# Visual Studio 2026 v18.9: Copilot Thinking Effort Dial > Source: > Published: 2026-08-22 17:08:10+00:00 GitHub switched Copilot to token-based billing on June 1, and the reaction was not subtle. Developers reported bills jumping from $29 to $750 a month; one engineer documented a [$180 charge on the very first day](https://visualstudiomagazine.com/articles/2026/06/02/slammed-by-copilot-usage-based-billing-on-day-1-facing-180-bill-for-june.aspx) of the new model. Visual Studio 2026 v18.9, released August 11, is Microsoft’s direct response: a three-level thinking effort dial that puts token spend back under developer control. That is the headliner. But v18.9 also ships a Git review agent and org-level custom agent distribution that enterprise teams should not overlook. ## Set How Hard Copilot Thinks The new thinking effort control appears in the model picker and the expanded Model management view. You choose from three levels: **Low**— Quick responses with minimal reasoning. Right for autocomplete, simple questions, and routine suggestions. Burns the fewest tokens.**Medium**— The default. Balanced reasoning depth and speed for everyday coding tasks.** High**— Deep reasoning for complex algorithms, architecture decisions, and hard bugs. Costs more. The framing matters: this is not a quality dial, it is a cost dial. Setting it to Low does not make Copilot worse at simple tasks — it tells the model not to over-reason before answering. High effort is appropriate when you are genuinely stuck on something hard. Using it for boilerplate generation is burning tokens you do not need to burn. VS Code added a nearly identical control months ago. VS 2026 is finally catching up. The [official announcement](https://devblogs.microsoft.com/visualstudio/tell-your-model-when-to-think-harder/) frames this as developer empowerment, which it is — but it is also Microsoft acknowledging that AI reasoning has a meter running, and somebody has to watch it. ## Review Code Before the PR v18.7 added the ability to review pull requests inside Visual Studio without leaving the IDE. v18.9 moves one step earlier in the workflow with the Git review agent. Switch to the Git agent in Copilot Chat and ask it to review your uncommitted changes. It surfaces inline comments directly in the editor alongside a navigable list in the Git Changes panel. You can also attach a specific commit as context and ask for a review — the results appear in a commit-specific view, kept cleanly separate from your working files so there is no confusion about what is reviewed versus what is live. The conversation does not end there. After the initial review you can follow up, ask about specific findings, or work through suggested fixes in the same chat thread. The feature works with both GitHub and Azure DevOps repositories. The practical implication: you no longer have to push code to a remote branch or open a PR to get AI feedback. Catch issues in the Git Changes window while you are still in flow. ## Org Agents Reach Developers Automatically This one is primarily for enterprise teams, but it is genuinely useful once you understand what it does. GitHub organization and enterprise owners can now define custom Copilot agents and commit them to an org-level repository. Visual Studio detects them automatically when any developer opens a repository belonging to that organization — no individual configuration required. The agents appear in the agent picker. Hovering over one shows its description and which organization it comes from. A definition button opens the underlying agent file so developers can inspect what the agent is instructed to do. Platform teams can use this to distribute internal tools: a compliance checker, an internal API assistant, a codebase conventions enforcer. The agents reach every developer the moment they open the repo. ## The Model Picker Is Also Cleaner The model picker was redesigned to reduce clutter. You can pin frequently used models to the top and collapse ones you rarely use. Selecting Manage Models opens a full view showing cost information, capabilities, and context window size for each model — including custom models connected via Ollama. The Ollama addition is worth noting. The Language Models editor connects to your local Ollama instance, discovers every model you have pulled, and groups them by provider. Local models cost zero tokens. If your team is finding Copilot bills uncomfortable, switching routine tasks to a local Ollama model is a concrete cost-avoidance path that v18.9 makes easier to reach. Here is [more context on the billing situation that prompted these changes](https://devops.com/github-copilot-bills-hit-800-visual-studios-june-update-adds-real-time-usage-alerts-and-mcp-trust-checks/). ## Upgrade to 18.9.1, Not 18.9.0 v18.9.0 shipped with a scroll bug in the Copilot Command preview dialog — users could not scroll to the Allow or Deny buttons, which effectively blocked command approval. A second bug affected the Test Agent. Both were fixed in v18.9.1, released August 18. If you installed 18.9.0, update before using either of those features. The [full release notes](https://learn.microsoft.com/en-us/visualstudio/releases/2026/release-notes) cover both bugs and the complete feature list. The broader trajectory is clear: token-based billing forced Microsoft to ship cost-visibility tools in June, and now it is shipping cost-control tools in August. The thinking effort dial is the right feature at the right time. Use Low for routine work, Medium for most coding, and reserve High for the problems that genuinely need it. Your Copilot invoice will reflect the difference.