GPT-5.6 in M365 Copilot — and What It Costs Anthropic

GPT-5.6 Is Now the Default Model in Microsoft 365 Copilot

OpenAI announced on July 9th that GPT-5.6 Sol is now the preferred model powering Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat, and Cowork. This is not a preview toggle — it is the production default for tenants already on paid Copilot plans.

If you have a governance posture built around predictable model behavior — system prompts tuned to earlier GPT-4o outputs, sensitivity label workflows validated against a specific response pattern — you need to re-validate now, not at your next quarterly review. Sol is a frontier-class reasoning model. It will produce longer, more inferential outputs than your users have seen before. That changes how your DLP rules fire, how your Purview audit trails read, and how your helpdesk volume will shift in the first two weeks of rollout.

What to do: Pull a sample of Copilot interaction logs from the past 30 days and run them through Sol manually before your next change window. Look specifically at outputs touching HR documents, legal templates, and anything that touches classified or FOUO content categories. The model upgrade is silent — your users didn't get a dialog box.

Fable Gets Another Extension — This Time Because of Sol

The same Sol release is why Anthropic extended Claude Fable 5 access again. As Simon Willison reported, Anthropic has pushed the Fable cutoff on all paid Claude Max plans to July 19th. The original rationale was compute constraints, but Willison's framing is sharper: Anthropic is responding competitively to Sol landing at the same capability tier.

The extension keeps Claude Code's weekly rate limits 50% higher through July 19th, which matters directly for engineering teams running Claude Code on agentic workflows. Fable remains capped at half your weekly usage limit; beyond that you're burning usage credits or dropping to another model.

For clients with Claude Code deployed at scale, the July 19th cliff is a real operational date. Background agent jobs that queue over the weekend of July 18th–19th could hit degraded throughput if your team hasn't planned for the model transition. Set a calendar reminder now, not after your on-call gets paged.

Claude Code Ships Four Releases in One Week

Between July 6th and July 11th, Claude Code shipped v2.1.202, v2.1.203, v2.1.205, and v2.1.207. The individually unremarkable version numbers obscure how much moved.

The operationally significant changes:

• Auto mode is now generally available on Bedrock, Vertex AI, and Foundry (v2.1.207). Previously this required the CLAUDE_CODE_ENABLE_AUTO_MODE environment variable opt-in. If you have Claude Code deployed in a Bedrock enclave with locked-down IAM policies, auto mode is now active by default unless you explicitly set disableAutoMode in settings. Check your configurations before this bites you in a regulated environment.
• A security consent bug was fixed (v2.1.207): remote managed settings from non-interactive claude -p SDK runs were being permanently recorded as consented without ever showing the security consent dialog. If you've been running headless Claude Code pipelines in production, your audit trail may show phantom consents. Review your SDK invocation logs.
• Dynamic workflow size controls landed (v2.1.202): a new /config setting lets operators set small/medium/large agent count guidelines for dynamic workflows. This is the first explicit operator-level throttle on multi-agent spawning. For clients running Claude Code in cost-sensitive CI pipelines, set this to small and raise it deliberately.
• OpenTelemetry workflow attributes added (v2.1.202): workflow.run_id and workflow.name are now emitted on telemetry from workflow-spawned agents. If you have a SIEM or OTel collector ingesting Claude Code data, update your dashboards to capture these — they're the keys you need to reconstruct full multi-agent session trees.

GCC-High Finally Gets Copilot Usage Reporting

The Microsoft 365 Roadmap confirmed this week that GCC-High and DoD tenants will receive a dedicated Microsoft 365 Copilot usage report in the admin center, GA in July 2026. A companion Copilot Chat usage report covering GCC, GCC-High, and DoD follows the same timeline.

This has been a governance gap that's forced federal clients into PowerShell workarounds and manual Graph API pulls for months. The new reports include total enabled and active users, per-app breakout, and anonymized user-level activity insights — which maps directly to what ISSO teams need for their ATO evidence packages.

If you're a federal client mid-deployment, start building your reporting cadence against these native reports now rather than hardening your custom scripts further. The native report structure will align with what CISA and agency IG offices will eventually expect to see in Copilot governance documentation.

Copilot Studio: SharePoint Lists as a Knowledge Source

The Microsoft 365 Roadmap added SharePoint lists as a native knowledge source for Copilot Studio agents, entering preview in July 2026 with GA targeted for September. This closes a meaningful gap: agents previously had to extract list data through Power Automate flows or connector hops.

The practical use case is structured operational data — procurement tables, asset inventories, project trackers — that teams currently maintain in SharePoint lists precisely because they don't want a full Dataverse schema. Grounding an agent directly against that data without a pipeline in the middle reduces both latency and the synchronization failure modes we've seen in several client deployments.

One caution: SharePoint list permissions are column-level only in limited scenarios. Confirm that your list's item-level permissions model is what you actually intend before you let an agent ground against it at query time. A misconfigured list permission becomes a data exposure surface the moment the agent starts serving answers across role boundaries.

---

Watch next week for whether Anthropic announces Fable's permanent availability or access terms post-July 19th — that decision will set the compute cost baseline for every enterprise Claude Code contract going into Q3.