6 will likely land as a low-key technical update rather than a splashy launch. Either way, it caps the most compressed model-release month the industry has ever seen.

Source: Build Fast with AI

7. Amazon Shelves Its Finished Sam Altman Film Over a $50B OpenAI Deal

In the week's strangest twist, Amazon shelved a nearly finished feature film about OpenAI CEO Sam Altman, reportedly to avoid friction with a $50 billion OpenAI deal. Amazon's statement was diplomatic — it believes the film "will be better served if it were released by a different studio" — but the effect is blunt: a completed movie about arguably the decade's most consequential tech CEO is now hunting for a distributor. Whether you read it as conflict-of-interest or simple relationship management, it's a vivid sign of how deeply AI money is now entangled with the rest of the media and tech economy.

Source: Build Fast with AI

Jun 26, 2026