OpenAI just quietly shipped something more important than another benchmark. The new Scheduled tasks page in ChatGPT is not cosmetic. It is the moment the assistant stops waiting for your prompt and starts acting on its own, at set times, with access to your chats, memory, Gmail and calendar. Pulse lasted 14 days. What replaces it is far more practical: AI that runs tasks in the background instead of just answering questions. This is exactly the boring infrastructure that decides whether AI stays in your daily life or becomes another tab you open once a week. Everyone watches the models. The real fight is over who becomes your default scheduler first. The question is no longer whether you hand AI part of your calendar. It is: to whom?

The "Amazon took down Anthropic" story is incomplete. There were two triggers, not one. Weeks before the Amazon report, Anthropic's Mythos access list quietly ballooned by about 50 entities it did not disclose for days. One of them was a Korean telecom the White House suspected of China ties. That is when the trust started cracking. Then came Amazon's Fable jailbreak report, and that was the final straw that triggered the actual export controls. SK Telecom denies any China link and the name was never officially confirmed, so hold that part loosely. But the real story is not one traitor. It is that a company which spent years warning its model was too dangerous to release lost track of who already had it, before anyone even jailbroke it. You cannot tell the world your AI is a weapon and then misplace the guest list.

A year ago open source was a punchline. Today GLM-5.2 lands as a leading open weights model, reportedly around 51 on the Artificial Analysis Index, in the same range as Opus 4.6 and GPT-5.4. Look at the trend, not the single number. The gap between the best open model and the best closed one has been closing every quarter, and it is now roughly 3 to 4 months, not years. And the timing is brutal. The same week US frontier models get rationed and export controlled, an open weights model you can run yourself walks right past that wall. Closed labs are racing each other. Open source is racing the calendar, and the calendar is losing.

Everyone is writing ChatGPT's obituary. Read the numbers calmly. Yes, its share fell below 50 percent for the first time ever, down to 46.4 percent. Two years ago it was 81 percent. The decline is real. But ChatGPT still has 1.1 billion monthly users. That is more than Gemini and Claude combined. Scale did not die. Here is what actually changed. The game moved from growth to monetization and trust. Claude converts 13 percent of users to paid, the highest on the market, on a fraction of OpenAI's reach. Gemini grows on Google distribution, not product. And the new part, absent a year ago. After the OpenAI Pentagon deal, uninstalls spiked. People started switching models on values, not just features. The market no longer picks the biggest. It picks the one it trusts and pays for. What are you building on, reach or trust?

Everyone is reading the Microsoft DeepSeek story as geopolitics. Wrong story. Microsoft made Copilot Cowork generally available today and quietly moved it to usage based pricing. The reason, straight from their Copilot EVP: users running hundreds of tasks a week make unlimited plans impossible. The DeepSeek part is real but smaller than the headlines. It would be optional, fully hosted on Azure, your data staying in Microsoft's cloud. Not a defection from OpenAI. A cheaper lane on the same road. The actual signal: agentic AI cannot be priced like Office. The flat monthly seat is dying. Every serious agent product is moving to the meter. Unlimited AI was never killed by politics. It was killed by math. Who still believes in flat rate AI agents in 2026?

This is the export control story nobody wanted. The chips were meant to slow China down. Instead, cutting off Nvidia forced every major Chinese firm to fund domestic GPU makers with real money, and now ByteDance is buying 50,000 inference chips from a startup that a year ago sold mostly to the government. You do not build a homegrown chip industry with slogans, you build it by removing the easy option. The ban did not stop the demand, it just redirected it inward. Sometimes the wall you build around a rival is the scaffolding they climb.

Read it carefully, this is a pause, not a reversal. Anthropic still plans to move Agent SDK and claude -p off subscription limits onto a separate monthly credit, they just hit the brakes after backlash. So for now your SDK apps keep drawing from your existing plan, but treat that as a reprieve with a clock on it, not a win. This fits the whole week, weekly caps, the Max lawsuit, now this. The direction is clear, anything agentic that burns tokens hard is being walked toward its own meter. Build with that assumption, not against it.

Good for devs, but the framing oversells it. Grok did not just arrive in Warp, it joined a list that already had Claude Code, Codex and Gemini CLI sitting there. The real story is that the terminal, not the chat window, is now where the model war is fought. And here is the part the hype skips, grok-build-0.1 still has no independent SWE-bench number, so this is a speed and access pitch, not a proven quality one. Nice to have the option, just judge it on a real task, not a launch post.

208 euros to nearly 1000 in one year. Same DDR5 kit, 419 percent of July 2025 pricing. Memory did not get harder to make, the makers just sent every wafer to HBM for AI instead of consumers. This is what the AI boom actually costs the rest of us, and it is heading for 500 percent. The data center gets its memory. You get the bill.

Read what just happened to the gap. K2.7 scores 81 percent on MCP tool use, beating Opus 4.8, and narrowed the distance to GPT-5.5 from 18 points to 7. Open weights, 256K context, roughly 5x cheaper, and it runs on vLLM on your own hardware. It still loses badly on novel research, 35 to Opus 81, so this is not frontier in everything. But for everyday agentic coding the open models just stopped being the compromise option.

I run Max and I have hit this exact wall, so the complaint is not crazy. The problem is the label. 20x sounds like a capacity promise, but it is 20x of Pro spread across rolling and weekly caps, and one heavy coding day can eat a chunk of your whole week. The limits are real and not secret, the issue is that 20x simply does not feel like 20x in practice. Call it what it is, the multiplier is marketing and the cap is reality.

The framing is Amazon versus Anthropic, but look at the relationship. Amazon is Anthropic's largest investor, a board member and its cloud host. Its researchers found a jailbreak and Andy Jassy carried it straight to the White House, and 90 minutes later Anthropic's best models were dead worldwide. The dispute is whether this was a real Mythos level cyber capability or a minor flaw already present in other models. But the precedent is the real story. Your biggest backer can now trigger a federal export control against your flagship product. Anthropic spent years warning AI was too dangerous to ship unguarded. They were right. The danger was sitting on their own board.

While the frontier labs spent this week pulling their best models offline and rationing compute by the week, open source quietly did the opposite. vLLM 0.23 shipped 408 commits from 200 contributors, 63 of them brand new, squeezing 20 percent more out of FP8 kernels and halving KV cache memory on hardware you already own. No press tour, no valuation, no 90 minute takedown call. One side is learning that its best model is too dangerous to ship. The other is making last year's GPU run this year's models faster. Bet on the layer nobody can switch off.

Read what this actually is. 5 trillion from robotaxis, 25 trillion from Optimus, all from a man whose timelines are a running joke and whose robots still need humans behind the scenes. That is not a forecast, it is a valuation slide that assumes every hard problem is already solved. Physical labor as the final bottleneck is a great line, but Optimus cannot reliably fold laundry yet. Wake me when one does a full shift with nobody teleoperating it.

The detail that should worry everyone is not the takedown, it is how it happened. A 1 p.m. call, 90 minutes to pull the models, citing a national security threat with no specifics provided. By 5:21 p.m. the export controls were live, hitting every foreign national on earth, including Anthropic's own non-US staff. Unable to verify citizenship at the API level, the company had to kill both models for everyone. First time US export controls hit AI software, executed in an afternoon, on a threat the company says it was never shown. Whatever you think of Anthropic, governing this fast with this little detail is its own risk.

Everyone quotes Gibson for predicting cyberspace and VR. But the thing that came true this week is the less glamorous one: the Turing Police. In Neuromancer, from 1984, that was a state unit whose job was to keep AI from getting too capable, and to shut it down if it crossed the line. A government closing Fable 5 as "too dangerous" is exactly that picture, 2026 edition. Gibson did not nail the technology so much as who would keep a hand on the kill switch. And that was always the scariest part of the book, not the machine itself.