An AI Agent Just Got 76x Cheaper to Run, and More AI News

Noted

5–7 things worth understanding, Today by Stackari

· October 12, 2026

Read online | Sign up | Advertise

Your Daily Note,

AI is learning to do more than answer questions. It can navigate websites, take action through text messages and help security teams investigate threats. But the systems behind that progress need more than clever models.

In today’s Noted:

An AI agent just became 76 times cheaper to run
What is Jev, and why is its maker valued at $7.5 billion?
AI companies are dropping NDAs to win over data-centre neighbours
AI agents are helping security teams investigate threats in seconds
Growing use of AI Assistance in daily life

ASANA × OPENAI

this one’s big →

An AI agent just became 76 times cheaper to run

We noted:

Asana has been working to make its browser agent more practical for real business workflows. The surprising part is that the biggest improvement did not come from simply switching to a more powerful model. It came from changing how the agent handles information while browsing. In a study published by OpenAI, Asana says it used GPT-6 Astra through Codex to investigate and optimize a browser agent running on GPT-6.1 Sol. The optimized workflow was 76 times cheaper and five times faster than the original production setup used for comparison.

Finding wasted work. The team discovered that the agent repeatedly sent its growing browsing history, including page text and screenshots, with new model requests. Much of that information had already been processed.
Improving caching. The team changed how the agent reused previous context so repeated information could cost less to process.
Keeping more browsing history. Instead of repeatedly trimming text and removing screenshots, the optimized workflow retained more information and removed older screenshots in batches.
Running experiments with an AI coding agent. Asana’s CTO used GPT-6 Astra in Codex to inspect the code, test different settings and compare results. Work estimated to take one or two months manually took about a week.

Why it matters

When evaluating an AI agent, ask how much a completed task costs, how often it succeeds and how much time it saves. The cheapest model is not always the cheapest solution if it needs several attempts to finish.

TYPESAFE AI

Look out →

What is Jev, and why is its maker valued at $7.5 billion?

We noted:

Jev is a new type of AI model developed by TypeSafe AI. It is not designed primarily to generate paragraphs, images or code. Instead, it produces probabilities that can be used to make decisions, which the company calls “calibrated decisions.” That difference helps explain why the startup has attracted so much attention. TypeSafe says Jev is designed for automation, where a system needs to choose an action or outcome rather than write a natural-language response. According to TechCrunch, TypeSafe raised $870 million at a $7.5 billion valuation just weeks after Jev’s launch. The funding round was led by Andreessen Horowitz, with Sequoia and existing investor DCVC also participating.

Moving beyond text generation. Jev is based on transformer architecture, but TechCrunch reports that it is not a large language model. It produces probabilities rather than text.
Targeting automated decisions. TypeSafe positions the model for tasks where a system must select an outcome or action, rather than generate an explanation.
Promising efficiency. The company says Jev can operate faster and use far fewer tokens than language models for the tasks it targets.
Attracting early enterprise interest. TypeSafe claims that a third of Fortune 500 companies are already using Jev. That is the company’s claim, not an independently verified adoption figure.

Why it matters

Jev points to another possible market: models designed specifically to make decisions inside automated systems. If specialized models can handle particular tasks more efficiently than general-purpose language models, businesses may eventually combine several types of models rather than relying on one model for everything.

AMAZON AND MICROSOFT

AI companies are dropping NDAs to win over data-centre neighbours

Transparency in infrastructure negotiation

We noted:

The AI boom requires more than chips and models. It also needs enormous data centres, electricity, cooling systems and land. As more communities push back against proposed developments, technology companies are being forced to explain their plans more openly. Amazon has said it will no longer use nondisclosure agreements, or NDAs, while negotiating with local governments over data centres. Microsoft made a similar announcement earlier in 2026. An NDA can restrict what parties are allowed to disclose about a negotiation. Removing that restriction could make it easier for local officials to discuss proposed projects with the people who will live near them.

Communities are asking harder questions. Residents and local authorities have raised concerns about electricity demand, water use, property values and the wider effects of large data-centre projects.
Companies are facing greater scrutiny. Some communities have considered temporary pauses on data-centre development while they study the implications.
Amazon is changing part of its negotiating approach. Its announcement was included in a broader effort to argue that data centres can benefit local communities.
Transparency is becoming part of the debate. TechCrunch’s discussion points out that removing NDAs does not automatically settle disagreements over the costs and benefits of a project.

Why it matters

This does not mean data-centre construction will stop or that every negotiation will become completely public. It does mean the industry is under pressure to explain who benefits from these projects, who bears the costs and what commitments companies are prepared to make. If companies want communities to accept more data centres, they may need to demonstrate concrete benefits rather than treating expansion as inevitable.

SOPHOS × OPENAI

updates →

AI agents are helping security teams investigate threats in seconds

openai.com/daybreak/

We noted:

Cybersecurity teams face a growing volume of signals, alerts and suspicious activity. Investigating each case takes time, and a slow response can give attackers more opportunity to act. Sophos says it is using OpenAI models through a programme called Daybreak to help investigate and respond to security incidents. According to OpenAI’s case study, Sophos reduced the average response time for cases handled by these agents from around 38 minutes to 89 seconds.

Gathering evidence for each case. An investigation agent collects customer context, security detections, indicators of compromise and relevant threat intelligence.
Planning and carrying out investigations. The system creates an investigation plan, performs steps and prepares a summary with recommended responses.
Automating some responses. Other agents can handle parts of the response process, rather than leaving every step to a human analyst.
Keeping human oversight. Sophos uses different operating modes, from recommending an action to acting on a customer’s behalf. Potentially destructive actions still require appropriate oversight.

Why it matters

For businesses, AI-assisted security could mean faster triage, quicker investigations and more time for human experts to handle unusual or high-risk incidents. It does not mean every security task can safely be handed over to an AI agent. The quality of the underlying data, the permissions granted to the system and the escalation process all matter. If you are choosing security software, ask what the AI is allowed to do automatically, which actions need approval and how incidents are recorded for later review.

PRODUCTIVE AI ASSIST

More than Chat →

Growing use of AI Assistance in daily life

We noted:

A growing number of AI assistants are moving into messaging apps. Instead of downloading another productivity app or opening a chatbot, users can text an assistant in a familiar conversation and ask it to do something. The latest wave goes beyond answering questions. These assistants aim to remember context, connect to other services and carry out multi-step tasks.

Folk works through iMessage, WhatsApp and Telegram. It can remember personal context, manage reminders, research topics, track flights and handle tasks. It also runs on its own private cloud computer, allowing it to execute code and multi-step workflows.
Instinct is building an always-available assistant that can connect to email, calendars and Google Workspace. Its proposed tasks include planning trips, buying groceries and handling bookings. It has also introduced dedicated email addresses for assistants to use when interacting with services.
Miso focuses on travel. Users can plan trips through iMessage with help from an AI assistant and a dedicated travel team, including support for flight arrangements and personalized recommendations.
Ohai and Ollie focus on family life, helping organize calendars, emails, household chores, school information and reminders.
Poke is a general-purpose assistant that can handle everyday tasks involving calendars, planning, smart-home controls and other services. It was acquired by AI coding company Cognition in 2026.

Why it matters

The chatbot era trained people to visit AI and ask for answers. The agent era is trying to reverse that relationship: tell AI what needs doing, then let it work across the services involved. But giving an assistant access to email, calendars, accounts and purchases raises important questions. Before connecting an AI assistant to your accounts, check its permissions, data policies and approval settings. Start with low-risk tasks before allowing it to send messages, make bookings or spend money.

Product Worth Knowing

H

HireKey

HireKey helps technology job seekers assess role fit, tailor their resumes, explain their experience and track applications. It also offers an interactive AI voice coach that responds to spoken answers and gives feedback for interview practice.

K

kōdlo

kōdlo brings prospect research, lead management, personalized outreach, proposals and sales documents into one workspace. Its Knowledge Hub can use a company’s product information and other documents to help prepare messages and materials.

O

OpenScreen 2

A free, open-source screen-recording tool for Windows, macOS and Linux. It adds planned zooms, smooth cursor movement, webcam layouts and transcript-based editing. Its standout feature is a local MCP server that lets a compatible AI coding agent help edit recordings. You can also edit manually by removing words from the transcript. The project requires no account and adds no watermark.

Everything else in AI

AI research labs are rethinking who gets access to scarce GPUs

The institute says demand for computing capacity was regularly two to three times greater than supply. Its earlier priority-based system encouraged problems such as teams marking too many workloads as high priority or leaving resources tied up in jobs that were not actively doing useful work. In a 30-day period, the institute reports that teams received 98% of the GPU hours they were allocated, while overall cluster occupancy remained at 98%. It also says the new system reduced the 90th-percentile wait time for short debugging jobs from two hours to 30 seconds on the measured workloads.

Apple hires from AI podcast startup Huxe

Apple has struck a deal to make employment offers to certain members of Huxe’s team and license some of its technology. It is not a full acquisition. The arrangement was disclosed in a filing with the European Commission, and the filing does not confirm which employees will join Apple or what products the company plans to build with the technology. Huxe was founded by former Google team members who had worked on NotebookLM’s AI-generated audio features. Its app created personalized audio briefings using information such as a user’s emails, calendar and interests. Users could also generate podcast-style discussions about topics they wanted to explore. Huxe announced in May 2026 that it was shutting down its service. Apple’s newly disclosed deal raises the possibility that some of its technology or expertise could inform future Apple audio experiences, although Apple has not announced specific plans.

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. A new week, a fresh Note!.

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

Stackari · More news · Privacy