OpenAI launches GPT-5.6 — and a new vision for ChatGPT Work
OpenAI dropped GPT-5.6 today alongside a broader “ChatGPT Work” push — positioning it as an agent that can stay with a project for hours, take action across apps and files, and turn a vague goal into finished output. It’s also now the preferred model powering Microsoft 365 Copilot, which is either a vote of confidence or a contractual obligation, depending on how cynical you’re feeling about the Microsoft relationship lately. The “most ambitious work” framing is doing a lot of heavy lifting in the marketing copy, but the underlying capability jump looks real. We’ll see if it earns the hype or just generates more of it.
The ChatGPT browser is already dead
Atlas, OpenAI’s agentic browser, launched in October and is now being “sunset” — a lifespan of roughly eight months, which in AI product years is practically a geological era, but in normal-people years is embarrassingly short. OpenAI says agentic browsing features are migrating to its desktop app and a Chrome extension, which raises the question of why Atlas needed to exist as a standalone product in the first place. The answer, probably, is that it needed to exist long enough to figure out what should replace it. Pivot complete.
Anthropic Wants You to Pay Up for Claude Fable 5
Anthropic is moving to usage-based fees on top of subscriptions for access to its best consumer model — Wired calls it a sign that “the golden era of AI subscriptions is ending,” and I think they’re right. The flat-fee, all-you-can-eat model was always a loss leader; the companies needed users to build habits before the bill came due. That bill is now arriving. Expect every major AI lab to follow suit within the next six months, and expect the current round of subscriber complaints to get very loud very fast.
Fidji Simo steps down from leading OpenAI’s AGI work due to illness
Fidji Simo, who was OpenAI’s No. 2 executive overseeing AGI deployment, is transitioning to a part-time advisory role after her medical leave for a neuroimmune condition extended longer than expected. This is genuinely unfortunate on a human level — nobody chooses to step back from a role like this — but the timing is rough for OpenAI, which is navigating an IPO runway and an increasingly fierce enterprise competition with Anthropic. That’s two senior departures in as many days, and “part-time advisor” is the corporate equivalent of a slow fade.
Anthropic found a hidden space where Claude puzzles over concepts
This is the most genuinely interesting research story of the week: Anthropic built a tool called the “Jacobian lens” that gives researchers a clearer look at what’s actually happening inside Claude’s activations as it processes a question — a kind of cognitive MRI for LLMs. What they found ranges from the mundane to, as MIT Tech Review puts it, “the unnerving.” We don’t yet have full details on what the unnerving part is, which is doing a lot of work in that sentence. Interpretability research like this matters enormously — you can’t align what you can’t see.
Microsoft’s carbon emissions went up 25 percent last year
Microsoft’s 2026 sustainability report reveals a 25% jump in carbon emissions last year, totaling 34 million metric tons — driven, the company says, by data center expansion for AI. To be fair, Microsoft adds the qualifier “without select interventions,” which is corporate for “we’re counting the carbon offsets, please don’t panic.” But the gap between Microsoft’s public climate commitments and its actual emissions trajectory keeps widening, and at some point the interventions need to actually show up in the numbers. The AI buildout has a very real environmental cost, and the industry keeps finding creative ways to footnote it.
Patreon Blocks Crawlers From Stealing Creators’ Work for AI Training
Patreon CEO Jack Conte announced a Cloudflare partnership to block AI crawlers from scraping creator content for training data, with a statement that included the phrase “the crawlers can stay the fuck off Patreon.” Points for clarity. This is part of a growing creator platform backlash against AI companies treating the internet as a free buffet, and Conte’s framing — credit, compensation, and consent — is the right framework for this fight. Whether it actually works technically is a separate question, but as a statement of values it’s refreshingly direct.
LinkedIn and X Are Flooded With AI Spam, Browsing Data Suggests
An AI detection company analyzed real browsing data and found that the volume of AI-generated content users encounter daily on LinkedIn and X is “shockingly high” — which, if you’ve scrolled either platform recently, will surprise approximately no one. The interesting part is that this is measured from actual browsing behavior, not just posted content counts, meaning the AI slop is reaching eyeballs at scale. We built tools to generate infinite content and platforms with infinite scroll. Turns out those two things together are a problem.
Bottom Line
The AI industry spent this Friday launching new models, killing old products, raising prices, and emitting a lot of carbon — a fairly complete picture of where we are in 2026.