Issue 004 · 29 October 2026
Gemini went on defense, NVIDIA gave away speech, and a $200 cartoon beat a $1.5M AI film
Google shipped its strongest model yet for cyber defense, NVIDIA open-sourced a 40-language speech model with a 463,000-hour dataset, GitHub made agentic workflows generally available, and a homemade cartoon outperformed China's $1.5 million AI feature.
This week the frontier moved in two directions at once. Google shipped its strongest model yet, built for cyber defense and long jobs, while OpenAI pushed a cheaper model into the middle of its lineup. Around them, the tools got more open and the box office delivered a blunt lesson about what audiences actually pay for.
1. Gemini 4 Carbon is built to find and fix security holes
Why it matters:
Google announced Gemini 4 Carbon, its strongest model yet, pitched at cybersecurity and long-horizon work. It ranks first on the Artificial Analysis Intelligence Index v5.0 at 59. The security numbers are the story: 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching. It also jumped 89 Elo points to fourth place on WebDev Arena at 1768. Carbon builds on Gemini 4 Argon, which launched two weeks ago. Google is rolling it out carefully: a subset of cyber defenders this week, then more trusted partners through its Fairwind Program, with feedback on guardrails before general developer access. It is part of the U.S. government's voluntary pre-release testing process. Price stays at Argon's introductory level: $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. A token is a chunk of text, roughly three quarters of a word. Google says its own teams have run Carbon for days on multi-step engineering tasks.
How to use this:
If you run security operations or work with a defense contractor, ask about Fairwind access now; this is the model that will be doing vulnerability triage and patch validation. If you are a normal business, the practical move is to ask your security vendor what they are building on Carbon and when it reaches their product. Do not plan around general access yet. When it opens, the first job to give it is the boring one: point it at a staging environment and have it find and patch known vulnerability classes before an attacker does.
The Bottom Line:
Same price as Argon. The saving is security labor: one exploitable flaw found before disclosure can pay for a year of tooling. Google publishes no hours figure, and access is limited for now.
2. NVIDIA open-sourced a 40-language speech model and its training data
Why it matters:
NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model that handles transcription, speaker attribution, translation and audio question answering in one checkpoint. Open-weight means the model files can be downloaded and run yourself. It transcribes 40 languages, translates 31 of them to or from English, and labels speakers in 13. It handles recordings up to an hour long. NVIDIA says it is first among open-weight models on the Hugging Face Open ASR Leaderboard for English, posts the lowest error rate on David AI's 21-language speaker-attribution leaderboard, and ranks first in VoiceArena's 14-system comparison. The release includes Granary v2, a 463,062-hour multilingual audio dataset, plus the curation pipeline and fine-tuning tools. In a medical case study, fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prior model. Quantized versions run locally on RTX hardware.
How to use this:
If you pay per minute for meeting transcription, call notes or subtitles, this is the replacement to test. Run it locally on one sensitive recording first and compare the transcript to your current vendor's. Prompt it for speaker labels and translation in the same pass, and fine-tune on your own jargon if the base model trips on product names. Check the license before you ship it inside a product. The privacy gain is real: audio that never leaves your machine cannot leak from someone else's cloud.
The Bottom Line:
The model is free to download. The saving is the per-minute transcription fee on your own volume, plus the review time you stop spending on bad speaker labels. Measure one week of audio before you cancel anything.
3. GitHub made agentic workflows a standard part of Actions
Why it matters:
GitHub's Agentic Workflows are now generally available. The open-source project from GitHub Next lets you automate reasoning tasks such as issue triage, CI failure analysis, documentation updates and test writing by running coding agents inside GitHub Actions. You define the automation in plain Markdown, and it compiles to standard Actions YAML, so it reuses your existing runner groups and policy constraints. You can use GitHub Copilot or bring your own coding agent. An agent is software that takes a sequence of steps toward a goal instead of answering one question. The safeguards matter: read-only permissions by default, a sandboxed container behind an Agent Workflow Firewall, validated safe outputs, and a threat-detection job that scans proposed changes before they are applied. GitHub also added budget forecasting, model routing and stronger agent isolation. Accenture is quoted saying one team lead saved about 29 hours a week, team velocity more than doubled, AI authored 91% of that team's commits, and production errors fell from roughly 11,000 a day to under 1,000.
How to use this:
Pick one low-risk workflow and run it for two weeks: issue triage, release notes or flaky-test analysis. Keep the default read-only permissions and make a human approve anything that writes. Use the budget forecast to see what the agent costs before you widen access. The Accenture numbers are vendor-provided and come from a mature engineering team, so treat them as a ceiling, not a promise. Measure your own before and after.
The Bottom Line:
The claimed saving is 29 hours a week for one lead. Your number will depend on how much of your backlog is repetitive reasoning. Start with one workflow and count the hours it actually removes.
4. Anthropic now shows companies what their AI spend buys
Why it matters:
Anthropic launched reporting tools for enterprise customers that show what a team's token spend produced. Managers can view a team's work over a date range, what it delivered, what it cost, and where workflows broke down, using internal analytics and employee surveys. Anthropic says the data is not used for model training by default. The reports also recommend how to improve results, and the company added pooled group budgets so teams can manage token spending together. The context is striking: Anthropic recorded an operating profit in the second quarter and 900% net revenue retention, meaning spending from the same customers in the third quarter grew nine times year over year. A Jellyfish survey of 276,000 customers found Claude Code and other Anthropic-based tools account for nearly a third of proposed code changes, with spending up tenfold since the start of the year.
How to use this:
If you are on Claude Enterprise, turn the reports on for one team and compare cost per finished output, not cost per seat. Use the workflow breakdown to find the prompts that burn tokens without producing anything you keep. Set a pooled budget before the next big project, and confirm the no-training default in writing. The point is not to shame heavy users. It is to find the ten percent of usage that produces most of the value.
The Bottom Line:
The 900% retention figure is Anthropic's investor number. The practical lever is cost per finished task. No independent savings figure exists yet, so measure one team for a month.
5. GPT-6.1 Terra puts frontier capability at half the price
Why it matters:
OpenAI released GPT-6.1 Terra on Monday, a fast, efficient model built around detailed visual reasoning and precise computer use. OpenAI says it is a significant step up from GPT-5.6 Terra and excels at the computer-based tasks that fill a workday: drafting proposals, interpreting technical drawings, filling out forms and editing videos. It sits in the middle of OpenAI's lineup, below Sol and Astra on price. The price is the practical part: $1 per million input tokens and $5 per million output tokens, half of GPT-6.1 Sol. Simon Willison notes that Terra's position is the same as Anthropic's Sonnet, which makes the choice between models harder, not easier.
How to use this:
Test Terra on the high-volume, low-glamour work first: invoice extraction, form filling, document comparison, first-pass video edits. Put it head to head with Sol on your own examples and count errors, not benchmark scores. At $1 per million input tokens, roughly 750,000 words costs a dollar, so the volume math changes for back-office work that was too expensive to automate before. Keep a human check on anything that touches money or customers.
The Bottom Line:
Half the input and output price of Sol. A million input tokens costs $1. The saving shows up when you run the same task thousands of times, not once. Measure the error rate before you move a workflow over.
6. AI replicas are out-earning the creators they copy
Why it matters:
Fast Company reports on AI-generated copycat accounts that recreate a creator's videos almost exactly and collect the views and ad money. One account, Duke of Ashcombe, posts videos modeled on Kjetil Krogstad, a dancer and speaker with 1.9 million followers across platforms. Krogstad has posted for five years in a recognizable uniform: dark suit, white shirt, red tie, teacup in hand. The replicas keep the costume and the bit, and the audience often cannot tell. When Krogstad posted side-by-side comparisons asking when inspiration becomes stealing, the videos got over a million views and most commenters sided with him. The legal question is unresolved. Copyright protects specific expression, not a style or a persona, and platform rules have not caught up.
How to use this:
If you publish video or build a personal brand, keep dated originals and register your key work where the law allows. Add visible and invisible provenance, such as watermarks or metadata. Monitor for replica accounts and use platform impersonation reports early, before the copy ranks above you. If you buy influencer or creator content, check whether the account is a real person with a history, not a synthetic replica with a nice backstory. For contracts, add a clause about digital replicas and likeness.
The Bottom Line:
No published revenue figure for the replica account. The cost is lost audience and ad income, which compounds quietly. The legal bill is still unknown, and that uncertainty is the business risk.
7. A $200 cartoon beat China's $1.5 million AI movie
Why it matters:
China's first fully AI-generated feature film, Sanxingdui: Future Past, opened on 23 October on more than 10,000 screens. It was backed by Bona Film Group, made by nearly 100 people over more than two years, and used AI for characters, voices and backgrounds. It has taken about $181,000 at the box office against a $1.5 million budget. Meanwhile Niu Lai, a rough animated film reportedly made for about $200 by a mother-and-son team, has grossed nearly $10 million. The cheap film won on charm and story, not polish. The expensive one had the technology and the screens and still could not fill them.
What you can do with this:
Watch both if you can find them, and study why the cheap one traveled. If you are using AI to make content, spend your effort on the story and the character before the render quality. Test with a real audience early, while the project is still cheap to change. The tool is not the product. The thing people retell is.
The Bottom Line:
The $1.5 million film lost money. The $200 film made nearly $10 million. The return came from taste, not compute. Zero money saved for you, and a useful reminder about where the value sits.
That is the week. The models got stronger and cheaper, the tools got more open, and the box office delivered the reminder that audiences buy stories, not pixels. Pick one workflow, one model, one story, and test it before you scale it.
Which of these would you actually use this week? Reply and tell me. I read every response.