AI Weekly
AI News ·
This week in AI — Gemini 4 Carbon, GPT-6.1 Terra, and Anthropic files to go public
Official announcements only, Oct 24–30, 2026. Google's new top model goes to cyber defenders first, OpenAI ships computer use for everyone and rewrites its Model Spec, Genie 4 adds sound, and Anthropic files an S-1.
From official announcements, Oct 24–30, 2026: Google released Gemini 4 Carbon to a subset of cyber defenders first, with launch API pricing of $2/$10 per 1M input/output tokens. OpenAI introduced GPT-6.1 Terra, pitched as fast computer use for everyone, and an updated Model Spec built on five core values. Google DeepMind announced Genie 4, a world model that generates audio natively. Anthropic filed an S-1 for a proposed Nasdaq IPO under the symbol ANT.
We only cover what companies published on their own blogs and newsrooms, and every item links to the original. Benchmark claims below are the companies' own numbers, not independent tests.
OpenAI's article pages blocked our reader this week, so for OpenAI items we only use the summary in OpenAI's official RSS feed. Nothing beyond that.
1. Google's Gemini 4 Carbon goes to cyber defenders first (Oct 28)
Two weeks after Argon, Google announced Gemini 4 Carbon, which it calls its strongest model, with a focus on cybersecurity and long, multi-step tasks.
- Who gets it: Starting this week, only a subset of cyber defenders. Access then widens to trusted partners through Google's Fairwind Program. Developers, businesses and consumers get it "soon," with no date.
- Launch pricing: $2 per 1M input tokens and $10 per 1M output tokens, with cached input 95% off. After the introductory period, $4 and $20.
- Safety claims: 0.0% attack success on the Gray Swan prompt-injection benchmark (Argon: 0.7%), and high-severity misalignment alerts in internal coding work down almost 80%.
- The other claims: #1 on the Artificial Analysis Intelligence Index (59), 88.2% on SEC-bench Pro for finding vulnerabilities — all reported by Google.
What it means for you: You can't use it in the Gemini app yet. The pattern is now clear: models with strong hacking skills go to defenders before everyone else.
Source: Gemini 4 Carbon — Google
2. OpenAI introduces GPT-6.1 Terra (Oct 26)
OpenAI's feed describes GPT-6.1 Terra in one line: "Bringing fast and capable computer use to everyone." Computer use means the model looks at a screen and clicks and types for you.
- We couldn't open the article, so we can't confirm plans, pricing or rollout dates.
What it means for you: Handing an AI the boring clicks — forms, dashboards, settings pages — is getting closer. When it touches your accounts or payments, watch what it clicks.
Source: Introducing GPT-6.1 Terra — OpenAI
3. OpenAI rewrites its Model Spec around five values (Oct 29)
The Model Spec is OpenAI's public rulebook for how its models should behave. Per the feed, the new version sets out five core values meant to steer more capable AI toward sound judgment, human agency, and faithfulness to human intent.
- The same day, OpenAI also published a post on how it detects misalignment in frontier models.
- We couldn't read the full list of five values.
What it means for you: When ChatGPT refuses something or pushes back, this document is where the reasoning is supposed to come from. Worth a skim if you rely on it daily.
Source: An updated Model Spec — OpenAI · How we detect frontier model misalignment — OpenAI
4. Genie 4 builds worlds with sound (Oct 27)
Google DeepMind's Genie 4 is a world model: it generates an interactive world you can move through in real time. This version generates audio along with the world, which Google says is a first for its real-time world models.
- Who gets it: A research preview for trusted testers now. "Coming soon" to Google AI Ultra subscribers in Project Genie and to early-access partners on the Gemini API and AI Studio. There's a waitlist.
- Labeling: All video and audio from Genie 4 will carry a SynthID watermark.
- Limits Google admits: objects can shift or drift over time, and legible text, spatial accuracy and long-term consistency are still being worked on.
What it means for you: Unless you're on Ultra, you'll be watching demos for now. As generated video and sound get more convincing, watermarks like SynthID matter more.
Source: Genie 4 — Google DeepMind
Also this week
- Anthropic files for an IPO (Oct 26): Anthropic filed a Form S-1 with the SEC for a proposed IPO on the Nasdaq Global Select Market under the symbol ANT. Share count and price range haven't been set, and the registration isn't effective yet. Source — Anthropic
- Mistral Large 4 open weights (Oct 29): Mistral's new multimodal model is on Hugging Face with a 1M-token context window, under a license that limits model-as-a-service use. The 2-bit quants run on a single Apple Silicon Ultra machine with 512 GB of RAM. Benchmarks are Mistral's own. Source — Hugging Face
The takeaway
The strongest models now ship to defenders first, AI is learning to drive your screen, and labs are publishing the rules their models are supposed to follow. The more you hand off to AI, the more it matters what you check yourself — find out your AI type in 3 minutes.
Sources
- Gemini 4 Carbon: our next frontier model — Google (2026-10-28)
- Introducing GPT-6.1 Terra — OpenAI (2026-10-26)
- An updated Model Spec for more capable and beneficial AI — OpenAI (2026-10-29)
- How we detect frontier model misalignment — OpenAI (2026-10-29)
- Genie 4: A new generation of world simulation — Google DeepMind (2026-10-27)
- Anthropic files S-1 with the SEC for proposed initial public offering — Anthropic (2026-10-26)
- Welcome Le Chonk: Mistral Large 4 on Hugging Face — Hugging Face (2026-10-29)
- OpenAI News RSS — OpenAI
Written with AI (Claude) and checked against the linked sources.