What Changed in AI This Week:OpenAI Dots, GPT 6.1 Sol, Claude Sonnet 5.5, Microsoft Copilot and Meta Enterprise AI
What Changed in AI This Week: OpenAI Dots, GPT 6.1 Sol, Claude Sonnet 5.5, Microsoft Copilot and Meta Enterprise AI
Persistent agents, cheaper models, more choice inside Copilot and a new enterprise AI player. Here’s what this week’s biggest announcements mean for businesses.
There’s a lot to unpack this week. OpenAI made more than 20 announcements at DevDay, Anthropic launched a faster (and cheaper) everyday model, Microsoft added more model choices to Copilot, and Meta made a much bigger play for enterprise AI.
OpenAI turned ChatGPT into a shared workspace for people and agents, Anthropic made its everyday model faster and cheaper, Microsoft put both into Copilot, and Meta made a serious move into enterprise AI.
But one of the more interesting OpenAI stories actually happened away from the launch stage. The company held back its planned GPT-6.1 Astra release after safety testing raised concerns around how reliably it stayed within the scope and authorisation given by users.
That’s worth keeping in mind alongside all the new releases, because as businesses are getting access to AI that can do more, work for longer and connect to more of the tools they already use. The question you’re ultimately then left with is ‘how do I make use of that capability without giving AI more access or autonomy than it needs?’
We’ll be digging into exactly that in AI in Business: Part 2, including integrations, governance, data and what businesses should think about before putting AI into real workflows.
Headlines at a glance
OpenAI Dots and ChatGPT Space
Persistent agents can work in the background, while teams can share project context, pages, tasks and connected tools inside ChatGPT.
GPT 6.1 Sol
OpenAI says it delivers near Astra performance at one fifth of Astra’s standard input and output token prices.
Claude Sonnet 5.5
Anthropic says it is more than 30% faster than Sonnet 5, and costs up to 30% less per task, despite unchanged headline token prices.
Microsoft Copilot
GPT 6.1 Sol and Claude Sonnet 5.5 are rolling into Cowork, Copilot Studio and Microsoft 365 apps under different billing models.
Meta Enterprise AI
Meta plans to bring Muse, Meta Business Agent, APIs and coding tools to businesses under a new enterprise division.
Barclays and Claude
Claude already supports 16,000 colleagues and processes roughly 120,000 Global Markets emails a day, with wider developer adoption planned.
OpenAI Dots and ChatGPT Space move AI from chat to ongoing work
OpenAI’s DevDay headline was Dots: always-on agents powered by GPT-6 Astra, with their own cloud computer and access to connected applications. If you missed it, a dot can work across projects in the background, carry context between ChatGPT, Slack and Microsoft Teams, and ask for approval when an action affects an account or shares information. OpenAI says proactive background research is restricted to read-only access, while administrators and users can set rules that allow, block, or require approval for specific actions.
The obvious thing to be mindful of here is in how it changes the risk levels you’re faced with if you rely on AI to work. A chatbot/LLM waits for you to prompt it before delivering output, but a persistent agent monitors, remembers and acts.
You can put forward that the business case is stronger with something like Dots - or agents - rather than just a usual chatbot-type AI, because they don’t need you to hold their hand to ensure productivity, but the flip side of that coin is that mistakes can travel infinitely further before someone notices. OpenAI itself says consequential work should still be reviewed, for what it’s worth - so you don’t just have to take my word for it.
Alongside Dots, OpenAI launched ChatGPT Space, Pages and shared Team Tasks. Space gives teams and agents shared project context; Pages creates editable documents for human and agent collaboration; Team Tasks can run recurring work on a schedule or respond to events such as a new email or Slack message. Business and Enterprise users can also mention ChatGPT directly in Slack or Teams, including colleagues who don’t hold an individual ChatGPT licence. That’s handy.
Before enabling persistent agents, businesses should define approved data sources, permitted actions, named owners and the points where a person must sign off - regardless of if you’re Team ChatGPT, Team Microsoft, or Team Anthropic for all your AI needs.
GPT 6.1 Sol makes advanced AI workflows cheaper to repeat
OpenAI also released GPT-6.1 Sol for coding, computer use and professional work. Its standard API price is $2 per million input tokens and $10 per million output tokens, compared with $10 and $50, respectively, for GPT-6 Astra. OpenAI describes the model as delivering near-Astra performance, with a context window of just over one million tokens.
For those inclined - or if you just want to dive into the nitty gritty - the full specifications and pricing are set out on the official GPT 6.1 Sol model page.
DevDay also added computer use to the Agents API, introduced a limited preview of the Decisions API for constrained classification and routing, and announced Private Intelligence features, which are intended to give organisations tighter controls over sensitive data. OpenAI and AWS also launched Bedrock Managed Agents powered by OpenAI, allowing agents to run inside AWS and connect to AWS resources.
The commercial point is simple. A model that’s slightly less capable but materially cheaper can be the better business tool when a workflow runs hundreds or thousands of times. Pilot it with your real inputs, measure success and cost per completed task, and reserve the most expensive model for work where it genuinely can - and will - change the outcome.
Claude Sonnet 5.5 raises the everyday cost performance bar
Anthropic’s answer (conveniently) arrived the day before DevDay. Claude Sonnet 5.5 is positioned for well-scoped everyday work, bug fixing and polished documents, slides and spreadsheets. According to Anthropic, it produces output more than 30% faster than Sonnet 5 and costs up to 30% less per task because it uses fewer tokens, although the list price remains $2 per million input tokens and $10 per million output tokens.
Crucially, though, there’s an important caveat in Anthropic’s own positioning: Opus 5.5 remains stronger for complex, open-ended work that needs sustained judgment, so the sensible buying question isn’t sitting and asking ‘well, which model benchmarks best?’, when it requires an understanding of which one meets the quality bar for a particular workflow at the lowest reliable cost.
Sonnet 5.5 is available through Anthropic, AWS, Google Cloud and Microsoft Azure, with zero data retention available. Anthropic has also added stronger cyber safeguards because the model’s cybersecurity capability is materially higher than its predecessor’s. For most businesses, that further reinforces the case for approved enterprise access rather than unmanaged personal accounts.
Microsoft Copilot turns model choice into a managed feature
Microsoft are moving quickly. GPT-6.1 Sol and Claude Sonnet 5.5 began rolling out in Microsoft Copilot on the 30th of September. Both are available first in Copilot Cowork and Copilot Studio with usage-based billing, then across Word, Excel, PowerPoint and Chat as part of the user subscription licence, subject to limits, access and region.
What that means is that the same model can sit within fixed-price, everyday-use and metered, agentic workflows, while Finance and IT teams need two views of Copilot: licence adoption and consumption. One tells you who has access; the other tells you what autonomous or high-effort work is costing.
Microsoft also introduced a central plugin registry and, in its September Copilot update, highlighted general availability for Copilot in SharePoint and OneDrive, an AI receptionist for Teams Phone, reusable PowerPoint skills and stronger citations in Word. The plugin registry is particularly useful for administrators, who can discover, approve and manage Microsoft, partner and custom plugins centrally.
Model choice is welcome, but most users shouldn’t be forced to become model experts. Give your teams a small set of task-based defaults, publish approved plugins and monitor where metered usage is growing. A lack of guidance very quickly becomes inconsistent outputs and harder cost controls.
Meta Enterprise Platform aiming to make the market more competitive
Meta has launched a new enterprise AI division led by former MongoDB chief executive CJ Desai. Meta Enterprise Platform will initially bring together the Muse agent, Meta Business Agent, Muse API, Muse Code and other parts of Meta’s stack for businesses and developers.
The announcement seems, in theory, to be strategically important, but if you open the link to read the statement in full, you’ll see it’s pretty light on practical details right now - there’s no clear packaging, rollout timetable or complete administrative model just yet, so it’s more of a ‘watch and see’ story right now.
Meta’s existing reach across advertising, messaging and customer interactions could make the platform highly relevant, but we’ll wait for specifics on data handling, identity, permissions, auditability and commercial terms before unpacking this more
Barclays offers a useful reality check on enterprise adoption
The week also produced a concrete example of AI at scale, as Barclays said it is expanding Claude across its global operations, with Claude Code expected to reach half of its developer population by the end of 2026 and a majority in 2027. Its colleague knowledge assistant already serves more than 16,000 staff and has handled over one million searches, while a Global Markets workflow uses Claude to classify and route roughly 120,000 emails each day.
This volume of usage may be on the higher side of the scale compared to most businesses, for sure, but there’s a useful lesson in the shape of the deployment: start with repeatable, measurable work; ground the system in trusted information; keep human oversight; and expand only when the operating controls are proven.
Is it less glamorous than handing everyone a chatbot licence and letting them run wild (with parameters in place)? Sure, but it’s also considerably more likely to deliver value.
What businesses should do next
Choose one recurring workflow to test, rather than launching another general AI trial.
Measure cost per completed outcome, including usage-based credits and staff review time.
Create an approved list of models, connectors and plugins, with a named owner for each.
Require approval before agents send messages, change records, spend money or publish externally.
Review identity, audit logs, data retention and regional availability before enabling persistent agents.
Fifosys view
There’s a slightly strange contrast in this week’s AI news.
On one hand, we’ve got persistent agents, cheaper models, more choice within Copilot and another major player entering enterprise AI. On the other, OpenAI has been willing to hold back one of its next models because it wasn’t happy with how reliably it stayed within the boundaries users gave it.
That means there’s plenty worth experimenting with, particularly as AI becomes easier and cheaper to integrate into everyday work. But giving an AI access to your systems is very different from booting up ChatGPT and asking it to draft an email. Start with a useful problem, be deliberate about what the AI can see and do, and keep someone accountable for the outcome.
That’s also where we’ll pick things up in our next AI in Business webinar on October 23rd, looking at what happens when you move beyond experimenting with AI and start integrating it properly into the business.
We’ll see you there!
This week's AI news FAQs
What are OpenAI Dots?
Dots are persistent OpenAI agents designed to work across projects in the background using their own cloud computer and connected applications. They can retain context, monitor ongoing work and request approval before certain consequential actions.
What is ChatGPT Space?
ChatGPT Space is a shared environment where teams and AI agents can work with common project context. OpenAI has also introduced Pages for editable human-agent documents and Team Tasks for recurring or event-triggered work.
What is GPT 6.1 Sol?
GPT 6.1 Sol is an OpenAI model designed for coding, computer use and professional work. OpenAI positions it as delivering performance close to GPT-6 Astra at substantially lower standard API prices.
What is Claude Sonnet 5.5?
Claude Sonnet 5.5 is Anthropic’s everyday model for well-scoped work including coding, documents, presentations and spreadsheets. Anthropic says it is more than 30% faster than Sonnet 5 and can cost up to 30% less per completed task.
Can businesses use GPT 6.1 Sol and Claude Sonnet 5.5 in Microsoft Copilot?
Microsoft has begun rolling both models into Copilot. They are initially available in Cowork and Copilot Studio with usage-based billing, with availability across Microsoft 365 applications subject to licence, region and usage limits.
Why did OpenAI delay GPT-6.1 Astra?
According to reporting cited in this article, OpenAI held back the planned release after safety testing raised concerns about how reliably the model stayed within the scope and authorisation users gave it.
How should businesses prepare for persistent AI agents?
Businesses should define approved data sources, permitted actions and named owners, control plugins and connectors, monitor variable usage costs and require human approval before agents take consequential actions.
What happens when AI starts connecting to the business?
On 23 October, we’re looking at integrations, data, governance and the practical decisions businesses need to make when AI moves beyond isolated prompts and into real workflows.