Add/update skill card: claude-opus-4-8
Browse files- claude-opus-4-8.json +36 -0
claude-opus-4-8.json
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": "claude-opus-4-8",
|
| 3 |
+
"provider": "anthropic",
|
| 4 |
+
"capabilities": [
|
| 5 |
+
"coding",
|
| 6 |
+
"creative_synthesis",
|
| 7 |
+
"instruction_following",
|
| 8 |
+
"math_reasoning",
|
| 9 |
+
"planning_agentic",
|
| 10 |
+
"world_knowledge"
|
| 11 |
+
],
|
| 12 |
+
"skill_vector": [
|
| 13 |
+
0.9,
|
| 14 |
+
0.9,
|
| 15 |
+
0.89,
|
| 16 |
+
0.95,
|
| 17 |
+
0.85,
|
| 18 |
+
0.93
|
| 19 |
+
],
|
| 20 |
+
"source": "benchmark",
|
| 21 |
+
"confidence": [
|
| 22 |
+
"high",
|
| 23 |
+
"low",
|
| 24 |
+
"medium",
|
| 25 |
+
"high",
|
| 26 |
+
"medium",
|
| 27 |
+
"high"
|
| 28 |
+
],
|
| 29 |
+
"imputed_capabilities": [
|
| 30 |
+
"creative_synthesis"
|
| 31 |
+
],
|
| 32 |
+
"support": null,
|
| 33 |
+
"subset_hash": null,
|
| 34 |
+
"date": "2026-07-14",
|
| 35 |
+
"notes": "Benchmark-derived skill vector, recalibrated 2026-07-14 by a 5-model council then cross-model reconciled. Sourcing per dimension: coding=SWE-bench Verified 88.6%; instruction_following=estimated from 4.5/4.6 lineage (IFEval ~92, MMLU-Pro ~89) tempered by AA IFBench 62.2%; math_reasoning=AIME 2026 95.7% + USAMO 2026 96.7%; world_knowledge=GPQA Diamond 93.6%; planning_agentic=Terminal-Bench 2.1 74.6% + MCP-Atlas 82.2% (the tau2 Telecom 94.4% figure was de-weighted as display-only); creative_synthesis=imputed. Opus 4.8 sits just below the newer flagship Fable 5 overall but leads it on refresh-era math (USAMO/AIME 2026) and ties/edges it on GPQA. The matrix was rescaled to widen tier separation for routing. A `measured` extraction on the frozen probe set overrides this. See README.md."
|
| 36 |
+
}
|