What Changed in AI This Week? Claude Opus 5.5, ChatGPT Business Controls, Microsoft Foundry and Agentic Commerce
What Changed in AI This Week? Claude Opus 5.5, ChatGPT Business Controls, Microsoft Foundry and Agentic Commerce
The biggest AI news for businesses this week, from cheaper frontier models and voice-driven work to security agents and AI-powered shopping.
AI is getting cheaper, more connected and more capable of taking action across business systems. That makes model choice less important in isolation, while permissions, variable spend, connected data and governance become more important.
In a week where AI discovered a new enzyme system - its first scientific discovery - we’re looking into some of the other, business-led stories that tie into it.
Staying with Anthropic, they made high-end AI materially cheaper. Equally, OpenAI gave ChatGPT Business more ways to act across company tools, along with tighter administrator controls. Microsoft expanded the range of models and voice agents available through Foundry and Amazon, and Shopify took sharply different positions on whether external AI agents should be allowed to shop on their platforms.
It feels like we say this every week, but AI isn’t just for answering questions; it can (and does) do proper work. For businesses, the cost, permissions, platform choice and governance are more important than another leaderboard score.
Headlines at a glance
Claude Opus 5.5
Anthropic claims performance comparable to its top coding model on many tasks, with typical workloads costing 40% less than Opus 5.
ChatGPT Business
Voice can now work with plugins, connected apps, documents and browsers, alongside new administrator controls over external access.
Microsoft Foundry
The full GPT-6 family and Claude Opus 5.5 join Foundry alongside public-preview voice agents and further governance capabilities.
Microsoft Defender
An integrated security operations centre gives human teams and AI agents a shared foundation for threat detection and response.
Agentic commerce
Amazon blocked Meta’s Muse shopping agent while Shopify opened access, highlighting the emerging battle over AI-driven purchases.
Claude Opus 5.5 makes frontier AI cheaper to use
Anthropic launched Claude Opus 5.5 on the 22nd of September, positioning it as its new leading model for agentic coding, computer use and knowledge work.
The headline improvement isn’t just capability, but efficiency too.
Anthropic says Opus 5.5 costs 40% less than Opus 5 on typical workloads, accounting for both lower token prices and reduced token usage. Standard API pricing is $4 per million input tokens and $20 per million output tokens, down from $5 and $25, respectively. Cache reads, which can account for a large proportion of the cost of coding and agentic workloads, have fallen from $0.50 to $0.20 per million tokens.
Anthropic also says the model generates output more than 30% faster than Opus 5. Usage limits are being increased for Pro, Max, Team and seat-based Enterprise plans.
The company has published plenty of benchmark results, including claims that Opus 5.5 leads on several coding and professional work evaluations. As ever, businesses should treat vendor benchmarks as a useful starting point rather than a purchasing decision.
The commercially interesting part is the price-performance shift, as complex AI work is becoming less expensive surprisingly quickly. Tasks such as reviewing a large codebase, researching across hundreds of documents or running a long multi-stage analysis are starting to look more viable outside very large enterprises.
There is an equally important governance point, as Anthropic says Opus 5.5 is less likely than recent models to take hard-to-reverse actions or operate outside its instructions, and has stronger prompt-injection resistance. Those are meaningful improvements when a model is allowed to use tools or work for hours rather than simply draft an email.
Don’t automatically move every Claude workload to the newest model. Test Opus 5.5 against a small number of expensive or difficult tasks, then compare total task cost, completion time and the amount of human checking required. A cheaper token is useful, but a cheaper completed outcome is what actually matters.
ChatGPT Business can now do more through voice, with new admin controls
OpenAI’s announcements were less dramatic, but arguably more relevant to everyday adoption.
According to the ChatGPT Business release notes, ChatGPT Voice can now use plugins and connected apps on web, iOS and Android. Voice is also available in ChatGPT Work on web and mobile, allowing users to ask it to create documents, presentations and spreadsheets, use connected applications or complete work in a browser.
If a task is unfinished when the voice conversation ends, it can continue in text.
This sounds like a user-interface update, but the more important change is what’s sat under the hood, so to speak. Voice is becoming a way to initiate actions, not merely dictate prompts, and we actually find in some instances - from internal use - that you get a better quality prompt generated when you talk to AI, rather than just typing. For you, that could look like a user starting a piece of work while travelling, asking ChatGPT to gather information from connected systems, and returning later to review the output.
OpenAI also added new external-access controls for ChatGPT Business administrators. Global admins can manage whether ChatGPT Sites can use members’ connected apps and whether applications can access ChatGPT Ads (both permissions are off by default during the admin preview, and individual users must still authorise access through their own accounts).
A new Privacy Centre is also rolling out, bringing information about memory, personalisation, connected apps, data use and account security into one place.
It should go without saying, but a more capable assistant needs more access. Connected apps may contain customer records, internal discussions, commercial documents and other information that should not flow freely between systems just because a feature is convenient.
OpenAI says existing workspace permissions and action restrictions continue to apply. Businesses still need to decide whether those settings reflect their own policy.
Review connected-app permissions before encouraging voice-driven work. Confirm which applications are approved, what information users can expose to ChatGPT and when an action should require human approval. Voice makes delegation easier; it doesn’t remove the need for oversight.
Microsoft Foundry expands model choice and launches voice agents
Microsoft used the week to strengthen Foundry as a platform for businesses that do not want to build around a single AI provider.
The company announced that the full OpenAI GPT-6 family, including GPT-6 Sol and GPT-6 Luna, and Anthropic’s Claude Opus 5.5, are now available through Microsoft Foundry.
Microsoft’s argument, put simply, is this: the best model for a coding task probably isn’t going to be the best one for research, customer service or high-volume processing. Foundry is intended to let organisations compare models against their own workloads and switch based on quality, latency and cost without rebuilding the wider system.
It also introduced voice agents in public preview, which support more than 80 languages and 140 locales, with features such as natural turn-taking and recovery when someone interrupts. Microsoft says voice agents can be built, deployed, monitored and evaluated through the same platform and tooling as other Foundry agents.
For growing businesses, possible applications include customer service triage, appointment management, internal IT support, and field worker assistance - although a convincing voice experience, for example, isn’t automatically a trustworthy one. Authentication, disclosure, call recording, escalation and access to customer data all need to be designed into the process.
Microsoft also reiterated how it sees workplace AI being bought. A separate Copilot update distinguishes between predictable per-user subscriptions and usage-based billing for longer-running agentic tasks, metered through Copilot Credits.
That hybrid model is likely to become the norm across the market, and everyday assistance can fit within a licence; agents that run for extended periods, retrieve large amounts of context, and call several tools are closer to cloud consumption.
Avoid choosing an AI model in isolation. Define the business task first, then compare models using your own examples and include the whole operating cost: licences, variable consumption, integration, monitoring and human review.
Microsoft Defender brings AI agents into security operations
Staying with Microsoft, they also announced an integrated security operations centre, or ISOC, in Microsoft Defender, available in preview.
The idea is to bring security information and event management, threat protection and AI-assisted operations onto a shared foundation. People and agents can then work from the same signals, context and controls rather than passing incidents between disconnected systems.
For smaller IT and security teams, that could be a significant development - or at the very least, something worth keeping an eye on. Alert investigation, threat hunting and containment are precisely the kinds of high-volume workflows where carefully constrained automation can help.
(“Carefully constrained” is doing some work there…)
Security agents can analyse and respond much faster than a human team, but the consequences of a mistaken action may also arrive faster. Businesses evaluating autonomous security features should understand what an agent can observe, which actions it may take automatically and how those actions are audited or reversed.
Treat AI security agents as privileged operators and apply least-privilege access, require approval for high-impact actions and retain clear activity logs. A faster response is valuable only when the response itself remains controlled.
Amazon and Shopify split over AI agents buying for customers
One of the week’s most revealing developments happened in online retail.
Amazon blocked Meta’s Muse agent from browsing and purchasing on Amazon.com. Amazon told Axios that third-party applications purchasing on behalf of customers should identify themselves and respect whether a service provider chooses to participate.
Shopify, it seems, is taking a different route. Its documentation says eligible merchants can make products available through Meta surfaces, including Muse, with a Shopify-powered direct checkout. Merchants remain responsible for fulfilment, returns and customer service, and can deactivate direct checkout.
There is an important analytics wrinkle: third-party tracking pixels don’t fire when checkout is completed on Meta rather than the merchant’s own website, and Shopify attributes those orders to Meta inside Shopify Admin, but businesses may lose some of the customer-journey data they normally use for marketing analysis.
This is more than a disagreement between large technology companies. It’s an early version of a question many businesses will face: do you want AI agents to access your services, recommend your products and transact with you?
Allowing them could create a useful new sales channel, and it could also place an intermediary between the business and its customer, while complicating consent, attribution, support and responsibility for mistakes.
Retailers should start treating AI channels like any other marketplace. Decide where products may appear, who owns the customer relationship, what data is retained, how orders are attributed and what happens when an agent purchases the wrong item.
Quick answer: What should SMEs do about AI this week?
If you only take four actions from this week’s news, make them these:
Benchmark outcomes, not model names: Test new models against real work and measure total cost, accuracy and review time.
Audit connected-app permissions: Voice and agent features make it easier for AI to cross between systems, so check what users and tools can actually access.
Put limits around variable AI spend: Assign an owner, set budgets and monitor usage before long-running agents become routine.
Map your exposure to external agents: Consider whether AI assistants can browse, buy from or act within your digital services, and whether your current policies cover that behaviour.
Fifosys view
As we started by saying, AI is becoming more capable and more operational, week by week. Its upward trajectory isn’t going to change any time soon, it doesn’t seem.
Claude Opus 5.5 lowers the cost of demanding work. ChatGPT can act across more business tools through voice. Microsoft is making it easier to switch between models and build agents into customer and security workflows. But the Amazon-Meta disagreement? Well, it shows that those agents will not always be welcomed wherever they go.
The businesses that benefit will not necessarily be the ones that adopt every release first, but we’d wager that they will be the ones that know which work should be delegated, which systems an AI can access, how much it is allowed to spend and where a human still needs to make the decision.
That may sound less exciting than another model benchmark, but it’s also much closer to how useful technology becomes part of a well-run business.
This week's AI news FAQs
What is Claude Opus 5.5?
Claude Opus 5.5 is Anthropic’s latest high-end model for agentic coding, computer use and knowledge work. Anthropic says typical workloads cost 40% less than Opus 5 while the model also produces output faster.
What changed in ChatGPT Business this week?
ChatGPT Voice can now work with plugins and connected apps, while ChatGPT Work can use voice instructions to create files, work across connected applications and complete browser-based tasks. OpenAI has also introduced additional administrator controls over external access.
What is Microsoft Foundry?
Microsoft Foundry is a platform for building, deploying and managing AI applications and agents across different models. Its expanding model catalogue allows organisations to compare providers based on their own requirements for quality, latency and cost.
What are AI security agents?
AI security agents can help investigate alerts, analyse threats and support or automate response actions. Because those actions can affect live systems, businesses need clear permissions, approval thresholds, audit logs and ways to reverse high-impact decisions.
What is agentic commerce?
Agentic commerce describes AI agents helping customers discover products, make decisions and potentially complete purchases on their behalf. It raises questions around customer ownership, consent, attribution, transaction data and responsibility when something goes wrong.
How should SMEs control AI costs?
Businesses should measure the total cost of completing a task rather than looking only at token prices or licence fees. That includes variable model consumption, integrations, monitoring, human review and the cost of correcting poor outputs.
Turn AI capability into controlled business value
We can help you identify useful AI workflows, understand the systems and data involved, control access and build governance around what happens next.
Talk to our team
Bring us the workflow, AI tool or business process you’re considering and we’ll help you work through the practical next step.
AI for business
See how Fifosys approaches AI adoption, automation, governance and practical business use cases.