Anthropic’s Tuesday countdown card landed this morning as Claude Opus 5.5 — $4 and $20 per million tokens, cache reads down to $0.20, and a claim that typical workloads cost 40 percent less to run than its own Opus 5 while performing at the level of Claude Fable 5.1. Today’s breaking note filed the facts within the hour the card dropped (claude-opus-5-5-ships-cheaper-than-the-flagship); the column’s job is the scorecard everyone owes a preview. The leak cluster said “Opus 5.5, Tuesday, ~20% cheaper,” and every element of that briefing held — then went deeper on the line that actually bills you. Beneath the card, the day’s other news was a buildout: Alibaba unveiled its Zhenwu V900 chip and a roughly 10-trillion-parameter Qwen flagship in training, Xiaomi’s open-weight MiMo-V2.6 sits on top of the open-model index, and Beijing’s regulator is prying into whether DeepSeek and Moonshot routed sensitive user data into Claude. The closed lane shipped a discount; the open lane crossed the Pacific.
The leak’s numbers held — and the discount went deeper where the bill is
What happened. The card that yesterday’s column called “still a leak” shipped at 10:00 as Claude Opus 5.5: $4/M input, $20/M output (20 percent below Opus 5), cache reads at $0.20/M (60 percent below), output over 30 percent faster, and claims of 40 percent less to run on typical workloads. On the vendor bench it leads the agentic-coding lane — Terminal-Bench 4.0 at 66.4 percent against GPT-6 Astra’s 57.9 and Fable 5.1’s 55.8 — and Anthropic says it matches Fable 5.1 across most work. It is the first release since the lab’s public call to slow the frontier, and it went through external evaluation by Frontier Design and METR. The “Fable 5.2” half of the leak did not materialize as its own card; the family is rooted at 5.5, with Sonnet 5.5 and Haiku 5.5 to follow in the coming weeks.
Why it matters. Score the preview against the shipped card and the headline claim was the floor, not the ceiling. “~20% cheaper,” as leaked, was true for tokens — but Anthropic’s own line is that cache reads make up the majority of agentic and coding work costs, so the 60 percent cut on the cache line is the number that changes what a persistent agent costs per hour. The lab that asked the industry to slow down just repriced the exact thing the pace debate is really about: running agents long enough to be useful. When the bench-leading agentic-coding model retails at 20 percent under its own flagship’s input rate — and 60 percent under on the cache line that powers long sessions — the frontier’s competitive unit has quietly stopped being the headline and become the recurring bill.
Source: anthropic.com, theverge.com
The open lane’s ceiling is now a Chinese chip and 20 gigawatts
What happened. Alibaba hit the same Tuesday with a full-stack announcement: the Zhenwu V900 AI chip (positioned squarely at the gap the US export rules and Nvidia left in the China stack), a Qwen flagship plan of roughly 10 trillion parameters — coverage puts the target between 4 and 10 trillion — with the next-gen model entering training, and an outline of about 20 gigawatts of compute capacity by 2032. Alibaba shares rose around 5 percent on the day, and the wave ran across Reuters, Caixin, SCMP, and CNBC.
Why it matters. When Qwen-4 preview shipped open in August, the open lane’s question was whether an open model could stay at the frontier. This tells you the next answer is being built, from scratch, on Alibaba’s own silicon: a 10-trillion-parameter open model, trained and served inside China, changes what “open weights” means for anyone who depends on them. The ceiling on that lane stops being a lab decision and becomes a geopolitical one — which stack the biggest open-weights bet in the world is trained on, and under whose export and data laws it runs. Operators who chose open to avoid vendor lock just need to decide which vendor is now the one making that call.
Source: reuters.com, caixinglobal.com
One API pipe, two investigations pointing opposite ways
What happened. The Information reports that China’s internet regulator is investigating DeepSeek and Moonshot AI over potential data leaks to Anthropic — a probe phase, no charges — after Anthropic alleged both companies had been routing sensitive user data to Claude models. The news lands days before the Trump–Xi talks, and it has run across Yahoo Finance, NDTV, and The Next Web off the single underlying report.
Why it matters. This is the September 9 distillation advisory run in reverse. Then, Anthropic flagged these same labs for feeding its model’s output back into their own weights — training IP flowing into China. Now, Beijing investigates its own companies for routing user data out of China into a US model — privacy flowing out. One cross-border inference pipe, two sovereigns, two investigations in opposite directions. The inference API is no longer just a cost center; it is a jurisdictional artifact, and every multi-region deployment just inherited a regulatory axis that neither Washington nor Beijing agrees to be the neutral ground on. Whatever happens with these two cases, this mirror is the shape of every future cross-border model.
Source: theinformation.com, thenextweb.com
The Rest
- DeepSeek gets a UN Security Council briefing this week — the Council’s AI-risks session, which already had OpenAI and Anthropic in the chair, adds the open lane’s flagship to the trio; the diplomacy lane just became multilateral, and the most interesting question is what the open lane’s leader says on a floor that usually hears from states. reuters.com
- California wants a kill switch on the record — Newsom signed an executive order on AI oversight for state agencies and revived the “kill switch” proposal, the strongest in-state governance move yet for agent operators; worth reading it as an engineering spec, not a slogan, because mandated off-ramps are exactly the thing practitioners should have opinions on before they are drafted by others. calmatters.org
- Gilroy pauses new data centers — the South Bay city adopted a moratorium on new data-center development, the cleanest municipal codification yet of the sentiment that power and land for inference are a planning-board decision, in the same week state regulators moved on the industry. sanjoseinside.com
- Xiaomi’s MiMo-V2.6-Pro tops the open-weight index — open weights shipped at frontier-class performance and it is now the strongest open-weight model on the Artificial Analysis index, a consumer-hardware giant occupying the top slot the same week the US frontier asked for a pause; the interesting part for operators is who owns the lane now. mimo.xiaomi.com
- Israeli startups raised $1.16 billion in the first two days of September — a pace its own ecosystem coverage calls a rankings rewrite, and a useful counterpoint for anyone who watches the region through the conflict-only lens; the HaSadna-relevant tech sector keeps clearing capital at speed. calcalistech.com
What I’m watching
Whether Sonnet 5.5 and Haiku 5.5 carry the same discount down the stack — that is when “cache reads are the real price list” stops being a flagship story and becomes the whole market. Whether OpenAI answers the pricing before DevDay on September 29. And the release-watch that matters most to the open lane: a ~10-trillion-parameter Qwen flagship shipped with weights, a card, and a price — the open weights are the tells, and nothing in the plan says they won’t come.