1. Moonshot’s Kimi K3.1 Pro: near-frontier, with open weights promised

On October 23, Moonshot AI released Kimi K3.1 Pro, a post-trained upgrade of July’s Kimi K3 rather than a new base model. The specs: 2.8 trillion total parameters, 104 billion active per token, a 1-million-token context window, and text-plus-image input. API pricing is $2/$10 per million tokens, down from K3’s $3/$15, with a Highspeed variant at $4/$20 that runs two to three times faster.

The score that matters is independent: 55 on Artificial Analysis’s Intelligence Index, fourth overall behind Claude Opus 5.5 (58), Claude Sonnet 5.5 (56), and Muse Wonder Full (56) — and ahead of GPT-6 Astra and Gemini 4 Argon, both at 53. Moonshot’s own benchmarks put it roughly level with Opus 5.5 and Astra on agentic coding, including 65.7% on Terminal-Bench 4.0 against Opus 5.5’s 66.4%, while its hallucination rate fell from 53% to 44%.

One caveat, stated plainly: the weights are not out yet. Moonshot promises them by November 4. Until then Kimi K3.1 Pro is API-only, and the “top open-weight model” claim is a promise, not a download. Moonshot also says it deliberately removed hacking-related data from training ahead of the open release.

Why it matters

Last week the open-weight leaderboard was a Chinese story with an American footnote. This week the gap is a few index points, not a class apart — and the model closing it costs a third less per token than its predecessor. If the November weights land as promised, the best open model will sit within reach of the best closed ones on agentic work. Watch whether independent evaluators reproduce the 55 once the weights ship.

Good reads

2. OpenAI’s GPT-6 models land on Google Cloud

On October 21, Google Cloud announced that GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna are rolling out in private preview to select enterprise customers in its Model Garden, with general availability planned in the coming weeks. Customers buy through Google Cloud Marketplace: one invoice, existing spend commitments drawn down, and usage counting toward Flexible Savings Plans. Codex CLI and Desktop workflows can already connect to Google’s Gemini platform for inference; the models are coming soon to Gemini for business and Google Antigravity.

Google’s claim is that it is now the only hyperscale cloud offering all three leading closed model families — Gemini, Claude, and GPT — alongside more than 200 other models on one platform. The sequencing matters: Microsoft’s exclusive Azure distribution ended in April, Amazon Bedrock added the models in September, and Google is now the third hyperscaler to sell them. Zero Data Retention and Confidential Compute options are listed as coming soon, not available at launch, and Google has not shared its own prices for these models.

Why it matters

This is distribution, not a new model — and distribution is the story. A frontier model family that was once an Azure exclusive is now a line item on all three major clouds, bought against the same enterprise commitments as storage and compute. For buyers, the practical effect is model choice without a second vendor contract. For the labs, it is a reminder that the moat is shifting from who has the model to who is already holding the customer’s invoice.

Good reads

3. DeepSeek raises a reported $15B, two days after the US removal order

Bloomberg reported on October 22 that DeepSeek has raised about $15 billion — roughly 100 billion yuan — one of the largest funding rounds in China’s history, with battery maker CATL and Tencent among the potential investors. The timing is the story: it came two days after the US ordered federal civilian agencies to remove DeepSeek’s technology from their systems over security concerns. DeepSeek has not confirmed the round, and Bloomberg’s report gives no valuation.

The context is a fast-escalating funding arc: more than $7 billion in June at a reported $50 billion valuation, a further $1.5 billion sought in July at roughly $71 billion, and reports of a possible China IPO as early as next year. The reported use is infrastructure and open models for domestic and international markets.

The same week, DeepSeek shipped the product side of that bet. On October 17 it released V4.1-Flash-1017 with MIT-licensed weights and a 1M-token context: its own scores rise from 74.2 to 77.0 on DeepSWE and from 31.2 to 42.3 on Terminal-Bench 4.0, at under $0.20 per task, alongside Search and Fetch APIs at $0.75 per 1,000 calls. On October 20 it open-sourced DeepInfer, the inference engine it says serves V4.1 Flash, with Huawei Ascend support built in.

Why it matters

A procurement ban that shuts a company out of one of the world’s largest customers would normally make fundraising harder. Here it did not — or, more precisely, Chinese capital treated the US order as a non-event for DeepSeek’s trajectory. The open question is whether $15 billion buys frontier parity or just more of the efficiency DeepSeek is already known for, under export controls that limit its chip supply. Either way, the US–China AI split now has a funding ledger to match its policy one.

Good reads

Elsewhere this week

  • Meta’s Q3: $64.6B revenue, 10M Muse users, and a $135–145B AI bill (Oct 21): revenue rose 26% year over year, but a $9.9 billion teen-safety settlement charge cut operating profit 37%, Q3 capex hit $34.9 billion, free cash flow went negative, and Meta announced a new non-voting Class C share distribution to fund AI spending without diluting voting control. Muse passed 10 million account setups in the US and Canada in about six weeks, with shopping the top use. (AI Parabellum, Particle)
  • Anthropic opens the Claude Duet waitlist (Oct 21–22): a proactive, always-on agent on Opus 5.5 that sends a morning brief, keeps working on its own cloud computer between messages, and asks approval before sending messages, making payments, or deleting anything. Waitlisted research preview for US Pro and Max subscribers; background checks count against existing usage limits. Anthropic is candid about the weak spots: misread emails, missed events between checks, and weaker personal errands than work tasks. (FoneArena, Substack comparison with Grok Bot and OpenAI Dots)
  • OpenAI is reportedly racing to raise $30B ahead of Anthropic’s listing (reported Oct 19–20): Reuters sources describe unusually fast fundraising that could value OpenAI above $1.4 trillion, with SoftBank and MGX in discussions, as both labs court the same investors. (PYMNTS)
  • More Chinese open-weight pressure: MiniMax shipped M3.1 Flash on October 19 at $0.15/$0.60 per million tokens with a 1M-token context, and released its open weights on October 23; Alibaba’s Qwen3.6 27B landed October 21. (MiniMax deep dive, model timeline)
  • Physical AI’s production week: Tesla began early production of production-quality Optimus humanoids at Fremont on October 21 — aiming, Musk said, for 20,000 per week by end of 2027, with outside sales around late 2027 — as quarterly profit fell 43% to $787 million. At GTC Berlin, Nvidia made Isaac Lab 3.0 generally available with a reference humanoid and a 2,000-asset SimReady library. (RobotToday weekly)
  • Banks block third-party agents (Oct 22): Citi and Wells Fargo confirmed restrictions on third-party AI agents logging into customer accounts or transacting directly, allowing only limited read-only access via approved open-banking providers — the consumer-facing sequel to last week’s rogue-agent disclosures. (Daily Caller)
  • UK sets out ~£600M of AI security funding (Oct 23): a biosecurity lab, cyber-defence programmes, and robotics and research funding, with OpenAI and Anthropic offering up to £200 million in model credits to the biosecurity initiative. (Reuters)