diff --git a/.github/workflows/integration-test.yml b/.github/workflows/integration-test.yml
index 5df61fc4c..8ceb9dacc 100644
--- a/.github/workflows/integration-test.yml
+++ b/.github/workflows/integration-test.yml
@@ -954,11 +954,14 @@ jobs:
         # pinned claude package's own installer, which fetches the native
         # binary from the vendor's release endpoint — the same vendor trust
         # executing the agent itself carries. Keep the pins in lockstep with
-        # run-predict, run-evaluate, and run-backtest.
+        # the workflows that install the same CLIs: run-predict and
+        # run-evaluate (gemini), run-backtest (all three), and run-analytics's
+        # labeler (claude). Production claude cells run the CLI their action
+        # bundles instead, which the engine-actions-smoke leg exercises.
         run: |
           set -euo pipefail
           case "$ENGINE" in
-            claude-code) package="@anthropic-ai/claude-code@2.1.259" binary=claude ;;
+            claude-code) package="@anthropic-ai/claude-code@2.1.280" binary=claude ;;
             codex) package="@openai/codex@0.144.1" binary=codex ;;
             gemini) package="@google/gemini-cli@0.49.0" binary=gemini ;;
           esac
@@ -2868,7 +2871,7 @@ jobs:
         # expressible for `npm i -g`.
         run: |
           set -euo pipefail
-          npm install --silent --no-audit --ignore-scripts --global @anthropic-ai/claude-code@2.1.259
+          npm install --silent --no-audit --ignore-scripts --global @anthropic-ai/claude-code@2.1.280
           node "$(npm root -g)/@anthropic-ai/claude-code/install.cjs"
           claude --version
           path=$(command -v claude)
diff --git a/.github/workflows/run-analytics.yml b/.github/workflows/run-analytics.yml
index 24d3ec2a2..fb297ba82 100644
--- a/.github/workflows/run-analytics.yml
+++ b/.github/workflows/run-analytics.yml
@@ -139,6 +139,7 @@ on:
           - claude-haiku-4-5-20251001
           - claude-sonnet-4-6
           - claude-fable-5
+          - claude-opus-5-5
         default: claude-haiku-4-5-20251001
       # corpus-stats inputs (ignored by every other mode). All are optional; an
       # empty value means "no filter" / "no grouping".
@@ -1233,9 +1234,11 @@ jobs:
       # instead — the same npm pattern run-backtest.yml and integration-test.yml
       # prove on these runners, kept in version step with both — and hand the
       # action the executable, which skips its own install path entirely. The
-      # action bundles a slightly newer CLI for the cells it installs itself;
-      # if the labeler ever misbehaves under this pin, matching the bundled
-      # version is the first lever.
+      # pin runs ahead of the CLI the action bundles, because it is the floor
+      # the newest `label_model` choice needs (an older CLI refuses the model
+      # id). If the labeler ever misbehaves under this pin, the first lever is
+      # an action release whose bundled agent SDK matches it — not a rollback,
+      # which would drop that choice.
       - name: Install the pinned Claude Code CLI for the labeler
         id: claude-cli
         timeout-minutes: 5
@@ -1248,7 +1251,7 @@ jobs:
         # executing the agent itself carries.
         run: |
           set -euo pipefail
-          npm install --silent --no-audit --ignore-scripts --global @anthropic-ai/claude-code@2.1.259
+          npm install --silent --no-audit --ignore-scripts --global @anthropic-ai/claude-code@2.1.280
           # Claude's native binary is placed by its postinstall, which
           # --ignore-scripts suppresses; run the pinned package's own
           # installer explicitly.
diff --git a/.github/workflows/run-backtest.yml b/.github/workflows/run-backtest.yml
index a136f7ef2..e8c4c4e19 100644
--- a/.github/workflows/run-backtest.yml
+++ b/.github/workflows/run-backtest.yml
@@ -457,7 +457,7 @@ jobs:
         run: |
           set -euo pipefail
           npm install --silent --no-audit --ignore-scripts --global \
-            @anthropic-ai/claude-code@2.1.259 @openai/codex@0.144.1 \
+            @anthropic-ai/claude-code@2.1.280 @openai/codex@0.144.1 \
             @google/gemini-cli@0.49.0
           # Claude's native binary is placed by its postinstall, which
           # --ignore-scripts suppresses; run the pinned package's own
diff --git a/docs/pipeline.md b/docs/pipeline.md
index b8d8de25e..8b6ed81b7 100644
--- a/docs/pipeline.md
+++ b/docs/pipeline.md
@@ -471,6 +471,7 @@ queues behind the production run of the same mode. The modes:
   the measured block still reaches the step summary, and the job fails. The
   `label_model` dispatch input picks the labeler's model; a ceiling-sized run
   overrides the default for `claude-fable-5`, the one tier measured to finish.
+  `claude-opus-5-5` is offered too, with no measured pace or cost yet.
   See [qp-topic.md](qp-topic.md).
 
 ## `integration-test` — the infrastructure preflight