Claude Fable 5.1, Copilot Studio Human Approval, and GCC-High Gaps Closing

Claude Fable 5.1 Is Now the Default Model in Claude Code

Claude Code v2.1.257 shipped Monday and made Fable 5.1 the default Fable model — 1M context window, same $10/$50 per million token pricing as Fable 5, with cache reads at $0.25/Mtok. According to Simon Willison's writeup, Anthropic's headline number is a 52.6% score on the new Terminal-Bench-Science 0.1 benchmark, up from 24.7% for Fable 5. That benchmark skews toward multi-step reasoning under real execution environments — exactly what your engineering teams are running against production codebases in Claude Code sessions.

What changed: Fable 5.1 is the new default; you don't opt in, you opt out. Any Claude Code deployment that hasn't pinned a model ID is now running 5.1.

Why it matters: The 1M context window changes the economics of large-repo work. We've had clients hit the wall mid-session on a 200K-seat engineering org's monorepo. That ceiling just moved. The pricing didn't.

What to do: If you have cost controls or evals pinned to Fable 5 behavior, audit before the next release cycle. If you're running Claude Code in an enterprise managed deployment, confirm your CLAUDE_CODE_SUBAGENT_MODEL_FORCE settings account for the new model ID (claude-fable-5-1).

The same release also added a Containment Escape rule to auto mode: cloud metadata-credential fetches, egress evasion attempts, and cross-tenant reach are now denied by default unless your environment explicitly marks them expected. For any client running Claude Code in a shared CI environment or multi-tenant cloud, this is a meaningful security floor that was previously absent from auto mode.

---

Copilot Studio Gets Per-Tool Human Approval Gates

The Microsoft 365 Roadmap item 570434 published September 1 is the governance story Copilot Studio has needed since agents started touching live systems. Makers can now toggle a per-tool, per-agent approval gate. When a gated tool fires — sending an email, closing a ticket, processing a payment — the agent pauses, surfaces an approval request inline in Teams or the channel where it's deployed, and waits for a human to approve, approve for the session, or deny.

What changed: Approval is deterministic and independent of the agent's instructions. The model can't talk its way past a gated tool.

Why it matters: Every client I've talked to this year has the same blocker on agentic Copilot Studio rollout: "What stops it from sending the wrong thing to the wrong person?" This answers that question with a hard mechanism, not a prompt engineering suggestion. It's the difference between a demo and a production deployment for anything touching finance, HR, or customer comms.

What to do: This is GA in September. Start mapping your agent's tool inventory against risk tiers now so you're ready to configure gates on day one, not after your first incident.

---

Three GCC-High Gaps Close in September and October

Three roadmap items landed the same week that matter specifically to federal and regulated clients:

• Viva Insights and Copilot Analytics in GCC-High — GA September CY2026. Privacy-protected, role-based Copilot adoption and usage analytics, finally available in the GCC-High enclave. If you've been flying blind on Copilot adoption in a GCC-High tenant, this closes that gap.
• Organizational Data Service in GCC-High — GA September CY2026. HRIS connector ingestion for org/HR data, unified for Copilot and M365 apps. Relevant for any GCC-High deployment that needs org-chart context in Copilot responses.
• Copilot Chat in Teams for GCC, GCC-High, and DoD — GA October CY2026. Chats, channels, calls, and meetings — the full Teams Copilot Chat surface, not a subset.

What to do: If you have a GCC-High Copilot rollout that stalled waiting for analytics parity, the September GA gives you the reporting layer to restart that conversation with leadership. The October Teams item also means your end-user training materials can stop carrying caveats about which surfaces work in GCC-High.

---

Organization-Level Compliance Controls Surface in Both SDKs

Both the Anthropic TypeScript SDK v0.123.0/0.124.0 and Python SDK v1.3.0/1.4.0 shipped this week with overlapping enterprise governance additions: named types for organization compliance settings state, Claude Tag category and user breakdowns in usage reports, and workspace ID support on more endpoints.

What changed: You can now reference compliance settings by named type rather than raw string — which matters for any IaC or policy-as-code layer you're building on top of the API. The usage report additions (tag category, user breakdowns) mean you can finally attribute API consumption to teams or cost centers without building your own tagging middleware.

What to do: If you're managing a multi-workspace Anthropic deployment and reporting usage to business units, the v0.124.0/v1.4.0 usage report fields are worth a sprint to wire into your existing FinOps dashboards.

---

Claude Code Gets Org-Managed MCP Server Delivery

Claude Code v2.1.259 added managedMcpServers as an organizational managed setting. Admins can now push HTTP/SSE MCP servers to every user in the organization — same entry shape as .mcp.json — without requiring users to configure anything locally. Entries specifying a command to run are skipped (HTTP/SSE only), which is the right constraint for a managed fleet.

What changed: MCP server distribution shifts from per-developer configuration to org-level policy.

Why it matters: For a 5,000-seat engineering org, the difference between "each developer configures their own MCP servers" and "we push the approved server list from a managed setting" is the difference between a pilot and a governed rollout. This is the missing piece for treating Claude Code like a managed enterprise tool rather than a developer's personal install.

What to do: If you're scoping a Claude Code enterprise deployment, block out time this sprint to define your approved MCP server list and test the managedMcpServers delivery path before you go wide.

---

Watch next week for GPT-6 Astra's broader API rollout — Simon Willison flagged it's priced identically to Fable 5/5.1 and scoring above Fable on OpenAI's own benchmarks, which means the model selection conversation with your engineering teams is about to get louder.