@@ -27,7 +27,7 @@ Paths in this table are relative to `../paper/`.
 | Framework, Figure 3 | `figures/mll/figure3_v2.png` → unchanged `public/lagen/assets/figures/framework.png` |
 | Measured latency structure, Figure 5 | `figures/fig5_profiling_structure/data.csv` and `plot.py` → `src/lagen/data/profiles.json`, `public/lagen/assets/data/profiles.json`, Canvas/SVG |
 | Held-out calibration, Figure 4 | `figures/fig4_sim2real_calibration/heldout_30cases_20261002/case_errors.csv` and `plot.py` → `src/lagen/data/calibration.json` and inline SVG |
-| Core results, Table 1 | Layout and formatting: `tables/table1_benchmark_overview/table.tex`, `table_caption.tex`, and `render_table.py`. VLA values: `figures/fig1_teaser/data.csv` → `src/lagen/data/results.json`. GPT values: `tables/table1_benchmark_overview/coding_agents.csv` → `src/lagen/data/coding-agents.json`. |
+| Core results, Table 1 | Layout and formatting: `tables/table1_benchmark_overview/table.tex`, `table_caption.tex`, and `render_table.py`. VLA values: `figures/fig1_teaser/data.csv` → `src/lagen/data/results.json`. GPT values: `tables/table1_benchmark_overview/coding_agents.csv` → `src/lagen/data/coding-agents.json`. Claude values: Table 1 of `public/lagen/assets/paper.pdf`, updated on 2026-10-18. |
 | Mean-delay/profile comparison, Table 2 | `tables/mean_delay_training/data.json` and `table.tex` |
 | Cross-delay transfer, matched-delay view of Figure 6 | `figures/fig6_latency_robustness/matched_delay/cells.csv` and `plot_matched_delay.py` → `src/lagen/data/latency-robustness.json` and interactive heatmaps |
 | Cross-task transfer, Figure 7 | `figures/fig21_demon_mixed_cross_task/data.csv` and `plot.py` → `src/lagen/data/cross-task.json` and inline SVG |
@@ -390,22 +390,23 @@ wait for an external font stylesheet before initializing its interactions.
 
 ## Core-results table
 
-The table follows Table 1 at paper revision
-`8f7996b91940746107ca1b9b323bd5238199c0ce`, verified on 2026-10-14.
+The table follows Table 1 of `public/lagen/assets/paper.pdf` (page 8),
+verified on 2026-10-18. The PDF is committed in website revision `7cdbb42`.
 Tasks occupy columns under Games and Embodied Control.
-Model/condition combinations occupy nine VLA rows and two GPT reference rows.
+Model/condition combinations occupy nine VLA rows and four coding-agent reference rows.
 VLA values remain unnormalized mean episode returns from `results.json`.
 GPT values retain source precision from `coding_agents.csv`.
+Claude Opus 5.5 and Claude Haiku 5.5 values use the precision displayed in the PDF.
 Its `Intercept Grab Fast` key uses the website's `Intercept Grab (Fast)` name.
 
 Values with absolute magnitude at least 1,000 use two decimals and `k`.
 Other values use one decimal. Decimal ties round away from zero, matching
 the paper renderer's `ROUND_HALF_UP`.
 Target/Real cells use `#F0EEF2`, bold returns and `#B34A4A` improvement arrows.
-Model cells stay white. GPT model names are italic, while their conditions
+Model cells stay white. Coding-agent model names are italic, while their conditions
 and values use regular weight and `#46414F`.
 
-Both GPT models use Low reasoning effort and simulated OpenVLA latency.
+All four coding agents use Low reasoning effort and simulated OpenVLA latency.
 The caption and footnote preserve these evaluation conditions and comparison scope.
 The table stays visible at every viewport size, with local horizontal scrolling.
 
 
@@ -33,7 +33,7 @@ const formatReturn = (value: number) => Math.abs(value) >= 1000 ? thousands.form
     Zero/Zero is the original policy without delay. Zero/Real runs that policy during live inference.
     Target/Real first trains with the matched profile and then evaluates in real time.
     Cells report mean episode return on each task's native scale.
-    Both GPT models use Low reasoning effort to approximate the rapid responses required in game environments.
+    All four coding agents use Low reasoning effort.
   </figcaption>
   <div class="table-scroll" tabindex="0" role="region" aria-label="Core results by model, condition, and task">
     <table aria-labelledby="core-results-caption" aria-describedby="core-results-note">
@@ -71,7 +71,7 @@ const formatReturn = (value: number) => Math.abs(value) >= 1000 ? thousands.form
       })}
     </table>
   </div>
-  <p id="core-results-note" class="table-note"><sup>†</sup> GPT results use simulated OpenVLA latency and are not directly comparable to the VLA results.</p>
+  <p id="core-results-note" class="table-note"><sup>†</sup> Coding-agent results use simulated OpenVLA latency and are not directly comparable to the VLA results.</p>
 </figure>
 
 <style>
@@ -81,7 +81,7 @@ const formatReturn = (value: number) => Math.abs(value) >= 1000 ? thousands.form
   figcaption strong { color: var(--ink); }
   .table-scroll { margin-top: 16px; }
   table { min-width: 1100px; table-layout: fixed; font-size: 15px; border-top: 2px solid #46414f; border-bottom: 2px solid #46414f; }
-  .model-column { width: 116px; }
+  .model-column { width: 138px; }
   .condition-column { width: 98px; }
   th, td { padding: 6px 3px; border: 0; text-align: center; }
   thead th { font-weight: 600; }
 
@@ -32,5 +32,39 @@
       "Balance": 42.448506682145236,
       "Intercept Grab (Fast)": 0.23094618071979767
     }
+  },
+  {
+    "model": "Claude Opus 5.5",
+    "returns": {
+      "Demon Attack": 235.0,
+      "Asterix": 560.0,
+      "Atlantis": 1300,
+      "AirRaid": 451.3,
+      "Deadly Corridor": 378.7,
+      "Flappy Bird": 5.7,
+      "Ant": 0.3,
+      "Hopper": 179.9,
+      "Inverted Pendulum": 8.9,
+      "Walker2D": 150.0,
+      "Balance": 44.7,
+      "Intercept Grab (Fast)": 0.6
+    }
+  },
+  {
+    "model": "Claude Haiku 5.5",
+    "returns": {
+      "Demon Attack": 82.3,
+      "Asterix": 0,
+      "Atlantis": 100.0,
+      "AirRaid": 0,
+      "Deadly Corridor": -44.2,
+      "Flappy Bird": 5.1,
+      "Ant": -203.1,
+      "Hopper": 98.6,
+      "Inverted Pendulum": 0,
+      "Walker2D": 33.3,
+      "Balance": 44.7,
+      "Intercept Grab (Fast)": 0.3
+    }
   }
 ]