We present the results of a survey for extremely red objects (EROs) undertaken in the fields of 10 massive galaxy cluster lenses at z∼0.2, combining sensitive, high‐resolution Hubble Space Telescope imaging with deep, half‐arcsecond K‐band imaging from UKIRT. We detect 60 EROs with
(R−K)≥5.3, of which 26 have
(R−K)≥6.0 in a total image‐plane survey area of 49 arcmin2 down to
K=20.6, including one multiply imaged ERO. We use detailed models of the cluster lenses to quantify the lens amplification and thus correct the observed number counts and survey area for the effects of gravitational lensing. After making these corrections, we estimate surface densities at
K≤21.6 of
2.5±0.4 and
1.2±0.3 arcmin−2 for EROs with
(R−K)≥5.3 and ≥6.0 respectively. These ERO number counts agree with previous shallower surveys at
K≲19, and flatten significantly at magnitudes fainter than
K∼19–20. This flattening may be due to a transition from an ERO population dominated by evolved galaxies at
z∼1–2(K≲19.5) to one dominated by dusty starburst galaxies at
z>1(K≳19.5). We also compare our results with various model predictions, including a model that attempts to explain EROs in terms of a single population of elliptical galaxies formed at high redshift. We find that a formation epoch of
zf∼2.5 for this population matches the observed surface density of
(R−K)≥5.3 EROs quite well, and the
(R−K)≥6.0 sample less well. More sophisticated models, including semi‐analytic prescriptions, underpredict the ERO surface density by approximately an order of magnitude, suggesting that these models produce insufficient stars and/or dust at high redshift. This deficit of EROs appears to be a general problem with models that reproduce the median redshift from K‐selected redshift surveys. One possible explanation is that the current K‐selected redshift distribution may be incomplete beyond
z∼1.