Today’s three notable AI updates focus on OpenAI, Meta, and Anthropic.
Viewed together, they reveal a bigger shift:
AI companies are no longer satisfied with just having you ask a chatbot questions.
The next step is:
Let AI stay embedded in workflows and keep working on its own.
Once this starts happening, businesses will face two new questions:
How capable is it?
And how much will the compute power behind these Agents cost?
1. OpenAI DevDay: Agents evolve from one-off tasks to persistent workers
On September 29, OpenAI held DevDay 2026 in San Francisco.
The most notable new direction is called:
Dots.
Reuters describes this as an "Always-on Agent."
This means it’s not about asking a question and the agent completing a single task once.
Rather, it stays active within a work context for an extended period, continuously tracking changes.
For example:
- Updating when a sales proposal changes
- Moving on to the next step as a project status updates
- Continuously building demos based on existing data
- Completing tasks across different tools
It can also integrate with work environments like Slack and Teams.
This contrasts with traditional chat AI:
Previously:
People thought of a task → found AI → gave instructions.
Now:
The AI is already embedded within the work.
When things change, it proactively continues working.
Why is being “persistent” more important than just having a stronger model?
The biggest real-world waste isn’t that AI can’t write an email.
It’s that:
Yesterday’s email,
Today’s meetings,
Tomorrow’s deadlines,
CRM data,
New Slack messages,
all require humans to reconnect the dots.
If an Agent can maintain context and keep track of the same task,
its value is not just:
“Help me write this section.”
but instead:
“Keep an eye on this for me.”
This is where Agents could truly transform office work.
OpenAI takes another step toward enterprise adoption
OpenAI’s latest enterprise info also introduced:
OpenAI Marketplace.
Currently in Beta.
Qualified enterprise customers can allocate parts of their contracted OpenAI spending toward eligible partner software.
This approach is very similar to recent moves by Claude Marketplace.
AI companies don’t only want to sell models.
They now want to be:
The first point of entry when enterprises buy other AI software.
OpenAI has also launched GPT-6.1 Sol to enhance agentic coding, computer use, and professional work.
So the real takeaway from DevDay isn’t just new features.
It’s that OpenAI is aligning models, Codex, work tools, agents, and partner software into a unified enterprise work platform.
2. Meta Muse: Now targeting “tasks bosses don’t have time for daily”
On September 29, Meta pushed Muse forward with a new version:
Muse for Small Business.
The focus is very direct.
It’s not designed for large enterprise IT departments.
But for:
Stores, small businesses, studios, personal brands.
Meta’s approach is practical:
Small businesses aren’t short on work.
They have:
One person juggling 10 different roles.
Which tools can Muse connect to now?
Officially announced connectors include:
- Asana
- Box
- Canva
- Dropbox
- Figma
- QuickBooks
- Klaviyo
- Notion
- Shopify
- Slack
- Stripe
- Zoom
Along with Meta’s own business tools like Facebook Page, Instagram Business Account, and Meta Ads.
Once connected, Muse can understand:
What you sell,
How your brand communicates,
Common customer questions,
Sales performance,
Advertising effectiveness.
This is very different from typical chatbots.
Users don’t have to daily:
Download reports → paste into AI → explain context again.
The agent stays next to those data sources.
Realistic examples Meta provides
For example, a boss might ask:
How did sales, campaigns, and socials perform this year?
Help me organize next year’s growth plan.
Or:
I’m overwhelmed today; any emails, calendar events, or news I should prioritize?
Or:
Which ads are effective?
How should next week’s campaign adjust?
Even:
How did our finances do this month?
Any unusual expenses?
This isn’t just chatting.
It’s like:
A small company gaining a digital assistant that can check many systems.
But Meta deliberately sets one boundary
The company explicitly states:
Nothing publishes, sends, or spends without your approval.
Meaning:
Muse can check,
Organize,
Analyze,
Draft,
but for:
Publishing content,
Actually sending messages,
Actually spending money,
Human approval is still required.
This is more important than how many apps Muse can connect with.
Once agents enter companies, the key question becomes:
Which steps can they do autonomously?
And:
Which steps must pause for human approval?
Setting this boundary is crucial for small businesses adopting agents.
3. Anthropic’s other side: the huge compute cost behind AI’s growth
The third update isn’t about new features.
But it may reveal more about stress facing frontier AI companies ahead.
Reuters reported on September 29 after reviewing a confidential Anthropic IPO document.
The document shows:
Anthropic’s committed AI infrastructure spend over the next decade
is at least:
$518 billion.
More importantly:
Reuters notes about 80% of these commitments are binding,
meaning even if Anthropic doesn’t fully use the capacity, many costs can’t easily be canceled.
Important to clarify:
This is not a press release from Anthropic announcing a $518 billion investment plan.
The information comes from confidential IPO filings reviewed by Reuters.
So more accurately:
Anthropic has disclosed massive long-term infrastructure commitments in its IPO documents.
Why do AI companies lock compute early?
Because frontier AI companies face a paradox.
If usage of models and agents spikes in the future,
but they don’t secure now:
GPUs,
Data centers,
Cloud capacity,
Networking,
Power—
They might not get enough compute years later.
Conversely:
If they sign huge long-term contracts now,
but AI revenue underperforms,
these capacities become costly liabilities.
So AI companies aren’t just betting on:
Whether models get stronger.
They’re betting on:
How much AI the world will need in five or ten years.
This is why “AI is expensive” can’t be judged by token price alone
Recently, we’ve seen headlines about:
GPT getting cheaper,
Claude making tasks cheaper,
Cache becoming cheaper.
From a user perspective,
AI seemed to be getting less costly.
But the flip side is:
AI companies are signing hundreds of billions in infrastructure commitments.
The two perspectives don’t conflict.
Because AI companies want to:
Lower unit cost first, then scale usage massively.
If an agent no longer answers just 10 questions a day,
but rather all day:
Researching data,
Checking email,
Running code,
Generating reports,
Operating tools,
Updating projects—
Even if the cost per task is cheap, overall compute demand can explode.
Today’s three updates are connected
OpenAI Dots answers:
Can AI keep working continuously?
Meta Muse for Small Business answers:
Can AI truly integrate into everyday business?
Anthropic’s compute commitments answer:
If so, how much infrastructure is required?
This is the real shift to remember today.
The next phase of AI isn’t:
Everyone asking a chatbot a few more questions daily.
But possibly:
Everyone getting an agent working for them all day long.
If this model takes off,
it won’t just be AI usage surging.
It will also dramatically increase:
Permissions,
Security,
Costs,
Procurement,
and entire data center infrastructure.
