2026-09-26.md

Daily AI News — September 26, 2026

Saturday launch edition · 2026-09-26
Research cutoff: September 26, 2026, 1:56 PM Eastern. This briefing covers verified developments from September 24–25; the older items are labeled below.

AI is moving deeper into the workday: Microsoft is putting documents, coding and persistent agents in one place, Cognition reports a major commercial milestone, and Exa is pushing research agents toward more exhaustive searches. The interesting question is becoming what these systems can finish reliably, rather than how impressive one answer looks.

The three things to know

• The workspace is becoming the product. Microsoft's new Copilot combines familiar Office work with tools for building and delegating.
• Coding agents have a business signal to watch. Cognition says its annualized revenue run rate has passed $1 billion; that is a company-reported pace, not a year's realized revenue.
• The infrastructure story includes CPUs. A large Anthropic–Akamai agreement puts the computing around AI models in the spotlight.

1. Microsoft introduces a broader Copilot: Home, Code and Autopilot

Announced September 25 · Enterprise AI

Microsoft outlined three parts of its new Copilot experience. Home combines Chat and Cowork with Word, Excel and PowerPoint inside the experience. Code lets users build solutions using technology shared with GitHub Copilot. Autopilot is intended to provide a persistent, proactive personal agent.

Availability is phased: Microsoft says Home and Code begin rolling out through its Frontier program in the coming weeks, while Autopilot expands its private preview at month-end. These are rollout plans, not evidence that every customer can use every feature today. Source: Microsoft announcement.

Why it matters — our read: Distribution through existing documents and work habits could matter as much as model quality. For teams evaluating agents, the practical test is whether a task stays coherent across files, tools and collaborators.

2. Cognition says it has crossed a $1 billion revenue run rate

Announced September 25 · Coding agents

Cognition reports that its annualized revenue run rate has exceeded $1 billion, less than two years after Devin became generally available. Its announcement names GE Aerospace, Rivian, Rohlik and Exa among the teams using Devin.

The figure is self-reported. An annualized run rate extrapolates a current pace; it does not establish $1 billion of recognized annual revenue, profitability, or the durability of growth. The short announcement offers limited financial detail. Source: Cognition announcement.

Why it matters — our read: This is a substantial demand signal for software agents. The next useful evidence is repeatable customer value: work completed, time saved, review burden and retention.

3. Exa ships Agent Ultra for deeper research

Announced September 25 · Search and research

Exa introduced Agent Ultra as the highest-effort setting for Exa Agent, available through its API tooling. The company describes a system that divides research into subtasks and coordinates agents across large numbers of sources, with particular emphasis on exhaustive lists and entity research.

Exa also published benchmark comparisons against other research systems. Those are vendor evaluations, with methodology and harness choices described in the post; they should not be read as a universal ranking of research quality. Source: Exa announcement and evaluation details.

Why it matters — our read: Finding ten plausible answers is different from finding nearly every qualifying answer. Research products are increasingly competing on coverage and evidence, not just prose quality.

Still worth your attention from September 24

4. Anthropic commits $11.6 billion to Akamai infrastructure over seven years

Announced September 24 · Infrastructure · Carryover

Akamai announced an $11.6 billion contractual commitment from Anthropic over seven years, focused on growing CPU workloads. The relationship could expand by another $9 billion, bringing the potential total to approximately $20 billion. That expansion is an option, not an already-committed $20 billion deal. Source: Akamai announcement.

Why it matters — our read: AI capacity is broader than accelerators. Serving products and operating agent workflows also creates demand for conventional compute and distributed infrastructure. This agreement is a useful counterweight to an exclusively GPU-centered picture of the market.

5. GitHub makes the case for interfaces beyond chat

Published September 24 · Developer tools · Practical read

A GitHub walkthrough explains canvases in the GitHub Copilot app: small applications that communicate with an agent, providing a more concrete interface for tasks such as inspecting data or steering development workflows. The post demonstrates examples rather than announcing a new model or claiming a new feature launch that day. Source: GitHub's canvas walkthrough.

Why it matters — our read: A chat transcript is a poor control panel for many recurring tasks. Purpose-built interfaces can make progress and human decisions easier to see. That is a product-design direction worth testing, even when the underlying agent stays the same.

What to watch next

• How Microsoft's announced capabilities behave as access expands beyond previews.
• Whether coding-agent customers report durable gains after accounting for review and correction.
• Independent evaluations of research coverage, citation accuracy and cost per completed task.

Sources and method

This edition uses the five primary sources linked beside their stories. Announcements, company metrics and vendor benchmark claims are attributed to their publishers. “Our read” is editorial interpretation. This first issue includes two clearly labeled September 24 carryovers; later editions will avoid repeating stories without a material update.

Published in Daily AI News on Stuff. A small briefing for a very loud industry.

Open with JavaScript for the full viewer.