AECI
A proprietary benchmark suite used by Anthropic for internal model evaluation, covering dimensions such as reasoning, coding, long context and tool use. It sets pass-rate thresholds as the basis for dividing capability tiers and is not publicly released.
Timeline (last 90 days)
According to the Meter report, versions such as Model 2, Model 3 and Model 4 exist inside Anthropic, and Model 4 has already fully exceeded 50% on Anthropic's internal benchmark AECI.
AECI is a proprietary benchmark set Anthropic uses for internal model evaluation; it is not publicly released and covers dimensions such as reasoning, coding, long context and tool use.
METR released an independent evaluation of Anthropic, stating that its models' capability improvement on well-defined tasks is about twice the previous rate from March 2024 to October 2025, and that AECI benchmark scores have been growing at 1.9 times the previous rate since October 2025.
Meter's assessment found that the rate of capability improvement of Anthropic's models on well-defined tasks (such as tasks related to the AI R&D process) has reached about twice that of some previous historical interval, and the current rate may be closer to 1.3 to 3 times the historical rate.
All Facts (4)
METR released an independent evaluation of Anthropic, stating that its models' capability improvement on well-defined tasks is about twice the previous rate from March 2024 to October 2025, and that AECI benchmark scores have been growing at 1.9 times the previous rate since October 2025.
60%UnverifiedAccording to the Meter report, versions such as Model 2, Model 3 and Model 4 exist inside Anthropic, and Model 4 has already fully exceeded 50% on Anthropic's internal benchmark AECI.
50%UnverifiedAECI is a proprietary benchmark set Anthropic uses for internal model evaluation; it is not publicly released and covers dimensions such as reasoning, coding, long context and tool use.
50%UnverifiedMeter's assessment found that the rate of capability improvement of Anthropic's models on well-defined tasks (such as tasks related to the AI R&D process) has reached about twice that of some previous historical interval, and the current rate may be closer to 1.3 to 3 times the historical rate.
50%