Google Launches Agentic Video Understanding for Gemini Flash Models, Cutting Video Tokens by Up to 88%

Video has been the most expensive modality to reason over. A Gemini model handed a 90-minute lecture has, until now, ingested the whole thing at a fixed one frame per second, whether the question was ‘summarize this’ or ‘what time does the speaker switch to the pricing slide?’ That single-pass design forces a bad trade: pay for the full timeline in context, or pre-chunk the video and risk dropping the detail that mattered.

This week, Google launched agentic video understanding across its Flash models. Instead of ingesting the timeline, Gemini navigates it deciding what to watch, at what frame rate, and through which modality. Google reports up to 88% fewer tokens, up to 66% lower cost, and up to 7% higher accuracy on standard video benchmarks.

Subaru Is Dropping Gas Cars For EVs In Europe By The End Of 2027

Tesla Unveils Roadster With a Nod to SpaceX Merger

Is it deployable? Yes, but only as a hosted API feature. There are no open weights and nothing to self-host. It ships through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, works with both file uploads and public YouTube URLs, and bills at standard Gemini API token pricing with no additional feature fee.

What actually changed

Static processing, still the default on every Gemini model, extracts frames at 1 FPS in a single pass, processes audio at 1 Kbps single channel, and inserts timestamps every second. Agentic processing replaces that with a loop. The model pairs its own reasoning with native video tools to search, scan, and inspect target segments across frames, audio, and transcripts, loading only what the prompt requires. Developers could already assemble this by hand; the change is that Gemini runs the loop internally, which is where the development overhead disappears.

Across Google’s evaluations, Gemini 3.7 Flash with agentic understanding lands on the accuracy-to-cost Pareto frontier for video analysis among the models tested. The efficiency gains concentrate on long-form content, from 10-minute how-to guides to multi-hour recordings.

=60?(Math.floor(m/60)+’h ‘+(m%60)+’m’):m+’ min’); vs.textContent=fmt(stat)+’ tok’; va.textContent=”~”+fmt(ag)+’ tok’; fs.style.width=”100%”; fa.style.width=”12%”; note.innerHTML=’Static is exact: frames at 1 FPS (‘+(res===’low’?’66’:’258′)+’ tokens each) plus 32 tokens per second of audio and metadata, which Google documents as roughly ‘+per+’ tokens per second of video at ‘+res+’ media resolution. The agentic bar shows the best case Google publishes — up to 88% fewer tokens on long-form content. Real usage varies with the prompt and the navigation path, and reasoning tokens are billed separately as thought tokens.’; resize(); } dur.addEventListener(‘input’,calc); lo.addEventListener(‘click’,function(){res=”low”;lo.setAttribute(‘aria-pressed’,’true’);hi.setAttribute(‘aria-pressed’,’false’);calc()}); hi.addEventListener(‘click’,function(){res=”high”;hi.setAttribute(‘aria-pressed’,’true’);lo.setAttribute(‘aria-pressed’,’false’);calc()}); calc(); /* panel 3 */ var pick={},resEl=document.getElementById(‘res’); [].slice.call(document.querySelectorAll(‘.q’)).forEach(function(q){ var k=q.dataset.k; [].slice.call(q.querySelectorAll(‘.opt’)).forEach(function(o){ o.addEventListener(‘click’,function(){ [].slice.call(q.querySelectorAll(‘.opt’)).forEach(function(x){x.setAttribute(‘aria-pressed’,’false’)}); o.setAttribute(‘aria-pressed’,’true’); pick[k]=o.dataset.v; decide(); }); }); }); function decide(){ if(!(pick.len&&pick.need&&pick.lat)){return} var out; if(pick.need===’every’){ out=”Static. Frame-level precision across the whole clip is the case the docs reserve for static mode. It is also the only mode that accepts clipping offsets and a custom FPS value.”; }else if(pick.len===’short’&&pick.lat===’ttft’){ out=”Static. On clips under five minutes, agentic navigation adds internal reasoning and tool round-trips before generation starts, which can push time to first token up.”; }else{ out=”Agentic. Long-form content or a query aimed at specific moments is the case it was built for. Use streaming or background execution so long runs do not hit connection timeouts.”; } resEl.innerHTML=out; resize(); } /* panel 4 */ [].slice.call(document.querySelectorAll(‘#cards .card’)).forEach(function(c){ c.addEventListener(‘click’,function(){ var open=c.getAttribute(‘aria-expanded’)===’true’; c.setAttribute(‘aria-expanded’,open?’false’:’true’); setTimeout(resize,340); }); }); /* resize */ function resize(){ var h=document.body.offsetHeight+40; if(window.parent){window.parent.postMessage({mtpFrame:’gemini-agentic-video’,height:h},’*’)} } window.addEventListener(‘load’,resize); setTimeout(resize,120); setTimeout(resize,600); window.addEventListener(‘resize’,resize); })();

Credit: Source link