What Changed In AI This Week? GPT-5.6, ChatGPT Work, Claude Fable 5, Meta’s Model API, And The Agentic AI Cost Question
It’s Friday, the 10th of July 2026, and this has been yet another busy week for business AI. The headlines this week aren’t simply that ‘the models are getting better’ (though they are), but we’re taking a look into how AI is moving further into everyday work: Microsoft 365, ChatGPT, development workflows, cyber operations, sales processes, finance teams and cloud marketplaces.
As we’re finding out more and more in recent weeks, the question most businesses are asking have gone above and beyond just “Should we experiment with AI?” Instead, it’s becoming: “Which tools are already entering the business, what can they access, what do they cost when used heavily, and who is responsible when they act on our behalf?”
Headlines at a glance
OpenAI launched GPT-5.6, with three model tiers, new multi-agent capabilities, and clearer token pricing for developers.
OpenAI also introduced ChatGPT Work, an agentic work mode designed to create documents, spreadsheets, slides, web apps and recurring workflows across connected tools.
Microsoft confirmed GPT-5.6 is available in Microsoft 365 Copilot, bringing OpenAI’s latest model family into Word, Excel, PowerPoint, Chat and Cowork.
Anthropic’s Claude Fable 5 and Mythos 5 remain important for businesses to watch, not only for capability, but because Anthropic is changing how high-end model access, pricing and business data retention work.
Meta launched Muse Spark 1.1 and the Meta Model API, giving developers public preview access to a new agentic coding and multimodal model.
Google and AWS both pushed further into practical agent infrastructure, with Google making AlphaEvolve generally available and AWS expanding ways to buy and deploy production-ready AI agents.
The EU AI Act is getting closer to its next major date: most of the Act becomes applicable on the 2nd of August 2026, with transparency obligations for generative AI also applying from that date.
OpenAI GPT-5.6: Better Models, But Also A Clearer Cost Conversation
OpenAI launched GPT-5.6 this week, describing it as a new family of models: Sol as the flagship, Terra as the balanced everyday option, and Luna as the most cost-efficient model.
GPT-5.6 isn't as basic as being just another model update, but it should be looked at how OpenAI are also putting much more emphasis on cost and performance, which reflects how businesses are actually using AI now. Instead of asking the odd question, teams are running longer research tasks, generating documentation, writing code and using AI across multiple steps of a workflow.
OpenAI says GPT-5.6 is available across ChatGPT, Codex and the API. For developers, the models are priced per million tokens: Sol at $5 input / $30 output, Terra at $2.50 input / $15 output, and Luna at $1 input / $6 output. OpenAI has also introduced more predictable prompt caching and says its Responses API now supports Programmatic Tool Calling and a multi-agent beta.
In plain English, that means AI systems are becoming less like an assistant answering a question, and more like a small team that can use tools, split up work, check intermediate results and bring back a finished answer.
What this means for your business: GPT-5.6 looks like a serious step forward for software, analysis and knowledge work. But the bigger lesson is cost design. If teams start using agents for long-running tasks, businesses will need to monitor usage, define approved workflows and check whether the time saved is worth the variable consumption cost.
ChatGPT Work: OpenAI Wants To Move From Chatbox To Working Environment
Alongside GPT-5.6, OpenAI introduced ChatGPT Work, which it describes as an agent that can take action across apps and files, stay with a project for hours, and turn a goal into finished work.
ChatGPT Work can connect to all those business tools you rely on, such as Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, CRMs and project trackers. OpenAI says it can create sheets, slides, docs and web apps, and can run scheduled tasks such as reviewing Slack updates, checking dashboards, summarising website changes or updating a presentation when new feedback arrives.
While having the potential to be incredibly useful, it’s also exactly the kind of thing that needs governance.
OpenAI says Enterprise and Edu admins can manage who has access, what company context ChatGPT can use, which tools it can connect to and what actions it can take. There are also spend controls for ChatGPT Work usage.
What this means for your business: ChatGPT Work could be very useful for recurring admin, sales preparation, finance analysis, reporting and internal project tracking. But it shouldn’t be introduced casually. Before enabling connected tools, decide which data they can access, which actions require approval, who will review outputs, and how usage will be monitored.
Microsoft 365 Copilot Gets GPT-5.6
Microsoft has made OpenAI’s GPT-5.6 available in Microsoft 365 Copilot, including Word, Excel, PowerPoint, Chat and Cowork.
For businesses already invested in Microsoft 365, this may be more immediately relevant than the standalone OpenAI announcement. Most employees don’t want (or even need) another place to work - they just want AI to help in the documents, spreadsheets, meetings and chats they already use.
The question you’ll likely have about how good this news is all depends on whether GPT-5.6 can make Copilot more reliable at producing usable first drafts, analysing messy spreadsheets, summarising business context, and helping teams get from “rough idea” to “something we can actually use”.
The caveat is familiar: better AI inside Microsoft 365 is still not the same as automatic business value. If Copilot has access to poorly governed SharePoint sites, outdated files or over-permissive Teams channels, it may surface information people should not see or base answers on material nobody trusts.
What this means for your business: if you already use Copilot, this is a good moment to review permissions, data quality and usage reporting. The model is improving, but Copilot is only as useful and safe as the Microsoft 365 environment it sits inside.
Anthropic Claude Fable 5: Capability, Pricing, and Data Retention Are Now Linked
Anthropic’s Claude Fable 5 and Claude Mythos 5 remain one of the most important AI stories for business users, even though the original launch was in June. The model was redeployed on 1 July after access disruption, and Anthropic says Fable 5 is now available everywhere, while Mythos 5 remains restricted to selected trusted-access use cases.
Fable 5 is positioned as Anthropic’s most capable generally available model, especially for software engineering, knowledge work, vision, long-context tasks and scientific research. Pricing is $10 per million input tokens and $50 per million output tokens.
The detail businesses should pay attention to is, once again, data handling. Anthropic says it will require 30-day retention for all traffic on Mythos-class models, on both first- and third-party surfaces, for safety reasons. It also says this data will not be used to train new Claude models or for non-safety purposes.
While being a sensible safety argument, it changes the procurement conversation. Some businesses have become used to treating enterprise AI as “no training on our data” and assuming that is the whole privacy story, but frankly, it’s not. Retention, human access controls, deletion, audit logs and third-party hosting all matter.
Anthropic also published business examples this week, including UST bringing Claude into physical AI and engineering workflows, and the Government of Alberta using Claude Code to scan 466 million lines of code in 20 hours and help find and fix vulnerabilities.
What this means for your business: Claude remains a serious option for complex technical and analytical work. But for high-capability models, read the data terms carefully. The question is not just “Is the model good?”, but it must also crucially encompass “What happens to our data, how long is it retained, who can access it, and does that fit our risk profile?”
Meta Muse Spark 1.1: AI Competition Is Becoming A Pricing Story
Meta launched Muse Spark 1.1 this week and opened a public preview of the Meta Model API for developers.
Meta says Muse Spark 1.1 is built for agentic tasks, computer use, coding, multimodal reasoning and long-context work. It can work with a one-million-token context window, support multi-agent orchestration, handle images, video and PDFs, and operate across tools.
OpenAI, Anthropic and Google have dominated much of the serious business AI conversation, but Meta now appears to be pushing harder into developer and agent workloads, not just consumer AI inside Facebook, Instagram and WhatsApp.
If Meta prices aggressively, it could put pressure on the market, which is good for buyers - but it also creates another decision point. If you think choosing a streaming platform to turn to at night is hard work, imagine the conversation around AI?
Just remember, the cheapest model is not always the best for regulated work, client data, support expectations, or long-term reliability.
What this means for your business: if you build AI into products or internal tools, Meta’s Model API may become worth testing. But the usual procurement questions still apply: data handling, regional availability, support, model stability, auditability and whether the model is good enough for the specific workflow.
Google And AWS Push AI Agents Toward Production
Google made AlphaEvolve generally available on the Gemini Enterprise Agent Platform. It is aimed at code optimisation and algorithm discovery, with examples across logistics, semiconductors, genomics, high-performance computing and financial services.
For most SMEs, AlphaEvolve may sound too specialist. But the broader signal matters: AI agents are moving from “help me write a paragraph” into highly specific business and technical workflows where optimisation has measurable value.
AWS made a similar push with its post on production-ready AI agents at scale. AWS is leaning into AgentCore, identity, observability, memory, browser use, code execution, S3 Vectors and marketplace access to AI agents and tools.
That tells us where cloud AI is heading. The next phase is not just model choice. It is the infrastructure around agents: permissions, monitoring, memory, deployment, procurement and integration with existing systems.
What this means for your business: if you are still at the “random AI pilots” stage, this is a reminder to slow down and structure the work properly. Useful AI agents need secure data access, identity controls, monitoring and clear ownership. Otherwise, they become clever prototypes that nobody is comfortable putting into production.
EU AI Act: The Compliance Clock Is Getting Louder
The EU AI Act is now close enough for businesses to pay attention to - it’s no longer a distant regulatory topic. The European Commission says the AI Act entered into force on the 1st of August 2024 and will be fully applicable from the 2nd of August 2026, with some exceptions.
For businesses using generative AI, the transparency rules are particularly relevant. The Commission’s Code of Practice on Transparency of AI-Generated Content says that Article 50 obligations apply from the 2nd of August 2026 and cover the marking, detection, and labelling of AI-generated or manipulated content, including deepfakes and certain public-interest text.
This won’t affect every business in the same way, but if your organisation sells into the EU, publishes AI-generated content, uses AI in customer-facing workflows, or deploys AI in HR, finance, education, healthcare, public services or other sensitive areas, you should be checking obligations now.
What this means for your business: AI governance is becoming a compliance issue, not just an IT preference. Keep an inventory of AI tools, document where AI is used, decide when AI-generated content needs labelling, and make sure staff know what data can and cannot go into public tools.
Quick Answer: What Should You Do About AI This Week?
If you only take three actions from this week’s news, make them these:
Review which AI tools are already connected to business systems such as email, Teams, Slack, SharePoint, Google Drive, CRMs and finance platforms.
Put cost controls around agentic AI before usage scales, especially for tools that run long tasks, use multiple agents or perform scheduled work.
Update your AI policy to cover data retention, approved tools, human review, AI-generated content labelling and who owns each workflow.
The real story this week is that AI is becoming more capable and more embedded at the same time. That is powerful, but it also makes casual adoption riskier.
Fifosys View
The most useful AI tools are becoming less like chatbots and more like junior digital colleagues: they can read, write, search, click, schedule, analyse and sometimes act, and that’s why they’re becoming genuinely valuable.
Yet, it’s also why they need management.
For business leaders, the right response is not panic, and it equally isn’t blind enthusiasm. Put simply, it’s structure. Pick a small number of high-value workflows. Use approved tools. Set access controls. Monitor spend. Keep humans in the loop for important decisions. Review outputs before they reach clients, suppliers or regulated processes.
AI is moving quickly, but the businesses that benefit most will probably be the ones that make it boring enough to manage properly.