Daily AI brief · May 31, 2026
Anthropic's June 15 Billing Split Is the Agent Tax You Didn't Plan For
Google's Gemini 3.5 Flash goes GA for agents and coding, Anthropic reprices agentic Claude workloads starting June 15, Copilot Studio graduates from chatbot builder to governed agent platform, and AI-powered phishing hits SMBs harder than ever.
1. Anthropic's June 15 Billing Split: The "Agent Tax" Is Real *(The New Stack / Codersera / Beam AI)*
On May 14, Anthropic announced a major restructuring of its Claude subscription billing, effective June 15. The Agent SDK, the `claude -p` command, Claude Code GitHub Actions, and all third-party agent applications will be removed from the existing subscription usage pool and migrated to a brand-new, separately billed "Agent SDK Credit pool."
Pro users get $20 in credit, Max 5x users get $100, Max 20x users get $200. Credits expire monthly and do not roll over.
The real sting:
agentic workflows with large context windows can burn through 100,000–200,000 tokens per session
— which means that $20 credit evaporates fast.
What it means for SMBs / MSPs: If any of your clients (or your own team) are running automated Claude pipelines, CI/CD agents, or scheduled Claude Code tasks,
if they don't know about it before June 15, they can find themselves with stopped automations, surprise overage charges, or both.
2. Google I/O: Gemini 3.5 Flash Goes GA — No Waitlist, Agent-First *(Google Cloud Blog / MarkTechPost)*
Google pushed Gemini 3.5 Flash into the Gemini API, Google AI Studio, and Vertex AI simultaneously with the I/O 2026 keynote announcement on May 19 — no waitlist, no preview tier.
3.5 Flash offers a balance of performance and speed, ideal for tackling long-horizon agentic tasks, often at less than half the cost of comparable models.
Notably,
Xero has deployed agents for complex, multi-week workflows, and Ramp uses it for smarter OCR on invoices.
What it means for SMBs / MSPs:
The feature management toggle for Gemini 3.5 Flash will not be available after June 8, 2026 — starting that date, it is enabled by default and cannot be turned off for users in the Gemini Enterprise app.
If your clients are on Google Workspace, the model underneath their AI tools just changed. Audit before they call you.
3. Copilot Studio: Chatbot Builder No More — Governed Agent Platform *(Windows Forum / Microsoft Copilot Blog)*
Microsoft Copilot Studio's May 2026 wave of Copilot and Power Platform changes recasts the product from a low-code chatbot designer into a governed enterprise agent platform for building, orchestrating, testing, and operating AI agents across Microsoft 365, business applications, web interfaces, and desktop workflows.
The big structural shift:
the redesigned workflows experience introduces a more intuitive visual designer for building and orchestrating agentic automation in one place — instead of stitching together disconnected tools and logic across multiple surfaces, you can design workflows end-to-end on a unified canvas.
Meanwhile,
the computer-using agent (CUA) feature, running on Claude Sonnet 4.5 (beta), expanded to all commercial geographies in Microsoft Power Platform.
What it means for SMBs / MSPs:
Platforms that promise to simplify work often become new places where work accumulates. Copilot Studio will not eliminate the need for integration design; it will make integration design available to more people, at higher speed, with AI making some decisions that used to be encoded manually.
This is a billable services opportunity — and a governance liability if you leave clients to figure it out alone.
4. AI-Powered Phishing Is Outpacing SMB Defenses *(Mainstay Technologies / MSSP Alert)*
Microsoft reported in April 2026 that attackers used generative AI to create hyper-personalized lures tied to the victim's role — including themes like invoices, RFPs, and manufacturing workflows — while pairing those lures with automation that improved the odds of account compromise at scale.
Cyber threats against SMBs rose sharply in 2025, with incidents nearly doubling compared to the prior year and over one quarter of all American SMBs experiencing a cyberattack within the past 12 months.
What it means for SMBs / MSPs:
The old advice to "watch for bad spelling" is no longer enough — the phishing email may look polished, reference your real vendors, and arrive at exactly the right moment in a workflow.
Mailbox-level AI security is no longer optional for client stacks.
5. Claude Opus 4.8 Drops — 41-Day Cycle Is the New Normal *(ChatForest)*
Claude Opus 4.8 landed May 28, 2026 — less than seven weeks after Opus 4.7. On SWE-Bench Pro it rises from 64.3% to 69.2% (vs. GPT-5.5 at 58.6%), and GDPval climbs from 1753 to 1890.
Two new features ship alongside: Effort Control (Low/Medium/High/Max, defaulting to High) and Dynamic Workflows for Claude Code, a research preview that runs hundreds of parallel subagents on massive tasks.
What it means for SMBs / MSPs:
The model you picked three months ago may already be outdated. Build your product stack to swap models without rebuilding everything.
Copilot Studio's Computer-Using Agent (CUA) — now GA in Power Platform
Microsoft's push toward computer use automation lets agents interact with web and desktop applications even when a clean API is unavailable.
That means agents that can navigate legacy web portals, fill out vendor forms, or extract data from tools that have no API — all from inside a Microsoft 365 environment your clients already pay for.
- Who it's for: MSP technicians and internal AI champions at M365-licensed clients, especially those with legacy line-of-business apps.
- How long a first test takes: 60–90 minutes to build a basic screen-navigation task in Copilot Studio's visual canvas.
- First step: In your Copilot Studio tenant, open Agent Settings → Computer Use and confirm CUA is enabled. Then create a new agent, add a computer-use action, and point it at a real internal task — data entry into an old web form is a perfect first proof of concept.
If you run Claude automations, check which ones move to the new credit pool on June 15. Anthropic is moving Agent SDK use, the claude -p command, Claude Code GitHub Actions and third-party agent apps out of the regular subscription pool. Miss it and automations can stop or bills can jump. Ask whoever built your automations to count last month's programmatic calls.
*For MSPs:* A free 30-minute Claude usage audit that shows overage exposure and model-routing options is a useful first conversation.
Don't treat Gemini 3.5 Flash benchmark claims as settled fact yet.
All of Google's benchmark numbers for 3.5 Flash are self-reported — Google evaluated its own model. Independent validation from Epoch AI or comparable third parties isn't available yet.
Wait for third-party replications before migrating production agentic workflows — or at minimum, run your own task-specific evaluation before pitching clients on a Gemini-based agent stack.
The "agentic tax" is coming for every AI platform — not just Anthropic.
What Anthropic just codified with its June 15 billing split is the beginning of a universal repricing of AI agent infrastructure.
Companies are spending dramatically more on AI even as per-unit costs drop — average enterprise AI spending jumped from $1.2M in 2024 to $7M in 2026, a 483% increase, while per-token API prices fell roughly 280x over the same period.
The reason:
an agent working through a multi-step task requires the model to re-process the entire conversation history before generating each next action, and re-sent context accounts for 62% of total agent inference bills.
Every AI vendor is looking at the same math. The era of flat-rate subscriptions subsidizing heavy agentic workloads is ending — and the MSPs and AI champions who help clients understand *actual* agent cost structures before the invoices arrive will be the ones clients trust when the next billing surprise lands.