diff --git a/.github/workflows/evaluation.yml b/.github/workflows/evaluation.yml
index 06e20b291b..79848eac6d 100644
--- a/.github/workflows/evaluation.yml
+++ b/.github/workflows/evaluation.yml
@@ -18,11 +18,11 @@ on:
       model:
         description: 'Copilot model to use'
         required: false
-        default: 'claude-opus-4.5'
+        default: 'claude-opus-4.6'
       judge-model:
         description: 'Judge model to use for evaluation'
         required: false
-        default: 'claude-opus-4.5'
+        default: 'claude-opus-4.6'
       runs:
         description: 'Number of runs per test'
         required: false
@@ -51,9 +51,10 @@ concurrency:
 env:
   GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
   DOTNET_VERSION: '10.0.x'
-  DEFAULT_MODEL: 'claude-opus-4.5'
-  DEFAULT_JUDGE_MODEL: 'claude-opus-4.5'
-  DEFAULT_RUNS: '3'
+  DEFAULT_MODEL: 'claude-opus-4.6'
+  DEFAULT_JUDGE_MODEL: 'claude-opus-4.6'
+  # Use more runs for main branch to get more stable benchmark data, but fewer runs for PRs to reduce quota pressure.
+  DEFAULT_RUNS: ${{ github.ref == 'refs/heads/main' && '5' || '3' }}
   DEFAULT_PARALLEL_SKILLS: '3'
   DEFAULT_PARALLEL_RUNS: '3'
 
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index cf1965c72f..3b3ab9cf71 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -224,12 +224,15 @@ node dist/index.js --tests-dir ./src/dotnet-msbuild/tests ./src/dotnet-msbuild/s
 node dist/index.js --tests-dir ./src/dotnet-msbuild/tests ./src/dotnet-msbuild/skills/common-build-errors
 
 # Fewer runs for faster iteration (default is 5)
-node dist/index.js --runs 1 --tests-dir ./src/dotnet-msbuild/tests ./src/dotnet-msbuild/skills
+node dist/index.js --runs 3 --tests-dir ./src/dotnet-msbuild/tests ./src/dotnet-msbuild/skills
 
 # Use a specific model
-node dist/index.js --model claude-sonnet-4.5 --tests-dir ./src/dotnet-msbuild/tests ./src/dotnet-msbuild/skills
+node dist/index.js --model claude-opus-4.6 --tests-dir ./src/dotnet-msbuild/tests ./src/dotnet-msbuild/skills
 ```
 
+> [!WARNING]  
+> If you share the results in a Pull Request, make sure to have `--runs` configured to at least 3 but better 5 for reliable results.
+
 ### CI evaluation
 
 Tests run automatically on pull requests that modify files under `src/`. The evaluation workflow discovers changed components and runs the skill-validator for each one. Results are posted as a PR comment and uploaded as build artifacts.