AI Weekly #7 — Sep 7–13, 2026

Three companies shipped three different answers to the same question this week: where does the agent loop run, and who pays for the tokens it burns getting there.

Sakana AI's answer is that the loop is the model. Fugu Ultra v2 and Fugu Max are orchestration systems trained to route each task to the leanest model in a pool that can solve it, and Ultra v2 scores 48.3 on Chartography against 27.3 for Opus 5 and 29.5 for Fable 5 — with Fable 5, Fable 5.1 and GPT-6 Astra explicitly excluded from the pool it draws on. OpenAI's answer is that the loop is infrastructure: the Agents API, in public beta since Thursday, runs the loop on OpenAI's own servers, coordinating model calls, tool use and context, with hosted sandboxes for code and files. DeepSeek's answer is architectural. V4.1-Flash activates roughly 8B parameters per token while reading input and 16B while generating — a design shaped around the fact that agents read enormously more than they write.

None of those is a benchmark story, though all three arrive with benchmarks attached. They are three bets on which layer is worth money a year from now.

The "we don't need your model" argument got a number

Sakana's exclusion is worth sitting with. It is one thing to build a router that sends easy work to a cheap model. It is another to publish a table showing the router beating two frontier models while stating that neither is in the pool it calls.

Treat the specific figure carefully. Chartography is one benchmark, chosen by the vendor, and "top two on seven of eight" is the sort of claim that rewards reading the eighth. But the structural point survives the scepticism: if orchestration over open and specialised models is competitive with frontier weights on real work, then the frontier lab's moat is not the weights. It is the distribution. That is a much less comfortable thing to defend, and it is the same position DeepSeek is attacking from the other side by putting a million-token multimodal model under an MIT licence.

The practical reading for anyone buying inference is that "which model" is turning into "which policy for choosing a model", and those are priced differently.

Jurisdictions showed up

Three institutions drew lines around models this week, and none of them were labs shipping product.

China's Supreme People's Court published 24 articles of judicial guidance on AI disputes on Monday — liability for deepfakes and voice cloning, algorithmic price discrimination, autonomous vehicle fault, training data, AI-generated evidence. It is the first national judicial guidance on AI from that court, and it works entirely through existing law: the Civil Code, the Copyright Law, the Personal Information Protection Law. There is no AI act underneath it. It also declines to settle whether AI-generated works are copyrightable, which is the question most people wanted answered.

Mistral raised €3 billion at a €21 billion valuation on an argument about control rather than capability — data inside your boundary, models you can customise, compute that is private and predictable. The Grand Duchy of Luxembourg is on the cap table. Whatever you make of the sovereignty framing, a national government investing directly in a model vendor converts it from a marketing line into a procurement precedent.

And Anthropic published capability evaluations on tactical intelligence targeting and conventional weapons development: geolocation, account linkage across platforms, drone terminal guidance under GPS jamming. Frontier models geolocated outdoor photographs with median errors of 37–47 km where human expert baselines sat at 151–174 km. A lab publishing its own numbers on military-relevant capability is a different kind of argument from a position paper, and it lands squarely in the open-weights debate — the open models tested trailed the frontier, but they were not absent from the results.

By the numbers

The directory took in 21 entries this week — 10 apps, 8 agent skills and 3 MCP servers — bringing the totals to 495 apps, 251 skills and 276 MCP servers.

The more useful number is four. That is how many of the 21 turned out to have shipped something during the week once we checked upstream. The rest are real, catalogued and searchable, but their actual release dates land anywhere between July 2025 and June 2026. Golioth's tinymcp was announced in July 2025; Arm's MCP server in October 2025; the five NVIDIA skills we picked up on Sunday are mirrored daily into NVIDIA's catalogue from product repositories where they have lived for months.

That gap between indexed this week and released this week is the entire reason the picks section exists, and it is wider than it looks from outside. A directory reporting its own intake as news would have run 21 items and been wrong about seventeen of them. The skills catalogue is the clearest case: it is now dominated by vendor-maintained collections that grow continuously rather than launch, which makes "new" a genuinely hard word to apply — and makes opening the upstream source the only honest way to apply it.

What we added

Four entries below shipped upstream during the week. Last week's issue covers what came before.