Speculation Surrounds Projected Capabilities of Unreleased Anthropic Model
Discussions within AI research communities have intensified following unverified calculations regarding the capabilities of “Model 4,” a designation used by the AI safety organization METR to refer to an unreleased model from Anthropic. According to data shared on Reddit and hosted on a personal GitHub repository, projections suggest the model may possess an Epoch Capabilities Index (ECI) of approximately 182.5.
The ECI is an aggregate benchmark score used to track AI progress, similar in concept to the Artificial Analysis Intelligence Index. The projected figure is derived from a mathematical correlation between the ECI and Anthropic’s own variant, the Anthropic ECI (AECI). The calculation, which suggests a high Pearson correlation coefficient ($r^2 = 0.997$), implies that the AECI and ECI are essentially rescaled versions of the same metric.
Discrepancies in Benchmark Projections
The projected ECI of 182.5 has drawn scrutiny from observers due to its significant scale. For comparison, users noted that Claude 3 Opus 5.5 reportedly has an AECI of 174.5 and an ECI of approximately 169. If the 182.5 figure is accurate, it would represent a substantial leap in capability, potentially exceeding the performance levels seen in other advanced models like Mythos during its training phase earlier this year.
However, some analysts have questioned the mathematical validity of the projection, noting that there is currently no clear evidence to suggest that the ECI should be higher than the AECI. The discrepancy between the two metrics remains a point of debate among those tracking model performance.
Implications for AI Safety and Competition
The discussion highlights ongoing debates regarding “sandbagging”—the practice of companies intentionally underperforming in public benchmarks to avoid scrutiny—and the difficulty of tracking rapid advancements in closed-source models. Some commentators suggest that because China cannot easily distill models that are not publicly available, the gap between leading Western models and international competitors may be widening.
While the projected leap in capability is significant, some observers suggest the practical impact may be limited to specific, time-sensitive domains such as cyberwarfare. In most other sectors, the lead held by a new model over its predecessors is expected to diminish as the broader industry catches up through distillation and further research.
At present, the ECI of 182.5 remains an unverified projection based on METR’s reported AECI data and has not been independently confirmed by Anthropic or other third-party benchmarking organizations.