Daily AI brief · July 25, 2026
Anthropic's Opus 5 Cuts Costs in Half; EU Forces Android AI Open
Anthropic launched a cheaper flagship rival to Fable 5, the EU ordered Google to open Android to competing AI assistants on equal footing, Microsoft's agentic Copilot Cowork went live with usage-based pricing, and Claude is now a selectable model inside Microsoft 365 Copilot—four moves that shift pricing, regulation, and where AI work happens.
Anthropic Launches Claude Opus 5 (Anthropic, July 24)
Anthropic released Claude Opus 5, which it says offers near-Fable 5 capabilities at half the cost .
The model performs well on coding and knowledge-work evaluations . What it means for SMBs/MSPs: You can now get nearly frontier-class reasoning at a price point that fits most team budgets—test it for your clients' highest-value analysis work before committing to more expensive tiers.
EU Mandates Rival AI Access on Android (European Commission, July 16)
The EU issued binding orders requiring Google to grant competing AI assistants the same system-level Android access that Gemini holds, including voice activation at the OS layer via custom wake words and system-wide entry points .
Google must also share anonymized search data with rivals including OpenAI starting in January 2027 . What it means for SMBs/MSPs: Competition is about to shift from "ChatGPT or nothing" to genuine OS-level parity; expect your clients to see native Copilot and Claude as options on European Android devices by mid-2027, changing device recommendations and support conversations.
Microsoft Copilot Cowork Goes Live with Usage Billing (Microsoft, June 16–July 1)
Copilot Cowork reached general availability on June 16, 2026, after a Frontier preview, with Claude Sonnet 5 added to Cowork and PowerPoint on July 2 .
As of July 1, 2026, Cowork stopped working for any tenant without usage-based billing configured . What it means for SMBs/MSPs: Cowork is no longer an experiment—it's an agent that executes real multi-step work. Usage-based pricing means you need to budget for it, but it also means ROI is measurable; pilot it on a defined process (data entry, report generation, approval workflows) with spend caps in place.
Claude Now Selectable Inside Microsoft 365 Copilot Chat (Microsoft, June 16)
Anthropic's Claude is now a selectable model inside Copilot Chat alongside Microsoft's own models, allowing users to choose Claude for complex analysis, document understanding and structured content generation, then switch back for everyday tasks . What it means for SMBs/MSPs: Your clients are no longer locked into a single model. You can recommend Claude for specific work types (long documents, complex reasoning) without asking them to switch tools—this is a real choice, not marketing.
Google's Gemini Delays Mount; Flash Tier Shipping as Stopgap (industry reporting, July 20)
Google has now delayed Gemini's flagship release three times, suggesting structural issues in training or evaluation; the reported stopgap is shipping a Flash-tier model to fill a Pro-tier gap, signaling that the top-end Gemini awaited by enterprise buyers is not imminent . What it means for SMBs/MSPs: Enterprise deals for Gemini's frontier tier are slowing. Clients waiting on Google should move evaluation cycles forward with OpenAI or Anthropic rather than hold; the competitive window is now.
Claude Opus 5 in Your Existing Workflows
What it is:
A thoughtful and proactive model that comes close to frontier intelligence of Claude Fable 5 at half the price .
Who it's for: SMB owners running cost-sensitive analysis or document-heavy teams; MSPs recommending a single Claude tier to clients on subscription budgets.
How long a first test takes: 15 minutes to swap the model in your Claude API call or platform settings; measure cost/token before and after on a sample job.
One concrete first step: Pick one client pain point where reasoning or analysis takes time (contract review, data synthesis, structured decision memos). Run it on Opus 5 instead of your current fallback model. Log tokens consumed and compare to your baseline spend. If cost drops 40%+ with acceptable quality, you've found a scalable swap.
If you use Copilot Cowork, set a spending limit before anyone runs another agent workflow. Since July 1, Cowork only works for tenants with usage-based billing, and it's metered per task, not per user. One agent running many multi-step workflows can burn through credits fast. Set the cap, pilot one small process, and check the actual cost per run before rolling it out.
Rushing into Copilot Cowork without budgets set.
Starting July 1, Cowork stopped working for any tenant without usage-based billing configured . Usage billing is metered per task, not per user. A single agent running 50 multi-step workflows in your finance team can burn through credits fast. Set spending limits first, pilot on a small process, and measure actual cost per execution before rolling out.
Agentic AI is moving from "toolkit" to "embedded layer."
40% of enterprise applications will contain task-specific AI agents by 2026, meaning AI agents are becoming a built-in layer across tools businesses already use . This isn't about buying a new platform. Your clients' existing CRM, ERP, and productivity suites are all getting native agents—and IT teams will need to govern them the same way they govern databases. Start mapping where agents will live in your clients' stacks now, before they surface in updates you didn't request.