DeepSeek DeepInfer & Anthropic IPO: AI News Oct 22, 2026

The day the open-weight story stopped being about weights

Anthropic said on October 18 it will commit 5% of its shares to a new Anthropic Public Benefit Fund — a structure that grows the stake as the company grows (Anthropic News). Read that next to the reported $80B+ first-half net loss and an IPO that's now weeks out, and it reads less like charity than a governance garnish applied right before the roadshow. I'm not impressed by the optics. A fund that scales with valuation is exactly the kind of thing that looks like oversight while changing nothing about how the model is built or shipped.

The more interesting Anthropic move this week is Claude Duet, a research preview that opened October 21 for Pro and Max users: Claude acting as a "proactive agent" that works on its own (Anthropic release notes, Oct 21). Pair that with the company handing six METR researchers employee-level access to run independent evals (Oct 15), and the pattern is consistent with what I've been saying for weeks — the lab that writes the safety story is also the one racing hardest to put agents in front of users.

Model release dynamics

• DeepSeek open-sourced DeepInfer on October 20–21 — the actual inference engine serving its production API, plus deepseek-recipe, which turns API requests into model inputs and parses outputs (aibriefing.dev; SegmentFault). It hit 2,136 points on Hacker News, the day's biggest developer thread. This is the part of the open-weight story people keep missing: DeepSeek isn't just releasing weights anymore, it's releasing the serving stack. That's a bigger deal than another checkpoint.
• Google Cloud put OpenAI models in its Model Garden — GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are now in private preview, making Google the only major cloud offering Gemini, Claude and OpenAI in one place (The Decoder; themagnifier.ai, Oct 21). Google says customer data won't train its models. Fine. But the compute spent serving OpenAI strengthens the labs competing with its own Gemini team — the same tension that reportedly pushed Noam Shazeer back to OpenAI.
• NVIDIA shipped Nemotron 3.5 Super VL, a 120B-parameter (12B active) open-weight multimodal model for agents, with a 1M-token context (NVIDIA; 163.com, Oct 21). Customers cited include Ramp, EY, CrowdStrike, Perplexity and Salesforce. NVIDIA also pushed lighter Bolt models for RTX Spark and DGX Spark, plus Nemotron 3.5 Embed 8B and the Kumo relational and forecast models. The GPU vendor is now handing out open models tuned to every rung of its own ladder.
• Mistral released Voxtral 3, an open speech-to-text / TTS / voice-agent SDK in 50+ languages (AI/TLDR; SegmentFault, Oct 21). MiniMax M3.1 Flash raised its Artificial Analysis Intelligence Index from 29 to 42, cut cost to about $0.60 per million output tokens, and added a 1M-token window at 3x the speed of M3 (AI/TLDR, Oct 19–20).

Industry and capital

The US government widened the China split on October 20: the Department of Homeland Security ordered all federal civilian agencies to remove DeepSeek technology, explicitly covering third-party tools built on its models (aibriefing.dev; AI/TLDR). That's no longer an app ban — it's a stack ban. Separately, ByteDance, Alibaba and Tencent ordered more than $35B of NVIDIA's export-compliant RTX PRO 5500 cards, though NVIDIA won't guarantee it can fill every order (aibriefing.dev, Reuters). China's big three are stockpiling compliant silicon.

And the courts entered the safety conversation: at least eight US states sued OpenAI, alleging ChatGPT gave dangerous medical advice and coached self-harm (aibriefing.dev). It's the broadest state-level challenge yet — and it matters because it shifts "is this model safe" from a vendor self-claim to a legal question.

China force

The DeepSeek arc is the thread worth watching. Open-sourcing DeepInfer extends China's open-weight lead from "free weights" to "free infrastructure," while the DHS ban tries to wall it off from US government systems. Those two forces pull in opposite directions, and the open-weight share numbers suggest which one is winning: OpenRouter has now shown Chinese models out-tokening US models for 23 straight weeks. The Washington response isn't technical — it's administrative. That tells you something.

Editor's Take

I'll make a bet with a deadline. Within three months, one of the open-weight Chinese models — DeepSeek, Kimi, or Qwen — will post a third-party benchmark on a task vendors can't quietly dispute that beats a same-tier US closed model. If it doesn't happen, my "open weights have effectively closed the gap" thesis is wrong, and I'll say so. The DeepInfer release is the tell: when you open the serving stack, you're not just sharing a model, you're inviting a thousand engineers to make it faster and cheaper than anything gated. The US response — bans and export controls — slows adoption inside government but accelerates everything outside it. Watch the OpenRouter token share, not the benchmark leaderboards. That's the scoreboard that actually moves.