MCP Lands in M365 Copilot as Claude Fable Resets Pricing Math

MCP Hits M365 Copilot in September — Governance Is Not Optional

The Microsoft 365 Roadmap entry for Federated Copilot Connectors confirms GA in September. That's five weeks away. What it means concretely: users can wire Copilot directly to third-party data sources — ServiceNow, Salesforce, whatever exposes an MCP server — using their own identity, with data retrieved in real time and never indexed in Microsoft's tenant. No copy stored. No graph ingestion lag.

The architecture is cleaner than the legacy Graph connector model, and the compliance story is genuinely different: because no customer data touches Microsoft storage, DLP policies that fire on indexed content don't apply here. That gap is where your compliance team will ask questions you need to be ready to answer. The admin center retains a kill switch per connector, and staged rollout is supported, but the default posture lets users self-authorize connections to any MCP server their identity can reach.

What to do before September: Audit which external SaaS environments your users have OAuth tokens for. Any system with an MCP server that a user can reach becomes a live Copilot data source the moment they authorize it. For GCC-High tenants, federated connectors are not in scope yet — confirm boundary before your communications go out.

The companion signal is the self-serve sync connector roadmap item, also GA in September. This one covers Jira and Confluence Cloud specifically, and it does index content — into the Microsoft 365 index, scoped to the user's own permissions. Two different plumbing paths, both landing the same month. Your helpdesk will see tickets on both if you don't get ahead of it.

Purview Retention for Copilot Memory Closes a Compliance Gap

Microsoft 365 Roadmap item 569612 adds retention and versioning for Copilot memory items — the saved preferences and inferred context that make Copilot interactions feel personalized. These items live in a hidden folder in the user's Exchange mailbox, and until this change, retention policies couldn't touch them. That was a problem for any org under legal hold obligations or e-discovery scope.

With this feature, admins can configure retention policies that preserve historical versions of memory items and surface them for investigation. The Exchange mailbox storage location means your existing Purview retention labels and litigation hold infrastructure applies — you're not building a new pipeline.

What to do: If you've deployed Copilot memory (or are planning to), add a retention policy scope for Copilot memory items before your next compliance review. Clients in financial services and healthcare should treat this as a prerequisite, not a nice-to-have.

Claude Fable Ends the "Wait for the Next Model" Strategy

Simon Willison's coverage of Drew Breunig's post surfaces something that matches what we're hearing from engineering teams: Fable is genuinely capable enough that teams can't cost-defer to a cheaper model for the hard tasks anymore. Breunig's framing is precise — prior to Fable, a new model at the same price would arrive and paper over context-engineering debt. That loop is broken. Fable is the ceiling, and it's expensive enough that the routing architecture matters.

The practical implication for a 200-engineer org running Claude Code: you cannot treat every task as a Fable task and stay inside budget. You need explicit model routing — Fable for architecture decisions and complex cross-file refactors, a cheaper model tier for test generation and boilerplate. That's not a product preference, it's a cost control requirement.

What to do: If your engineering team is on Claude Code with no model routing policy, build one before Q4 budget reviews. The ANTHROPIC_DEFAULT_MODEL environment variable added in Claude Code v2.1.236 is the lever — set the org-wide default to a cheaper model and require an explicit /model override for Fable. That override persists per session, so it's low friction for engineers who need it.

Claude Code v2.1.239: Data-Residency Workloads Now See True Cost

Claude Code v2.1.239 adds the 1.1× US-only-inference premium to cost estimates everywhere — the /cost command, the status line, and --max-budget-usd enforcement. If your team runs Claude Code in a data-residency workspace (common for clients with ITAR, FedRAMP, or contractual data-localization requirements), your displayed costs just went up without your usage changing.

This is accurate accounting, not a price increase. But it means any budget guardrails you set before this release were underreporting actual spend by that margin. Review your --max-budget-usd flags if you've deployed Claude Code in managed container environments with residency constraints.

The same release extends the fullscreen renderer to Bedrock, Vertex, and Foundry installs — worth knowing if your team is running Claude Code on Azure or AWS rather than direct Anthropic API.

Copilot Studio Evaluation Traces Are Worth a Second Look

The Copilot Studio evaluations improvement adds reasoning traces, cited knowledge sources, and cross-run comparison to the agent evaluation experience. For teams building agents in Copilot Studio, this is the gap between "the agent gave a wrong answer" and "here's which knowledge chunk it retrieved and why it ranked that way."

If you're running agents in production without a structured evaluation harness, this feature is a forcing function. The ability to generate test datasets from your own knowledge sources directly removes the most time-consuming part of evaluation setup.

What to do: Run an evaluation baseline before the next agent capability change you deploy. Comparison only works if you have a prior run to compare against.

---

Next week, watch for any Microsoft messaging around the September GA window — specifically whether federated connector governance controls ship complete or arrive post-GA as a follow-on update.