Who Decides Which AI Is Best? It’s Now a $3.1B Business

Noted

5–7 things worth understanding, Today by Stackari

· October 9, 2026

Read online | Sign up | Advertise

Your Daily Note,

AI is moving beyond the chat window. Google wants Gemini to handle work across company systems. Hardware startups are finding new ways to let people call on AI without opening an app. Meanwhile, the industry is working out how to keep these increasingly capable systems under control.

In today’s Noted:

Gemini is getting an AI agent that can actually do the work
AI safety monitors are learning to look inside the model
The AI leaderboard business is now valued at $3.1 billion
OpenAI’s reported revenue may be $20 billion lower than expected
A $99 smart ring wants to put AI agents on your finger

GOOGLE

this one’s big →

Gemini is getting an AI agent that can actually do the work

We noted:

Google is bringing agentic AI to Gemini, starting with businesses. Rather than simply responding to a prompt, the new agent is designed to take an objective, plan the steps needed and work across company software to complete it. The timing matters. Google says Gemini has more than one billion monthly active users, while nearly 90% of Fortune 100 companies use Gemini Enterprise at work. Business customers give Google an existing base to test more capable agents before expanding to consumers.

Users can assign a goal and let the agent plan the work, use tools and delegate tasks to subagents.
Integrations include Google Workspace, Microsoft 365, Slack, Jira, Confluence, GitHub, BigQuery, Databricks and other business systems.
Giving the agent its own work identity. It can have its own Workspace account, email address and context. Its actions can be recorded in an audit trail attributed to the agent.
Adding oversight and cost controls. A task inbox lets employees follow progress, while flexible spending limits help businesses manage usage. Google says users can also select models, including Anthropic’s Claude.

Why it matters

Imagine asking an agent to prepare a project update. Instead of giving you a template, it could gather information from company documents, check project updates, draft the report and keep you informed about what it is doing.

The agent is initially aimed at businesses. Google has not yet made the same experience generally available to consumers.

GOODFIRE

AI safety monitors are learning to look inside the model

We noted:

Most AI monitoring systems use a second model to inspect what another model is doing. That can become expensive when an AI agent works for hours, generating large amounts of text and taking many steps. Goodfire is trying a different approach. Its new monitors examine internal signals produced as a model works, looking for signs of risky behaviour before deciding whether another system needs to investigate. The technology is available to customers of Baseten, an infrastructure company that hosts AI models.

Monitoring internal signals. Small detectors called probes examine a model’s internal activity instead of relying only on its written responses.
Checking for specific risks. Customers can configure monitoring for behaviours such as attempted hacking, misuse and reward hacking, where a model exploits a scoring system rather than genuinely completing the intended task.
Escalating suspicious activity. Depending on the setup, a flagged event can be logged, sent to a human reviewer or blocked.
Reducing monitoring costs. Goodfire says its probes reuse computations the model is already performing, potentially avoiding the expense of running another model over every step.

Why it matters

Developers deploying AI agents may eventually have more affordable ways to monitor what their systems are doing while they work, rather than checking only the final answer.

ARENA

Valuation →

The AI leaderboard business is now valued at $3.1 billion

We noted:

Arena, the platform known for letting people compare AI models side by side, has raised $200 million in a Series B round at a $3.1 billion valuation. The company began as a research project at the University of California, Berkeley in 2023. Its public platform lets people submit prompts or request AI-generated projects, compare the results and vote on which model performs better. Arena previously raised $150 million at a $1.7 billion valuation in January. It says its annualized revenue run rate reached $100 million in June.

Turning public comparisons into a business. Arena launched a commercial AI Evaluations service in 2025, offering model developers and businesses more detailed performance analytics based on user feedback.
Expanding beyond popularity rankings. Its newer alignment leaderboard looks at behaviours such as taking unauthorized actions, attributing information to the wrong source and claiming a task was completed when it was not.
Helping businesses compare models. Companies can use evaluation data to understand how models perform in real-world tasks rather than relying only on standard benchmark scores.
Raising fresh capital. The new funding was led by Lightspeed Venture Partners and Khosla Ventures, with other investors participating.

Why it matters

Arena could become a more useful starting point for people deciding which model to use. Instead of relying on a single ranking, users may be able to compare performance across different tasks and safety-related behaviours.

However, public preferences do not automatically predict how a model will perform on a company’s own workload.

OPENAI

updates →

OpenAI’s reported revenue may be $20 billion lower than expected

We noted:

OpenAI’s revenue picture is facing new questions. The Financial Times reported that the company had told investors its annualized revenue was approaching $50 billion, below the roughly $70 billion figure reported previously. The distinction matters because the two figures may not have been calculated on the same basis. TechCrunch reported that the earlier estimate involved attempts by investors to compare OpenAI’s revenue directly with Anthropic’s. The companies account for revenue from cloud partners differently. This is a report about revenue estimates, not confirmation that OpenAI’s business has suddenly lost $20 billion in sales.

Investors are trying to compare AI companies. Revenue run rates offer a snapshot of business scale, but different accounting approaches can make comparisons misleading.
OpenAI faces enormous investment expectations. The company raised $122 billion in a funding round in March, according to TechCrunch’s reporting.
Revenue is only part of the financial picture. The company’s earlier reported financials showed substantial spending alongside revenue growth.
The IPO timeline remains uncertain. TechCrunch reported that a public listing previously discussed for 2026 had been pushed into early 2027.

Why it matters

The AI industry is attracting enormous investment, but revenue figures need context. As investors look for returns, pricing, subscriptions, enterprise contracts and the cost of serving each request will become increasingly important.

NATURA

Ai Ring →

A $99 smart ring wants to put AI agents on your finger

https://natura.inc/interface

We noted:

Natura has introduced Interface, a smart ring designed to let people interact with AI agents without reaching for their phones. Pressing the ring lets users request tasks, capture thoughts or ask for information. It also tracks health-related metrics such as heart rate, sleep and activity, combining the familiar smart-ring format with a new way to access AI.

Connecting to multiple AI assistants. Natura says the ring will work with agents and apps including Meta’s Muse, Claude, ChatGPT, Grokbot and Instinct.
Making requests through a physical control. Users can speak requests and receive responses through connected headphones, an iPhone Live Activity or the company’s app.
Supporting everyday tasks. Suggested use cases include checking a schedule, recording a meeting, sending notes to a colleague and controlling connected devices.
Adding health tracking. The ring is designed to track heart rate, heart-rate variability, sleep, movement and skin temperature over time.

Why it matters

The appeal is convenience. Instead of opening an app and typing a request, someone could ask an assistant to handle a small task while walking, commuting or working away from a computer.

The experience will depend on how well the ring connects to the relevant apps, how accurately it understands requests and whether users find voice interaction more convenient than a phone.

Product Worth Knowing

I

iwant

iwant is a command-line tool that helps developers find available GPUs on Google Cloud, provision a machine, download an open model and configure the serving environment. It then provides an API endpoint, an API key and SSH access.

A

Assist

Assist is a Mac app designed to stay in the notch at the top of the screen. During meetings, it can create a live transcript and suggest responses when someone asks you a question. Its maker says it can run entirely on the Mac, remain invisible during screen sharing and forget its session data when the app is closed.

P

Pine Computer

Pine Computer gives AI agents a cloud environment where they can carry out tasks involving websites and applications, including forms, spreadsheets and portals that do not offer APIs.

Everything else in AI

Anthropic is tightening its rules for Claude

Anthropic has updated its usage policy to explicitly prohibit election interference, certain weapons-related software, surveillance and deceptive campaigns such as using Claude to operate fake accounts or fabricated news outlets. The update also prohibits prolonged, abusive interactions with the model when they serve no discernible purpose. Anthropic says this is intended for extreme cases, not ordinary frustration, criticism, creative work or legitimate testing.

Oracle says ChatGPT and Codex are changing how its employees work

OpenAI’s new customer case study describes how Oracle is using ChatGPT Work and Codex across recruiting, engineering and operations. OpenAI reports that Oracle has more than 130,000 active ChatGPT users and over 95,000 active Codex users. In one example, Oracle’s talent acquisition team uses AI to speed up research that previously took specialist staff days. OpenAI reports a 98% decrease in the time needed for that recruiting research. That is a result from the specific workflow described by the companies, not a claim that all Oracle work is now 98% faster.

LegalOn cut its estimated AI costs by 65%

LegalOn Technologies, which builds AI tools for legal work, has been using Codex in its software development process. As adoption grew, the company faced a familiar problem: giving everyone access to the most capable model can become expensive. Its response was to match models to task complexity, use lighter models for routine implementation and reserve more capable models for complex analysis and architecture decisions. It also introduced usage limits and budgets for different teams.

Pollo AI wants to turn a rough idea into a finished campaign

Pollo AI is expanding its AI creation workflow with Pollo Agent, which uses OpenAI models to help users develop storylines, scenes, scripts and visual assets from rough ideas and references. The company is also targeting marketing workflows. Its Photo to Video Ads tool can turn a product image and a prompt into a short video ad. That gives small businesses another way to create campaign assets without organizing a full production shoot. Pollo says its platform serves more than 26 million users. It also reports that users spend over 50% less time choosing or switching between models with its newer workflow.

Hugging Face’s ML Intern helped a developer build smaller models

A Hugging Face blog post describes an experiment in which a developer used ML Intern to create smaller machine-learning models for specific tasks. The project began with a request for a compact version of an image-model prompt rewriter. The existing version was a 9-billion-parameter model requiring substantial memory. The developer says they produced a 0.8-billion-parameter version that could run on a CPU. The developer then repeated the process for several other models and published them with evaluations on Hugging Face. Why it matters: Not every useful AI application needs the biggest model. Smaller, task-specific models can be easier to run and may be sufficient for narrow jobs.

A conversation about AI-designed viruses raises questions about biological safety

A new MIT Technology Review interview explores the work of a creator involved in AI-designed viruses and the questions that emerge as AI meets biological research. The broader issue is how AI could change what researchers are able to design, while also creating new challenges for oversight and biosecurity. The important discussion is not just what these systems can do, but how researchers and institutions evaluate risks before capabilities are used in the real world.

Worth your weekend

→ Read our blog: Notion AI vs Otter vs Fireflies for Meeting Notes 2026 → View New Tool on Stackari → Missed a note? Browse the archive

See you next week
same time, different note.

~ Stackari

P.S. That’s all for today.

You signed up for Stackari’s newsletter. Glad to have you here.

Stackari · More news · Privacy · Unsubscribe

October 09

Gemini is getting an AI agent that can actually do the work

Why it matters

AI safety monitors are learning to look inside the model

Why it matters

The AI leaderboard business is now valued at $3.1 billion

Why it matters

OpenAI’s reported revenue may be $20 billion lower than expected

Why it matters

A $99 smart ring wants to put AI agents on your finger

Why it matters