Microsoft has quietly rotated the default model behind Copilot Studio more than once in the last eighteen months. The launch-era GPT-4 class model gave way to GPT-4o, then to GPT-4.1, and newer GA options including GPT-5 Chat, GPT-5.5 Chat, Claude Sonnet 4.6, and Claude Sonnet 5 now sit alongside it. Most enterprise teams noticed none of these transitions. That is exactly the problem.

A Copilot Studio model update is not a platform improvement you accept on faith. It is a release event, shipped by a vendor, into a runtime you thought was frozen. If you did not retest, you shipped a change you did not approve.

TL;DR

A Copilot Studio model update is a vendor-initiated release event that can silently alter agent behavior, tool selection, output format, refusal patterns, latency, and cost. Enterprise teams should treat every default-model movement with the same change-control discipline as an internal deploy, using Copilot Studio's native agent evaluations, multi-turn tests, and Power Platform automation to retest before accepting the new default.

Key Takeaways

• Microsoft periodically upgrades the Copilot Studio Default model tag, and retired models remain available only for a short grace period, typically around thirty days.
• Peer-reviewed research from Stanford and UC Berkeley documented GPT-4 accuracy falling from 84 percent to 51 percent on the same task in three months, with instruction-following degrading in parallel. Behavioral drift is real and measurable.
• Nine categories of regression risk apply to any Copilot Studio agent after a model swap, and each maps to a specific native test.
• Copilot Studio agent evaluations went GA in March 2026, and a Power Automate connector plus a Power Platform REST API make scheduled retesting achievable for low-code teams without a data-science function.
• Every model swap needs a documented disposition: accept, pin the retired model within the grace window, switch to an alternate GA model, or roll back and hold publish.

What Actually Changed, And Why It Counts As A Release

The Copilot Studio default model change pattern is now well established. Microsoft's own model-settings documentation defines the Default tag as periodically upgraded to the current recommended model, and the What's New in Copilot Studio ledger reads like a release calendar: GPT-4.1 general availability, GPT-5 Chat general availability, Claude Sonnet 4.6 and Opus additions, GPT-5.5 Chat in mid-2026, agent-to-agent protocol general availability in April 2026, agent evaluations general availability in March 2026, and Entra Agent IDs general availability in July 2026.

Each of those entries is a change to the runtime that executes production agents. Some are additive. Some move the Default. All of them can alter user-observable behavior in agents your business already depends on, and none of them file a change request against your internal change advisory board.

The retirement side is just as consequential. Once Microsoft promotes a new default, the prior model typically remains selectable for roughly thirty days before it disappears from the maker experience. That is your entire window to detect a regression, decide whether to pin the retired model, and prepare an alternative. Foundry's broader lifecycle policy is more generous, with a minimum sixty days of notice for GA model retirement and eighteen-month lifecycles for OpenAI models (twelve months for Anthropic, DeepSeek, Mistral, and Fireworks), but Copilot Studio makers experience the change at the platform surface, not at the Foundry layer.

The reason this matters is not theoretical. In How Is ChatGPT's Behavior Changing over Time?, researchers from Stanford and UC Berkeley showed that GPT-4's accuracy on a prime-versus-composite task collapsed from 84.0 percent in March 2023 to 51.1 percent in June 2023. Instruction-following degraded on the same schedule. Same service name, same API, materially different behavior. If you built a Copilot Studio agent against a model in one quarter and the vendor swapped it in the next, you inherited that drift whether you noticed or not.

The framing that unlocks executive support is simple. You would not tolerate a silent production push from an internal team. A vendor-initiated AI change control event deserves the same treatment. The vendor's improvement is the customer's regression event, and the burden of proof lives on the customer's side of the shared responsibility line.

The Copilot Studio Regression Risk Taxonomy

Generic AI agent behavioral drift commentary is not useful to a Copilot Studio maker. What is useful is a taxonomy that maps risk to the components you actually built. Nine dimensions cover the field, and each maps to a specific test you can run inside the platform.

Behavior categories that shift under a model swap

• Instruction adherence: Does the agent still follow the system prompt and topic instructions verbatim, or has it started paraphrasing, summarizing, or ignoring negative constraints. Test with multi-turn conversation tests that assert on required phrases and forbidden phrases.
• Tool and plugin selection: Does the orchestrator still choose the correct Power Automate flow, connector action, or MCP tool for a given intent. A shift here can silently reroute finance approvals, ticket creation, or lookups to the wrong destination.
• Knowledge-source grounding fidelity: Does the agent still ground answers in SharePoint, Dataverse, or configured knowledge sources, or has it become more willing to answer from parametric memory. Watch for citation drop-off in agent analytics.
• Topic routing: Does natural language still trigger the intended topic, or has intent classification shifted so that queries fall into the fallback topic or the wrong topic entirely.
• Entity extraction: Do custom entities and prebuilt entities still resolve at the same precision. Model swaps commonly change how loosely a value is coerced.
• Output format: Does the agent still produce the JSON structure, table, or bulleted response your downstream automation expects. Format drift breaks integrations without breaking the agent.
• Refusal behavior: Has the model become more or less willing to answer sensitive queries, and does that align with your policy. Both directions are risks.
• Latency: Has time-to-first-token or full-turn latency moved. Voice agents and real-time agents are especially exposed.
• Per-turn cost: Has Copilot Credit consumption moved. Standard-rate models bill roughly fifteen credits per thousand tokens, while premium deep-reasoning models can run six to seven times higher. A verbosity shift in the new default alone can move a monthly bill materially.

Treat the taxonomy as a checklist. Any agent that touches money, customer data, external systems, or regulated content should be evaluated against all nine categories before you accept a new default. Our field notes on enterprise AI agent evaluation go deeper on the test design side.

The Silent Release Framework: Detect, Classify, Retest, Dispose

Knowing that model swaps are release events is not the same as operating that way. The BabyBots Silent Release Framework gives a Copilot Studio program four repeatable steps to run every time Microsoft moves the Default. It works for a two-person IT team and scales to an enterprise center of excellence.

Detect

Subscribe to the What's New in Copilot Studio feed and the Microsoft 365 message center. Assign one owner in the center of excellence to triage every entry within one business day and flag any item that changes the model list, moves the Default tag, or retires a model. Every flagged entry becomes a candidate release event with a ticket number, an owner, and a due date. Nothing about this step requires new tooling.

Classify

Maintain a live inventory of every Copilot Studio agent with three attributes: which model it is running (Default or pinned), its business criticality, and its downstream dependencies. Agents that call other agents via the agent-to-agent protocol, invoke MCP servers, or feed automations through Power Automate carry cascade risk. A model swap in a shared orchestrator agent propagates to every connected agent, so classify by graph, not by individual bot.

Retest

This is where Copilot Studio agent regression testing lives, and the native tooling is now sufficient for most enterprise teams. Copilot Studio agent evaluations reached GA in March 2026, along with multi-turn conversation tests that assert on both intermediate turns and final outcomes. Build your golden set by exporting question and reaction lists from the agent analytics surface, which gives you real production traffic instead of hand-written synthetic prompts.

For scheduled and automated runs, the Copilot Studio connector for Power Automate exposes Evaluate Agent, Get Agent Test Run Details, Get Agent Test Runs, Get Agent Test Sets, and Get Agent Test Set Details as no-code actions. Teams with a DevOps function can wire the same operations through the Power Platform REST API into a pipeline, gating publish on evaluation pass rate. Either path gives you a repeatable Copilot Studio agent evaluation for model swap workflow that runs on trigger, on schedule, or on demand.

Dispose

Every candidate release event ends with a written disposition on each affected agent. There are four options. Accept the new default when regression tests pass and cost and latency stay within tolerance. Pin the retired model when a critical regression appears and you need time, using the thirty-day grace window to plan a controlled migration. Switch to an alternate GA model, such as Claude Sonnet 4.6 or 5 for balanced reasoning, GPT-5 Reasoning or Claude Opus for deep analytical tasks, when the new default underperforms on your workload. Roll back or hold publish for agents that fail evaluation outright, and treat the hold as a real incident, not a note in a spreadsheet.

The Governance Record: Making the Retest Auditable

A retest that is not documented did not happen. Copilot Studio now provides the identity and audit primitives to make a model-swap disposition auditable end to end, and enterprise teams should wire them together deliberately.

Entra Agent IDs reached GA in July 2026 and are created automatically for every new agent. That gives each agent a first-class identity in Entra, which can be scoped through Conditional Access policies, governed by DLP data policies at the environment level, and traced through Purview and Microsoft Sentinel audit logs. When you run Retest Copilot Studio agents after model change, attach the evaluation run ID, the golden-set version, the model version tested, the disposition, and the approver to the agent's governance record. That record is what internal audit and any external assessor will ask for.

The maker community needs its own communication surface. A short, standing template works well: a one-paragraph description of what Microsoft changed, the affected agents by name, the retest owner, the disposition deadline, and the fallback options available inside the grace window. Sending the same shape of message every time builds pattern recognition in a distributed maker population and stops citizen developers from learning about the change from a user complaint.

The Executive Checklist

Hand this to your Copilot Studio team the day a model change is announced.

1. Open a release ticket for the model change and assign a named owner.
2. Pull the current inventory of agents, tagged by Default versus pinned, business criticality, and downstream dependencies.
3. Freeze publish on business-critical agents until retest completes.
4. Run agent evaluations and multi-turn conversation tests on the golden set for every affected agent.
5. Compare results against the nine regression risk categories: instruction adherence, tool selection, grounding, topic routing, entity extraction, output format, refusal, latency, and cost.
6. Record a written disposition per agent: accept, pin, switch, or roll back.
7. For any pin, calendar the thirty-day grace expiration and schedule the migration decision.
8. Update agent documentation and internal knowledge base to reflect the current model.
9. Log the change, the retest evidence, and the approver in the governance record for audit.
10. Send the maker community update using the standing communication template.

None of these steps require a data-science hire or a custom test harness. All of them require discipline and an owner. Our overview of Copilot Studio capabilities in 2026 covers the surrounding platform context that makes this discipline executable.

The organizations that will get durable value from Copilot Studio are not the ones with the newest model. They are the ones that treat every vendor-initiated change with the same seriousness as their own deploys. At BabyBots, we build and operate Copilot Studio agents in production, and the pattern is consistent across engagements: the teams that ship reliable agents are the teams that retest on schedule, disposition every model swap in writing, and refuse to let the Default tag move their runtime without approval. A model swap is a release event. If you did not retest, you shipped a change you did not approve. The most dangerous release is the one you did not schedule.