review

Gemini Omni Review: What Our Own 40-Prompt Benchmark Found

Gemini Omni scored 4.82/5 in our testing, fourth of 16 models, with real synced audio on every lip-sync prompt and its weakest results on hands.

By Mara Voss · 24 September 2026 · Updated 21 October 2026

Gemini Omni scores 4.82/5 in our benchmark, fourth of 16 models and 0.02 behind the three models tied at the top. Its strongest category is lip-sync: it produces real spoken audio matching the requested line, synced to the mouth on screen. Its weakest is hands.

What we actually measured

We ran Gemini Omni through our fixed battery of 40 prompts across 10 categories, scored blind against the same rubric we use for every model. It came out at 4.82/5 overall, fourth of 16, behind Kling 3.0 Turbo, FLUX 3 and Seedance 2.0, which share first place at 4.84.

Every dimension sits at 4.8 or above. The weak point is a single category, hands, which pulls the overall score down by almost a tenth of a point.

By category

Hands is the lowest-scoring category. On the chopping prompt half of the pepper vanishes after the first cut, on the finger-counting prompt the hand flips over midway, and on the guitar prompt the shot cuts twice to a new framing during what should be one close-up. Liquids and animals come next, and those are middling rather than broken. The condensation on a melting-ice-cube prompt formed faster than physically plausible, a galloping horse showed some foot-sliding and a stiffer torso than a real gait would have, and a cat jumping onto a counter had a slightly floaty apex mid-leap. None of these were the kind of outright failure we have flagged in other models' runs, like an animal never leaving the ground or a hand losing a finger.

Lip-sync

Lip-sync is the highest-scoring category in our data for Gemini Omni (4.99/5). Many models still fail it: some generate a silent clip for a spoken-line prompt, some animate a mouth with no audio behind it, and Kling 3.0 produces no audio at all despite third-party marketing describing native dialogue.

Gemini Omni generated real spoken audio for all four of our lip-sync prompts, and the words matched what we asked for. A woman delivered our exact test line to camera with her mouth movements synced to the audio. A two-speaker exchange had both people say their assigned lines with natural back-and-forth timing. A singing prompt produced sustained, on-pitch vowel sounds synced to visible mouth shapes, not just a mouth flapping to a beat. We confirmed the audio track itself was real, not silent, on all 40 clips by checking the actual waveform, not just trusting a metadata flag.

This is Gemini Omni's clip for the direct-address lip-sync prompt.

Gemini Omni was the first model in our benchmark to clear this bar. Later models, including FLUX 3 and Vivix W1, have matched it.

What else stood out

• The 360-degree camera orbit completed for the first time in our testing. Every other model we have run either stops partway around the subject or breaks continuity mid-turn. Gemini Omni's orbit prompt passed behind the subject and returned to the front.
• The product-rotation prompt also completed a full turn, with the mug's handle disappearing and reappearing on the correct side, something none of the other models in our round managed cleanly.
• The multi-shot day-to-dusk prompt actually changed the lighting. Every other model we have tested on this exact prompt renders one continuous shot or a static frame instead of the requested lighting change between cuts.

This is the 360-degree orbit prompt, which Gemini Omni completed.

Who should use it, and what to check first

• Anything that needs a character speaking on camera: it reliably produces matching audio and lip movement, which many models still do not. If dialogue is the point of the clip, it belongs on the shortlist.
• Camera moves that need to complete, like a full orbit or a full product turn: strongest result we have measured for either.
• Close-up hands: its weakest category in our data, with an object that vanishes mid-cut and a hand that flips over. Worth a test clip before committing if the shot depends on hands, and the same goes, to a lesser degree, for liquids and animals.

Full clips, per-prompt scores and every category breakdown are on the Gemini Omni model page. The lip-sync leaderboard shows how every model we have tested compares on audio specifically. Methodology, including how we score and verify findings before publishing them, is on the methodology page.

Related: Google Gemini Omni

Mara Voss: Runs the prompt battery and scores every clip before it is published. Came to AI video testing from QA automation, where she spent years building test suites designed to break things on purpose.