Add Opus 5 and Kimi K3 benchmark scores

src/xrouter_llm/resources/config/models/claude-opus-5.yaml

@@ -0,0 +1,22 @@
+# One model per file. Load the registry with: --benchmark-profiles config/models
+# Benchmarks are PUBLISHED PERCENTAGES (0-100); featurizer divides >1 by 100.
+# Only sourced/official values are active; unverified numbers stay in comments.
+# Researched 2026-07-26.
+model_id: anthropic/claude-opus-5
+provider: anthropic
+aliases: ["claude-opus-5"]
+release_date: "2026-07-24"
+source_quality: official        # specs/pricing are official; gpqa/lcb are AA/Vals
+source_urls:
+  - https://www.anthropic.com/news/claude-opus-5
+  - https://platform.claude.com/docs/en/about-claude/models/overview
+  - https://openrouter.ai/anthropic/claude-opus-5
+  - https://artificialanalysis.ai/models/claude-opus-5
+  - https://www.vals.ai/models/anthropic_claude-opus-5
+context_length: 1000000
+max_output_tokens: 128000
+input_cost_per_1k: 0.005
+output_cost_per_1k: 0.025
+benchmarks:
+  gpqa_diamond: 93.2            # Artificial Analysis (93.232%, adaptive, max effort); Vals 93.434%
+  livecodebench: 89.0           # Vals LCB v6 (89.033%, max effort); AA has not run LCB

src/xrouter_llm/resources/config/models/moonshotai-kimi-k3.yaml

@@ -1,7 +1,7 @@
 # One model per file. Load the registry with: --benchmark-profiles config/models
-# Benchmarks are PUBLISHED PERCENTAGES (0-100). Served via OpenRouter. Researched 2026-07-18.
+# Benchmarks are PUBLISHED PERCENTAGES (0-100). Served via OpenRouter. Researched 2026-07-26.
 # Moonshot's launch table does not report GPQA Diamond or LiveCodeBench; the
-# capability score below is from Artificial Analysis' independent API run.
+# capability scores below are from independent AA/Vals API runs.
 model_id: moonshotai/kimi-k3
 provider: moonshotai
 aliases: ["moonshotai/kimi-k3", "kimi-k3"]
@@ -12,11 +12,11 @@ source_urls:
   - https://platform.kimi.ai/
   - https://openrouter.ai/moonshotai/kimi-k3
   - https://artificialanalysis.ai/models/kimi-k3
+  - https://www.vals.ai/models/kimi_kimi-k3
 context_length: 1048576
 parameters_b: 2800
 input_cost_per_1k: 0.003        # $3.00 / 1M cache miss; $0.30 / 1M cache hit
 output_cost_per_1k: 0.015       # $15.00 / 1M
 benchmarks:
   gpqa_diamond: 93.5            # Artificial Analysis (93.535%, max reasoning)
-# livecodebench: pending (AA reports no result as of 2026-07-18). Do not
-# substitute DeepSWE, FrontierSWE, Program Bench, or other agentic coding scores.
+  livecodebench: 87.2           # Vals LCB v6 (87.185%); AA has not run LCB

tests/test_yaml_profiles.py

@@ -49,7 +49,7 @@ def test_shipped_models_registry_loads() -> None:
     from xrouter_llm.paths import default_models_dir
 
     catalog = load_benchmark_profiles(default_models_dir())
-    assert len(catalog) == 14
+    assert len(catalog) == 15
     # model_id is the canonical OpenRouter slug; the bare id stays as an alias.
     opus = catalog.get("anthropic/claude-opus-4.8")
     assert opus.provider == "anthropic"
@@ -60,8 +60,12 @@ def test_shipped_models_registry_loads() -> None:
     assert catalog.get("z-ai/glm-5.2").benchmarks["livecodebench"] == 69.5
     assert "livecodebench" not in catalog.get("deepseek/deepseek-v4-flash").benchmarks
     assert catalog.get("openai/gpt-5.5").provider == "openai"
-    # 2026-07 additions: latest gemini flash/pro/flash-lite, sonnet 5, Kimi models
+    # 2026-07 additions: latest Gemini, Claude, and Kimi models
     assert catalog.get("claude-sonnet-5").model_id == "anthropic/claude-sonnet-5"
+    opus_5 = catalog.get("claude-opus-5")
+    assert opus_5.model_id == "anthropic/claude-opus-5"
+    assert opus_5.benchmarks["gpqa_diamond"] == 93.2
+    assert opus_5.benchmarks["livecodebench"] == 89.0
     assert catalog.get("google/gemini-3.5-flash").benchmarks["gpqa_diamond"] == 92.2
     assert catalog.get("google/gemini-3.5-flash").benchmarks["livecodebench"] == 87.6
     assert catalog.get("google/gemini-3.1-pro-preview").provider == "google"
@@ -72,7 +76,7 @@ def test_shipped_models_registry_loads() -> None:
     assert kimi_k3.model_id == "moonshotai/kimi-k3"
     assert kimi_k3.parameters_b == 2800
     assert kimi_k3.benchmarks["gpqa_diamond"] == 93.5
-    assert "livecodebench" not in kimi_k3.benchmarks
+    assert kimi_k3.benchmarks["livecodebench"] == 87.2
     # superseded models are removed from the registry
     removed = {"google/gemini-2.5-flash-lite", "anthropic/claude-sonnet-4.6"}
     assert removed.isdisjoint({p.model_id for p in catalog.profiles()})