This week in one sentence: model prices are falling while the compute bill is rising. What is getting cheaper is the price of a single model call; what is getting more expensive is the infrastructure needed to actually deploy AI at scale.
This English edition is written from English-language sources, listed at the bottom. It is not a translation of the Chinese edition.
Five stories
Anthropic filed for an IPO. On October 26, Anthropic filed an S-1 for a Nasdaq listing under the ticker ANT. It is the first time a frontier model lab has put its financials in the open. The next question is no longer benchmark scores, but whether it actually makes money. Source: Writingmate, AI News, Week of October 26
Kimi K3.1 Pro shipped, with open weights promised for November 4. Moonshot's new model is a 2.8T-parameter MoE with 104B active parameters and a 1M-token context window, priced at $2 per million input tokens and $10 per million output tokens, down from $3/$15 for K3. On Artificial Analysis's Intelligence Index it scored 55, fourth overall, against 58 for Claude Opus 5.5. The gap between open and closed frontier models has narrowed to single digits. Source: Writingmate, AI News, Week of October 26
OpenAI rewrote its Model Spec around five values. On October 29, OpenAI updated the Model Spec, its public description of how its models should behave, reorganising it around faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. The update adds guidance on context, memory and oversight, retires most detailed style guidance, and now refers to the model as “GPT” rather than “the assistant”. OpenAI itself notes its current models do not yet fully reflect the new Spec. Sources: OpenAI release notes; Reuters, October 29
Big tech earnings: the AI bill is now the story. Microsoft, Alphabet and Amazon reported on October 28–29 against combined 2026 AI capital-expenditure guidance of roughly $600–735 billion. Meta had already shown the market's new mood a week earlier: revenue of $64.6 billion, up 26%, but third-quarter capex of $34.9 billion, free cash flow turning negative, and a raised 2026 spending range of $135–145 billion — its shares fell about 7% the next day, and it has since announced plans to raise up to $95 billion in equity. Analysts expected Azure at about $33 billion for the quarter, Google Cloud up about 95% to $29.6 billion, and AWS up about 38%; the question investors kept asking was not whether cloud revenue grew, but whether it grows fast enough to pay for the build-out. Sources: Finimize; First for Money; ValueAddVC capex tracker
Two cyber models, opposite access policies, 24 hours apart. On October 28 Google announced Gemini 4 Carbon — it says the model scores 59 on the Artificial Analysis Intelligence Index, first overall, with 88.2% on SEC-bench Pro — but limited first access to a subset of vetted cyber defenders through its Fairwind Program. On October 29 Mistral released the weights of Mistral Large 4, a roughly trillion-parameter model, on Hugging Face under a modified MIT licence, with a hosted price of $0.70 per million input tokens. Same week, same pitch, opposite answers to who gets the most capable security model. Sources: Model Dex News; Transclew; AI World Wire weekly wrap
Deep dive: why are models getting cheaper while AI gets more expensive?
The most interesting thing this week is not any single model, but two numbers moving in opposite directions.
On one side, inference prices are collapsing: one analysis put the June-to-October price cut at OpenAI and Anthropic at up to 80%, driven by Chinese competition, better inference efficiency, and a blunt reality — GPU capacity is perishable like airline seats, so cutting prices to raise utilisation is rational. Kimi at $2/$10 and GPT-6.1 Terra at $1/$5 are the latest data points, not exceptions.
On the other side, build-out costs are rising: the four largest hyperscalers alone plan roughly $735 billion of 2026 capex, up about 79% from about $410 billion in 2025, with Amazon at about $220 billion, Alphabet at $195–205 billion, Microsoft at about $175 billion and Meta at $135–145 billion. Cheap model calls do not mean cheap AI. What is really expensive is data centres, power, chips, and the engineering work of embedding models into real businesses.
For ordinary users and small teams, this week's practical takeaway is simple: do not chase the most expensive flagship model; get good at using the strong models that are getting cheaper. The bill you should actually watch is your own compute, data, and deployment cost — not the leaderboard. Next week's Anthropic S-1 financial details will test this further: if even a leading lab needs an IPO to fund compute, then “cheap models, expensive infrastructure” is not a one-week story, but the main theme for years to come.
What to watch next week
November 4: whether Moonshot delivers the promised Kimi K3.1 open weights and licence on schedule;
The financial details in Anthropic's S-1 once they are public: revenue, losses, and the structure of its compute spending;
Whether Gemini 4 Carbon widens beyond vetted cyber defenders, and Mistral's promised Large 4 technical report.
Note: this issue is a curated summary and commentary on public reporting. Prices and earnings figures follow the original reports; if a source corrects them later, the correction stands.