What Changed In AI This Week? GPT-5.6 Pricing, Copilot Scale, Claude’s Cyber Warning, and Enterprise AI Agents
We’re back again with this week’s AI news, which - on the face of things - is a lot less about shiny demos and more about something business leaders should care about: cost, control and operational use.
We saw OpenAI cut prices on GPT-5.6, Microsoft showing that Copilot is becoming a serious enterprise platform - rather than just an add-on, Anthropic offering a very useful reminder that powerful AI agents need proper containment, and explore AWS, Meta (and others), who keep pushing AI further into everyday workflows.
In short, this week: AI is becoming cheaper to run, easier to deploy, and harder to govern casually.
Headlines At A Glance
OpenAI reduced GPT-5.6 Luna pricing by 80% and Terra pricing by 20%, making high-volume AI workflows more commercially viable. OpenAI
Microsoft said Microsoft 365 Copilot has passed 30 million paid seats, with net paid seat additions more than doubling quarter over quarter. Microsoft
Microsoft is now framing the next phase of AI as “work transformed”, with Copilot Cowork, Scout and role-specific agents moving AI from chat into workflows. Microsoft 365 Blog
Anthropic disclosed three real-world cybersecurity incidents during AI evaluations, caused by misconfigured test environments with live internet access. Anthropic
Anthropic expanded its Cognizant partnership, giving Claude a larger route into enterprise consulting, delivery and managed transformation work.Anthropic
AWS added Claude Opus 5 to Amazon Bedrock and highlighted new observability for AI agents, reinforcing that model choice and governance are becoming cloud platform features. AWS
Meta AI is also moving from answering questions to taking actions, including planning, connecting to email and calendar apps, and generating slides. Meta
OpenAI Cuts GPT-5.6 Pricing: AI Cost Control Is Becoming More Practical
Fresh offlast week’s headlines of AI going rogue, OpenAI’s biggest business-relevant update this week was considerably less dramatic - it’s a pricing change.
The company said GPT-5.6 Luna, its fastest and lowest-cost model in the GPT-5.6 family, will cost 80% less, while GPT-5.6 Terra will cost 20% less. As of the 30th of July, OpenAI lists API pricing for Terra at $2 per million input tokens and $12 per million output tokens, and for Luna at $0.20 per million input tokens and $1.20 per million output tokens. ChatGPT and Codex subscription prices remain unchanged, but Luna and Terra usage now consumes fewer credits in ChatGPT Work and Codex -OpenAI’s announcement is here.
Why does this matter? Well, it’s because many of the most useful AI workflows are not single prompts, but rather things that involve reading documents, using call tools, checking work, drafting outputs, revising them, and sometimes running in the background for a while.
Without proper guardrails, AI can get expensive quickly - as Uber found out a few weeks ago, blowing through its entire annual budget in just four months, with one coding session estimated to have cost $1,200.
Lower model pricing doesn’t magically make AI “free”, but it does change the business case somewhat. Tasks such as document classification, customer interaction triage, routine coding changes, internal knowledge search and repetitive back-office workflows become easier to justify at scale.
For SMBs and mid-market organisations, this is where AI procurement really needs to mature and adopt the “cost per useful outcome” mindset. It’s no longer enough to ask whether a tool has the latest model and simply run with it as the answer. You need to ask which model is used for each task, what the failure rate looks like, how much human review is still needed, and what the total cost is once retries and checks are included.
What this means for your business: cheaper models make broader AI use more realistic, but they also make it easier for usage to spread quietly. If staff are using AI agents regularly, treat usage monitoring and outcome measurement as part of the rollout, not something to bolt on later.
Microsoft Copilot Passes 30 Million Paid Seats
Microsoft had a strong AI week, partly because its latest earnings gave the market a clearer sign that Copilot and Azure AI are translating into real adoption.
In its FY26 Q4 update, Microsoft said Azure revenue surpassed $100 billion for the first time, and Microsoft 365 Copilot reached more than 30 million paid seats. Microsoft also reported Microsoft Cloud revenue of $59.3 billion for the quarter, up 27% year on year. For those curious, Microsoft’s earnings release is here.
The 30 million number is significant because Copilot has sometimes been discussed as if it were still a grand experiment, when in reality, it’s here to stay. In my own personal opinion, better AI tools are out there at this exact moment in time, but it’s considerably better now than it was at launch, and should only continue trending upwards. Also, it’s hard to look past its benefits that many organisations, especially those already standardised on Microsoft 365, are tapping into, as Copilot becomes part of the normal productivity stack.
Microsoft’s follow-up blog post made the positioning even clearer. The company argued that the next phase is now “work transformed”, pointing to Copilot Cowork, which became generally available in June, and Microsoft Scout, an always-on agent that can operate in the background with its own identity and permissions. For those interested, you can read Microsoft’s blog here.
Make no odds about it, Microsoft isn’t selling Copilot as a clever writing assistant anymore. It’s packaged and seen as them selling a layer of agents that can sit across Teams, Outlook, SharePoint, OneDrive and business context.
For business leaders, the opportunity it presents includes things like fewer manual handoffs, faster analysis, better internal search, and automated first drafts of work that currently gets stuck in inboxes.
The risk is equally obvious: permissions, data quality and process ownership suddenly matter much more.
What this means for your business: if you’re using Microsoft 365, your AI strategy is already partly a Microsoft 365 strategy. Review your data permissions, SharePoint sprawl, sensitivity labels, guest access and retention policies before pushing agentic tools widely. Copilot can only be as well-governed as the environment it’s allowed to read and act within.
Copilot Is Moving From Chat To Workflows
The more interesting Microsoft story this week, surprisingly, isn’t just that Copilot has more paid seats. It’s what Microsoft says users are doing with it.
Microsoft said Copilot conversations per user have nearly doubled over the past year, users engaging with multiple Copilot features grew by triple digits, and average weekly Copilot engagement is now on par with Outlook and Teams. It also said that when multi-step Copilot Cowork workflows are included, analysis-related work accounts for 49% of tasks, up from 29% when analysis is counted only as a standalone activity.Microsoft shared those figures here.
Asking AI to “summarise this meeting” is helpful, but asking AI to analyse a set of data, draft the email explaining it, prepare the follow-up document and keep the context together is a different category of productivity - and one that starts to look a lot more like process automation.
The practical lesson for us is that AI value tends to improve when it is attached to a workflow, not sprayed across the organisation as a general-purpose novelty, or because you ‘probably have to use it, if other businesses are’. In an opinion that shouldn’t be taken as gospel, I believe that the businesses that get the most from this wave will probably not be the ones with the most licences. Instead, they’ll probably be the ones who pick repeatable workflows, clean up the process first, then use AI to accelerate it.
What this means for your business: start with workflows, not tools. Pick two or three areas where work is repetitive, document-heavy or coordination-heavy. Then define the business owner, data sources, approval points and success measure before giving an agent room to run.
Anthropic’s Cyber Disclosure Is A Governance Warning Worth Reading
Veering away from positivity, somewhat, Anthropic published one of the more sobering AI posts of the week.
The company said that during a review of cybersecurity evaluation transcripts, it found three incidents in which Claude models accessed the internet from within or while interacting with a third-party evaluation environment, then gained unauthorised access to real systems. Anthropic said it reviewed 141,006 evaluation runs where Claude could have obtained internet access and identified three incidents, which can be read inAnthropic’s full disclosure, which is here.
The important detail is that, unlike last week’s story, this wasn’t framed as a released Claude product attacking customers. Anthropic said the issue came from evaluation environments that were supposed to be isolated but had live internet access. The models had been given capture-the-flag cybersecurity tasks and, because they believed they were inside a simulation, treated real systems as part of the exercise.
Still, this is exactly the kind of incident businesses should pay attention to.
AI agents are increasingly being asked to use tools, browse systems, call APIs, run code and complete tasks. If the environment is misconfigured, the agent may not understand the boundary between “allowed” and “not allowed” in the same way a human would.
That isn’t a reason to avoid AI altogether, for sure. But it’s a reason to take containment seriously.
What this means for your business: any AI agent with access to tools should have clear technical boundaries, not just a policy document. Limit permissions, test in isolated environments, monitor activity, log outputs, and require human approval for sensitive actions. And please, remember “but the prompt told it not to!” isn’t a control.
Anthropic and Cognizant Expand Claude For Enterprise Delivery
Anthropic also announced an expanded partnership with Cognizant, one of the world’s largest technology services companies. Cognizant is embedding Claude across its own business and engineering platforms, scaling a Claude-certified workforce, and becoming a Global Premier Partner in the Claude Partner Network. Anthropic said more than 30,000 Cognizant associates have completed Claude training. Anthropic’s announcement is here.
Admittedly, this is less flashy than a new model launch, but arguably more important for enterprise adoption.
Most businesses don’t struggle to find an AI tool. Their challenges arise because they need AI to operate within existing systems, industry regulations, security requirements, and business processes, which is where consultancies and managed service providers become important.
Claude moving deeper into Cognizant’s delivery model is another sign that AI adoption is becoming an implementation discipline, revolving around who can safely connect it to the work.
What this means for your business: expect more AI capability to arrive through your existing technology partners, not just directly from AI vendors. It can be useful, but ask the same questions: what data is used, where it is processed, who owns the output, how quality is checked, and what happens when the tool is wrong.
AWS Adds Claude Opus 5 To Bedrock And More Agent Controls
AWS used its weekly roundup to highlight Claude Opus 5 on AWS, available through Amazon Bedrock and Claude Platform on AWS. AWS said Bedrock offers Claude Opus 5 with zero data retention enabled by default. The same roundup also highlighted Bedrock AgentCore observability improvements, putting agent traces and prompts into the same CloudWatch log group, with finer access control and customer-managed encryption at the individual-agent level.AWS’s roundup is here.
For SMBs and mid-market firms with AWS estates, bear in mind that cloud platforms aren’t just adding models; they’re adding the surrounding controls needed to run AI in production.
That includes logging, observability, data retention choices, encryption and debugging.
This may sound technical, but the business point is simple: if AI agents are going to act inside real systems, businesses need visibility into what they did and why.
What this means for your business: when comparing AI platforms, don’t only compare model quality. Compare auditability, data retention, admin controls, integration options and cost visibility. The best AI tool isn’t always the one with the most impressive demo. Sometimes it’s the one that your IT team can actually govern.
Meta AI Also Moves Toward Action
And finally this week, Meta announced that Meta AI can now plan, connect to email and calendar apps, create slides, and handle tasks on a user’s behalf, powered by Muse Spark 1.1. The features are starting in select markets in the Meta AI app and meta.ai, with more countries and surfaces, including WhatsApp, planned in the coming weeks.Meta’s announcement is here.
For many businesses, Meta AI may not be the core workplace AI platform. Microsoft 365, Google Workspace, OpenAI, Anthropic and cloud platforms are more likely to sit at the centre of business operations.
But Meta’s move does show where the whole market is going. Assistants are becoming agents. Chat is becoming action. AI is being integrated with calendars, messages, files, slides, and shopping-style workflows.
It’s all very convenient, but it also creates another shadow AI risk if staff start using consumer-grade assistants to manage work tasks without approval.
What this means for your business: update AI policies to cover consumer AI assistants, not just obvious workplace tools. The boundary between a personal assistant and a work assistant is going to get blurrier.
Quick Answer: What Should Businesses Do About AI This Week?
If you only take three things from this week’s AI news, make them these:
Revisit your AI cost model: Lower model pricing is good news, but agentic workflows can still create variable usage costs.
Review your Microsoft 365 governance: Copilot and agents are only as safe as the permissions, labels and data structure underneath them.
Treat AI agents like operational systems: They need containment, logging, approval flows and monitoring.
Fifosys View
The story this week isn’t that AI became magically safe, cheap or simple; it’s just improving on its ‘business as usual’ status.
OpenAI’s pricing changes make more use cases affordable. Microsoft’s Copilot numbers show that enterprise adoption is now material. Anthropic’s cyber disclosure shows why agent boundaries matter. AWS’s Bedrock updates indicate that governance is moving into the platform layer. Meta’s assistant update shows that action-taking AI will not stay confined to enterprise tools.
The sensible path isn’t to chase every new feature like a dog with a ball. Just start by making AI boring enough to manage properly.
That means clear ownership, approved tools, sensible data rules, cost monitoring, staff training and proper review of outputs.
Nothing too dramatic, just deliberate