Oscillatory neuronal activity above 0.1 Hz plays an important role in coordinating the function of large-scale brain networks. fMRI has typically been used to measure slow timescale neural activity, as the hemodynamic response is generally thought to be sluggish. However, recent studies have suggested that the BOLD signal may contain faster dynamics than previously realized [1,2]. Here we use ultra high field imaging to test the limits of this temporal resolution, and study the physiological basis of detectability of high frequency oscillations using fMRI. Five subjects gave informed consent and were scanned on a 7T Siemens whole-body scanner with a custom-built 32-channel head coil array. Each session began with a 0.75 mm isotropic multi-echo MPRAGE. Functional runs were acquired as single-shot gradient-echo blipped-CAIPI SMS-EPI [3] with 15 oblique slices with 2 mm isotropic resolution targeting the calcarine sulcus (R=2 acceleration, MultiBand factor=3, matrix=120x120, CAIPI shift=FOV/3, TR=227 ms, TE=24 ms, echo-spacing=0.59 ms, flip angle=30°). Stimuli consisted of a 12 Hz counterphase flickering radial checkerboard displayed for 4 minutes. The luminance contrast of the stimulus oscillated at either 0.2 Hz or 0.75 Hz. The first run with a stimulus oscillating at 0.2 Hz was used as a functional localizer to identify the stimulated region of V1, which was then taken as the ROI for analysis of subsequent runs. This localizer run was also used to define the phase of individual response, with early and late-responding voxels defined as the 33% earliest and 33% latest voxels. A control ROI in gray matter that was not visually driven was manually defined. Data were slice-timing corrected, motion corrected, and high-pass filtered, and triggered responses to the oscillatory stimulus were computed by upsampling the BOLD timecourse and averaging it on every cycle of the oscillation. The response of linear canonical models was measured by convolving the stimulus function with the SPM hemodynamic response function. The balloon model was simulated using equations from [4], and the parameter set drawn from [5], using Matlab.The stimulus oscillating at 0.75 Hz resulted in a robustly detected BOLD oscillation in V1, with a steady-state amplitude 1.3% the size of the response to a 0.2 Hz stimulus (Fig. 1a,b). No oscillation was detected in a nearby gray matter ROI that was not visually driven (Fig. 1c,d), suggesting that the oscillation reflected neural activity rather than motion or physiological noise. The linear canonical model predicted that this steady-state oscillation would be 0.14% the size of the response at 0.2 Hz, indicating that these results were an order of magnitude larger than predicted by standard models (Fig. 2e). Simulations with the balloon model demonstrated that two factors could contribute to the larger response at high frequencies. When the time constant for the viscoelastic effect (tau_v) is set to large but physiologically plausible values [4,6], brief stimuli elicit narrower BOLD responses, which lead to increased high-frequency content (Fig. 2b-c). In addition, the presence of continuous ongoing activity could lead to higher baseline flow and thus a lower mean transit time, which would also cause a faster BOLD response (Fig. 2d). Finally, we conducted additional analyses examining individual voxel responses and observed that voxels with early and late responses could be consistently observed at both stimulus frequencies. The lag did not hinder detection of oscillations at 0.2 Hz (Fig. 3a) but introduced substantial phase cancellation at 0.75 Hz (Fig. 3b). The delay found in individual voxels did not vary smoothly across cortex (Fig. 3c), suggesting that spatial smoothing or low spatial resolution scans can interfere with the ability to detect oscillatory BOLD responses at high frequencies. Analyzing the early-responding voxels alone increased the amplitude of the detected oscillations by 60%, relative to averaging across the ROI. We found that fMRI can detect oscillations of up to at least 0.75 Hz in individual subjects and in single sessions. The fMRI response was ~10 times larger than predicted by canonical linear models, suggesting that alternate hemodynamic models should be used when neural activity is continuous and rapidly varying, rather than driven by long stimuli and long interstimulus intervals. Previous studies have demonstrated that individual voxels can exhibit relative vascular delays of hundreds of milliseconds [7] and our results suggest that these delays can introduce phase cancellation of these oscillations, suggesting an additional benefit to using small voxel sizes to study these effects. These data suggest that fMRI could be used to localize 0.1-0.75 Hz oscillations in the human brain, and help identify neurovascular dynamics that can be taken into account when analyzing fast fMRI data.