Unverified50% confidenceBenchmarkExact time
On the ARC-AGI-3 benchmark, the same GPT-5.6 model with only two configuration items in the Codex Harness changed saw its score rise from 13.3% to 38.3%, with output tokens reduced by a factor of six.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/11/2026
First Seen
Valid until: 12/10/2026
Sources
Related Entities
Related Claims
Unverified在完全相同的模型下，仅靠Harness策略优化，ARC-AGI-3得分从13.3%提升到38.3%，同时输出Token节省约六倍72% similarUnverifiedOpus 5在ARC-AGI 3上的得分是次优模型的三倍68% similarUnverifiedThe strongest configuration, GPT-5.5 paired with Codex, achieved a pass rate of only 28.0% on Harbor-Index67% similarUnverifiedThe original Codex achieved a pass@1 score of 28.8% on the HumanEval benchmark, whereas using GPT-3 directly yielded only 0%.66% similarUnverifiedThe median output token count per problem-solving attempt for Opus 4.7 is 60,000, while GPT 5.5 requires only 16,000.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/908811API
curl https://kongchang.com/api/v1/knowledge/claims/908811MCP
get_claim(id=908811)