The day the US government switched off Fable 5

Just after midnight Pacific time, Anthropic disabled Claude Fable 5 and Mythos 5 for every customer on the planet. The detail that captures it best comes from Simon Willison, who had a little script hitting the API once a minute just to measure how long he’d keep access. It cut out and he posted the exact timestamp of the moment he lost it. The reason, according to Anthropic’s own statement, is an export control directive.

The US government, citing national security authorities, ordered access suspended for any foreign national, whether they’re inside the United States or outside it, and that explicitly includes Anthropic’s own foreign national employees. The company says the only way to comply on that timeline was to pull the models entirely rather than build per-user gating overnight, so Fable and Mythos went dark for everyone. Their words were, we believe this is a misunderstanding, and we’re working on it. to restore access as soon as possible.

What made this more than a status page footnote was the scramble underneath it, because it showed how deep these models had already gotten into real production systems in just a few days. 8 and GPT-5.5 available. 8 at Opus Pricing. Replit did the same.

And the coding subreddits turned into something between a wake and a refund line, memorial threads, people asking if they could get their money back, a lot of RIP posts, all for a model that had been generally available for a grand total of three days. There are two questions hanging over the next week and neither one has an answer yet. The first is narrow and practical. Does that word, misunderstanding, mean this gets walked back in a few days, or does the export control framing harden into something permanent?

The second is, what does it tell every company outside the United States that a frontier vendor’s flagship model can simply disappear on a government say-so, with essentially no notice? Gergely Orosz said the quiet part out loud almost immediately. He pointed out that suddenly all that investment in Nonus Labs, the mistrals of the world, doesn’t look wasted at all. For a Nonus company, he argued, it now looks strategic to have a Nonus vendor in your stack.

That’s a sentiment that did not exist at this intensity a week ago. And that’s exactly the backdrop for the day’s other big release. VentureBeat reported that Xiaomi shipped something called MiMo Code, an open-source agentic coding harness, weights included, and they’re claiming it beats Claude Code on ultra-long tasks, the kind that run past 200 steps. Now, I’d take any vendor’s own benchmark with caution, but the metric they’re leaning on is the interesting one.

It’s not a single SWE-bench score. It’s horizon length, how many steps the agent can take before it loses the thread. That’s genuinely where most harnesses fall apart. They start looping, they forget what they were doing, they drift off the original goal somewhere past a few dozen steps.

So an open release that claims to hold coherence across 200-plus steps is the sort of thing teams actually fork and test rather than just nod at. And put it next to the access shock from Anthropic, and the pitch for open weights basically writes itself. If you can’t bet your entire roadmap on one American API… that a directive can switch off, a capable open harness you can run yourself starts looking less like a hobby and more like risk management.

The next story is quieter, but if you’re building agents, maybe the most directly useful thing today. GitHub published real production numbers on a problem every single agent builder is running into right now, which is, when should an agent delegate to a sub-agent, and when should it just do the work itself? Their post on smarter sub-agent delegation in Copilot CLI is unusually concrete. What they found was that eager delegation was the problem.

The main agent would spin up an exploration sub-agent for a task it already had enough context to finish, and then it would sit there idling while the sub-agent went and researched the whole repository and tripped over stale file paths. So they tighten the orchestration policy, start with the narrowest path that works, only escalate to a sub-agent when the complexity genuinely creates leverage, and treat sub-agents as a way to parallelize, not as a pause button where the main agent just waits. The results from their production A-B test, tool failures per session down 23%, search tool failures down 27%, P95 wait time down 5% and no quality regression, and the lesson there generalizes way past Copilot, more agents is not the same thing as better orchestration. The win came from the main agent handing off less, not from any individual model call getting faster.

Google pushed the agent story in a completely different direction. There’s a proposed standard called WebMCP that’s now entering origin trials in Chrome 149. The idea is that a website can expose tools, actual JavaScript functions and HTML forms, directly to an AI agent running in the browser, and the whole point is to kill DOM scraping and screenshot reading as the way agents act on the web. So, instead of an agent guessing at the CSS selector for a checkout button and breaking the moment the site redesigns, the agent just calls a tool that the site will not read, and the agent just calls a tool that the site will not read, and the agent just calls a tool that the site will not read, and the agent just calls a tool that the site will not read, and the agent just calls a tool that the site will not read, and the agent site author declared on purpose it’s basically the browser native cousin of mcp and if this actually catches on it changes how web automation gets built away from brittle playwright style scripts that break constantly and toward sites that advertise their own agent surface origin trials are early this is not shipping to everyone tomorrow but a real flag in a real chrome build is a much stronger signal than yet another spec document nobody implements a quick one on security tooling because it’s a useful gut check Semgrep published how they actually benchmark ai powered vulnerability detection and what current models really score their headline is three and a half times more true positives than their previous approach but honestly the more valuable part for practitioners is the methodology how they separate real findings from the confident but wrong noise because that noise is the thing that kills these tools in practice if you’re evaluating an ai scanner the false positive rate is the number that decides whether your engineers actually trust it or whether they start ignoring every alert by the second week a high true positive count means nothing if it’s buried in junk let me close with two voices that framed where this is all heading swix floated a deliberately provocative idea he said look after the pull request dies and after code review dies maybe git itself needs to die next his argument is that something like 20 to 30 percent of code spend goes into managing merge conflicts and that’s a workflow we cargo culted from before agents existed when humans collaborate in a notion doc or a linear board we don’t do line by line merges we just chat and an owner ships it so the future code base in his telling might look more like a database than a tree of git objects less efficient more scalable you don’t have to buy the whole thing to notice that version control is one of the very few parts of the stack that agents haven’t reorganized yet and separately Replit’s Amjad Masad used this moment to publicly refuse to build a token consumption leaderboard that enterprise customers had asked him for his point was that the whole token maxing craze rewards burning tokens instead of shipping outcomes and he’s not in the business of selling tokens for their own sake with the field still digesting last week’s numbers on DeepSeek versus Anthropic and raw token volume a vendor standing up and declining to optimize for tokens is itself a signal worth noting so here’s what i’m watching whether anthropic restores fable and mythos in a matter of days or whether that export control framing hardens into something that sticks the first version is a logistics hiccup you forget in a week the second version genuinely changes how anyone outside the united states sources frontier capability and if that’s the world we’re in MiMo Code is not going to be the last open weights release built to catch exactly that demand that’s the thread to pull on this week it’s not going to be the last open weights release built to catch exactly that demand that’s the thread to pull on this week