Unverified50% confidenceBenchmarkExact time
The same model's scores on DeepCWE range from 65.5 to 74.2 depending on the agent framework used; the official harness uses the upper end of the range, while under Claude Code's harness the score is only 69.8.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/27/2026
First Seen
Valid until: 12/26/2026
Sources
Related Entities
Related Claims
UnverifiedThe same DeepSeek v1.1 model scored between 69.8% and 74.2% on Terminal Bench 2.0 under different harnesses.69% similarUnverifiedIn the MoE architecture, the Mixtral 8x7B model has a total of 46.7B parameters, but only approximately 12.9B parameters (2 experts) are activated per token, making inference costs close to those of a 13B dense model while achieving performance comparable to the 70B level.62% similarVerifiedOn the DeepSWE benchmark, GPT-5.6 scored 45 at Low, 61 at Medium, 69 at High, 71 at X-High, and 73 at Max across reasoning tiers.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/954587API
curl https://kongchang.com/api/v1/knowledge/claims/954587MCP
get_claim(id=954587)