The frontier took Saturday off; the constraints around it did not. No new model card shipped overnight, and the federal seat that would set the pace of this industry is still empty, past a blown “three or four days” clock. But the forces that actually price AI moved on three fronts. Two Bay Area municipalities are squeezing the data-center buildout through zoning, with Oakland’s full council voting Tuesday on a 45-day moratorium. Anthropic slid its most capable model into the free tier without touching a rate card. And OpenAI retired a production model on twenty days’ notice, against a policy that advertises six months. On a day with no model news, the economics, the land and the lifecycle are the news.
The AI buildout just got a zoning board
What happened. Oakland’s city council votes Oct. 6 on a 45-day pause on new data-center projects; the measure already cleared its Life Enrichment Committee unanimously, and state law behind it is moving too — the seven data-center oversight bills Newsom signed this week, which give communities more control over water, electricity and land use. San Jose is running a resident-driven regulation process toward “Uniform Standards” with a community session Oct. 15, in scope a 444,400-square-foot, 99-megawatt South San Jose facility planned by Menlo Digital near residential streets. The precedents are already on the books elsewhere in California: Monterey Park passed a permanent data-center ban, the first city in the nation to do so, and Oakley became the first Bay Area city to halt new proposals.
Why it matters. The buildout’s binding constraint just moved from wafers to permits. Yesterday’s note watched DRAM move the local-AI price ladder; this is the other half of the same bill. Even with silicon in hand, standing up capacity in the Bay is now a city-council decision about power, water, noise and displacement — none of which track a GPU roadmap. That puts two costs into an operator’s model that nobody priced: a permitting timeline slower than any hardware cycle, and a power picture set by politics rather than by the grid’s actual headroom. The moratoria are pause-while-regulating, not bans; the durable number to watch is not the 45 days but what the regulations say when they land — cost pass-through is the lever that will actually show up in inference bills.
Source: oaklandside.org, sanjosespotlight.com
The capable tier just went free
What happened. Claude Sonnet 5.5 — the mid-tier model that shipped last week at unchanged prices — is now what runs the free tier on claude.ai. Simon Willison threw his standard capability probe at it, an HTML page rendering a three-dimensional pelican riding a bicycle in WebGL, and got a working page back, then flagged the contrast: ChatGPT’s free tier runs Luna 5.6, so Anthropic is currently offering the more capable free model of the two.
Why it matters. The price war this column has been tracking played out on the rate card — list prices, per-task cost, cache reads. The free tier is the frontier repricing itself without a rate card at all, and for a small builder it does shift the floor: prototyping and light workloads now run on a genuinely competent model at zero, deferring the “what does it cost at load” question to whatever scale you actually hit. Read it as what it is, though — a tap, not charity. Anthropic gets real prompts, habit, and the upgrade path, and “free” is how an incumbent defends share when it is losing on list price. The durable variables for anyone shipping are still per-task cost and the data terms, because a free tier surfaced to the frontier grade is a product decision with a meter on it, not the end of the bill.
Source: simonwillison.net
Twenty days’ notice on a production model
What happened. OpenAI’s deprecations page records that gpt-5.4-cyber was withdrawn from the API on Oct. 1, twenty days after its Sept. 11 announcement. The same page promises at least six months of notice for generally available models, and the same Oct. 1 update gives the other models in the batch — gpt-5.1, gpt-5.3-codex, gpt-5.4-nano — the full six months to an April 1, 2027 removal. The recommended replacement for the cyber model is the non-pointer: “the most capable cyber model available to you.”
Why it matters. This is the lifecycle risk hiding inside the model-name era, and the deprecation policy’s own carve-out — retire sooner on safety or compliance grounds, with “as much notice as reasonable” — is the door the twenty-day cut walked through. Price in the pattern: the more security-adjacent the model, the shorter the sunset, which is precisely the class of model you pin an agent to. Twenty days is shorter than a re-eval-and-migrate cycle for most stacks; a three-week clock on a model you’ve tuned against is a production incident with an announcement date. The durable fix isn’t a favorite model, it’s a router plus a test lane, treating a model id as a version pin with an expiry — the same lesson as the between_tools migration that silently bit Sonnet 5.5 adopters, only now it is the whole model disappearing, not a setting.
Source: developers.openai.com
The Rest
- Agent verification is moving from the fact to the source — Hugging Face’s write-up of ProvenanceGuard, a source-aware verifier for MCP agents: a claim that is true somewhere in the pooled evidence but attributed to the wrong tool passes a source-blind check, and a source-aware one catches it by reading the captured MCP trace instead of collapsing it into one anonymous context. When agents pull from a row of tools, attribution becomes a correctness property. huggingface.co
- LangSmith Engine v2 brings red-teaming to the evals layer — Engine now generates hypotheses about failures that have not surfaced in production, tests them against your traces and repos, and validates its proposed fixes before you deploy; Managed Deep Agents v0.8 adds user-level memory, user-owned credentials, and Slack/HTTP channels. The observability layer is finally catching up to the multi-agent workload people actually run. langchain.com
- The buildout’s power supply just grew a second naval chokepoint — Houthi forces have seized Yemen’s entire Red Sea coast, including the Bab al-Mandab strait; Yemeni armed forces report 474 attacks and 1,540 Houthi casualties in 24 hours, Saudi Arabia intercepted three ballistic missiles, and Washington is urging reimposed UN “snapback” sanctions on Iran. With Hormuz already contested, two straits now sit in an active maritime war — and energy volatility is a compute-cost variable, not a headline. aljazeera.com
- Israel’s Oct. 27 election is tightening — the latest ToI poll puts the pro-Netanyahu bloc at 51 seats with Winter’s party below the threshold and neither bloc near a majority. Political uncertainty is an ecosystem variable for the country that is a first-rank AI talent and capital hub. timesofisrael.com
- Triple-digit heat is the weekend’s power stress test — the first fall heat event across the Bay Area puts load on the same grid the data centers now under moratorium would draw from, a sharp little preview of the season guarding the buildout’s other constraint. kron4.com
What I’m watching
The AI-czar seat, which has now blown through the “three or four days” clock floated since mid-September — as of yesterday the float is intact, every wire still future-tense, the White House’s only comment that reporting before the President announces is “baseless speculation,” and the WSJ’s that the decision “isn’t final.” It stays armed because the instant it lands, it stops being politics and becomes the single most load-bearing governance fact for every lab and operator. Also watching Tuesday’s full-council Oakland vote and whether the San Jose resident-driven process, still just proposing “uniform standards” this week, hardens into rules with permitting teeth. And on the free tier: whether Anthropic publishes the data-retention terms for the Sonnet-powered claude.ai free tier, because a frontier-grade free model defaults to being a training and analytics hook unless the terms say otherwise.