Google AI Studio can write subtitles as text, not put them on your video. As of 17 November 2026, Google's Gemini API documentation says Gemini models take video and audio files, can transcribe the speech, and can add timestamps in MM:SS form. Ask for that transcript in SRT layout, save the reply as a .srt file, and you have a subtitle file you can upload to YouTube or burn in with an editor. The documentation does not describe any way to get back a video with captions on it.
This post is built from Google's own developer documentation, read on 17 November 2026. We did not open AI Studio or run a prompt for it, so nothing here is a test result. Google AI Studio is also not the Gemini app: the app's upload limits are in can ChatGPT, Gemini or Claude add captions to a video. AI Studio is the free developer playground at aistudio.google.com, and the limits below are the ones Google documents for the Gemini models behind it.
What Google's documentation says (17 November 2026)
| Question | What the docs say | Page |
|---|---|---|
| Does Gemini take video files? | Yes: MP4, MPEG, MOV, AVI, FLV, MPG, WebM, WMV and 3GPP | Video understanding |
| Longest video | Up to 3 hours at low media resolution, 1 hour at high, on models with a 1M context window | Video understanding |
| Largest uploaded file | 2 GB per file on the free tier; files are deleted after 48 hours | Video understanding, Files API |
| Does it take audio files? | Yes: WAV, MP3, AIFF, AAC, OGG, FLAC, M4A, Opus, WebM and more | Audio understanding |
| Longest audio | 9.5 hours of audio per prompt | Audio understanding |
| Can it transcribe? | Yes, "Transcription and translation (speech to text)" is a listed use | Audio understanding |
| Timestamp format | MM:SS, for example 01:15 | Video and audio pages |
| How fine are video timestamps? | "Timestamps are added every second" | Video understanding |
| Word-level timestamps | Only on the Gemini 3.5 Transcribe model, audio only, up to 30 minutes | Audio transcription |
| Structured output | JSON that follows a schema you give; Google says to validate the values | Structured outputs |
| Returns a captioned video? | Not described anywhere in these pages; the output is text | All of the above |
| Cost | "Google AI Studio usage is free of charge in all available regions" | Pricing |
Can you upload a video or audio file in Google AI Studio?
Yes, with a caveat about what we could confirm. Google's file input guide ends by telling readers to "try writing your own multimodal prompts using Google AI Studio", and the quickstart describes AI Studio as the place to "quickly try out models and experiment with different prompts". The quickstart does not show the upload control itself, and we did not open AI Studio, so we cannot tell you where the button sits or whether the screen sets a smaller file limit than the API does.
What the documentation does state is what the models accept. For video, the listed formats are MP4, MPEG, MOV, AVI, FLV, MPG, WebM, WMV and 3GPP. Models with a 1M context window "can process videos up to 3 hours long by default (at low media resolution), or up to 1 hour long at high media resolution". The video page puts uploaded files at 2 GB on the free tier and 20 GB on the paid tier, and the Files API page says files "are stored for 48 hours". The same video page says the Gemini API also accepts a public YouTube link in place of a file, as a preview feature, limited to 8 hours of YouTube video a day on the free tier.
For audio, the listed formats are WAV, MP3, AIFF, AAC, OGG, FLAC, MPEG, M4A, L16, Opus, ALAW, MULAW and WebM, and the stated maximum is "9.5 hours of audio per prompt". Google notes that audio is downsampled to 16 Kbps and that multi-channel audio is combined into one channel.
Two rules to know before you sign in: Google's troubleshooting page says "You must be 18 or older to use Google AI Studio", and AI Studio returns an access error outside its supported regions.
Can Gemini give a transcript with timestamps?
Yes, according to the docs, and the format they describe is MM:SS. The audio page says "To get a transcript, ask for it in the prompt", and gives the example "Provide a transcript from 02:30 to 03:29." The video page says to "use the MM:SS format (e.g., 01:15 for 1 minute and 15 seconds)" when you refer to a moment in a video. Google's own transcription example asks the model to "Provide accurate timestamps for each segment (Format: MM:SS)", along with speaker labels and an English translation when the speech is not in English.
That is the gap between a Gemini transcript and a subtitle file. An SRT cue needs a start and an end time written as hours, minutes, seconds and milliseconds, such as 00:00:01,000 --> 00:00:04,200. Google's video page says that in the default processing mode "Timestamps are added every second". Nothing in the documentation promises timing finer than a second for the general Gemini models. You can ask for SRT layout and the model can write text in that shape, but treat the milliseconds as formatting, not as a measurement, until you have checked them against the video.
What the docs say about structured output
The AI Studio quickstart says the Run settings panel lets you switch on structured output. With it, in the words of the structured outputs page, "You can configure Gemini models to generate responses that adhere to a provided JSON Schema." Google's transcription example uses this to get a list of segments, each with a speaker, a timestamp and the text.
This helps if you plan to convert the result to SRT with a script, because every segment arrives in the same shape. It does not make the timings correct. The same page says: "While output is syntactically correct JSON, always validate values in your application." For a one-off video, plain SRT text in the reply is simpler.
Gemini 3.5 Transcribe: word-level timestamps, with limits
Google's documentation also lists a separate speech-to-text model, Gemini 3.5 Transcribe, with a "Try in Google AI Studio" button on its model page. It is the one place in these docs where precise timing is described: it "generates precise start and end time offsets for each recognized word", and the sample output shows offsets such as 0.100s and 0.400s.
The limits matter for subtitles:
- It takes audio, not video. Export the audio from your video first.
- Audio can be up to 1 hour, but "is limited to 30 minutes when features like speaker diarization or word-level timestamps are enabled".
- Longer audio is cut off at the limit while the request still reports that it completed. Split a long recording into parts of 30 minutes or less, or you will lose the end without an error.
- Google's own note says "Enabling word-level timestamps may degrade overall transcription accuracy."
- The word timings come back as data attached to the response, not as an SRT file. Turning them into cues is your job or a script's.
We could not confirm from the documentation whether the AI Studio screen offers the word-timestamp setting or whether it is only available through the API.
What Google AI Studio costs
Google's pricing page says "Google AI Studio usage is free of charge in all available regions." The billing page adds that it stays free "unless users link a paid API key for access to paid features", and that new accounts start on a free tier with "access to certain models", up to each model's free rate limits. On 17 November 2026 the pricing page lists Gemini 4 Hydrogen and Gemini 3.5 Transcribe as "Free of charge" on the free tier, and says Gemini 4 Carbon and Gemini 4 Argon are not available on it. We did not find the free request limits on a public page; Google shows them in a signed-in dashboard.
The free tier has one condition creators should read. The pricing page lists "Content used to improve our products" under Free and "Content not used to improve our products" under Paid. If your video is unreleased or shows a client, that line matters.
For scale, here is what the paid tier would charge. Google counts a video at about 100 tokens per second at the default resolution, so a 10-minute video is about 60,000 input tokens. At Gemini 4 Hydrogen's listed $1.50 per million input tokens that is about 9 US cents, plus the output tokens for the transcript at $7.50 per million. That sum is our arithmetic from Google's figures, not a quoted price.
What Google AI Studio cannot do for subtitles
- It does not return a captioned video. Every page we read describes text or JSON coming back. Putting the words on the picture needs YouTube, an editor or a caption tool.
- It does not save a .srt file for you. The documentation describes text responses. You copy the text into a file yourself.
- It does not promise frame-accurate timing. The general models work with MM:SS and timestamps "added every second"; the word-level model is audio only and capped at 30 minutes.
- It is not a subtitle editor. There is no timeline where you drag a cue or see the caption over the video. You fix timings by reading numbers.
- It is not meant for live captions. For real-time transcription the audio page points to Google's Live API and the Cloud Speech-to-Text API.
- It does not keep your uploads. Files sent through the Files API are deleted after 48 hours.
How to get an SRT file from Google AI Studio
Because we did not run this, read it as a workflow built from the documentation, not a recipe we timed.
- Open aistudio.google.com, sign in with a Google account that meets the age rule, and start a new chat prompt.
- Pick a model your account can use on the free tier, and add your video or audio file to the prompt.
- Ask for the subtitles. A starting prompt: "Transcribe the speech in this file as an SRT subtitle file. Number each cue. Use HH:MM:SS,mmm --> HH:MM:SS,mmm for the times. Keep each cue to two lines of about 40 characters and under 5 seconds. Write the words in the language and script spoken. Return only the SRT text."
- Copy the reply into a plain text editor and save it as a .srt file in UTF-8. What an SRT file is shows the exact layout if the file needs repair.
- Play the video with the file loaded and check the first cue, a cue in the middle and the last cue against the speech. If the timing slides as the video goes on, fixing subtitles that drift out of sync covers the repair.
- Read every line for wrong words, names and numbers.
- Use the file. YouTube accepts basic .srt files in plain UTF-8, with no styling, as described in how to add subtitles to a YouTube video. To burn the subtitles into the picture, follow burning subtitles with HandBrake, VLC or FFmpeg.
The prompt is a starting point. A model can return cues that overlap, run past the end of the video, or skip a stretch of speech, and it will not warn you. Steps 5 and 6 are not optional. If YouTube or your editor rejects the file, why an SRT file does not work lists the usual causes.
Translation, Hinglish and Indian languages
Translation is where AI Studio does something a caption burner like CapsAI does not. Google lists "Transcription and translation (speech to text)" as a use of Gemini's audio understanding, and its example prompt asks for an English translation of non-English speech. If you need English subtitles for a Hindi video, ask for both the original and the translation, then have a speaker of the language check the result. Translating Hindi video subtitles to English explains what to look for.
For mixed Hindi and English speech, the Gemini 3.5 Transcribe page says it handles "intra-sentence and inter-sentential code-switching without manual configuration". Its language table includes Hindi, English (India), Bengali, Gujarati, Kannada, Malayalam, Marathi, Oriya, Punjabi, Telugu and Assamese. We did not find Tamil or Urdu in that table on 17 November 2026, and the page does not say which script it uses for Hindi words inside English sentences. The general Gemini model pages we read list no languages at all. So tell the model the script you want, Roman letters or Devanagari, and check the output. We did not test any language. If a Hinglish subtitle file is what you need, Hinglish SRT generator options compares the routes.
Google AI Studio or CapsAI
They produce different things. AI Studio gives you text that you turn into a subtitle file. CapsAI gives you a finished video with the captions burned in, and no subtitle file.
Pick Google AI Studio if:
- you need an SRT file, for YouTube closed captions or for an editor;
- you need a translation, which CapsAI does not do;
- your video is longer than 10 minutes, or you only have an audio file;
- you are comfortable fixing timestamps by hand and saving text files.
Pick CapsAI if the goal is styled captions on a Reel, a Short or a TikTok. CapsAI is a browser tool, on phone or desktop. You upload one MP4 or MOV of up to 10 minutes, it writes the captions from the speech in 11 spoken languages (Hinglish in Roman script, English, Hindi, Bengali, Telugu, Marathi, Tamil, Urdu, Gujarati, Kannada and Punjabi), you fix words in the editor over the video, choose a caption style such as Story, Base or Pop Out, and export one MP4. There is no prompt to write and no file to convert. A video of up to 3 minutes needs no sign-up to try; you sign in when you export. The 3 free minutes to start cover generating and editing; exporting needs a one-time pack: ₹49, ₹99 or ₹299 for 20, 60 or 180 minutes in India ($4, $9 or $19 elsewhere). Nothing renews, and watermark-free export comes with every paid pack.
What CapsAI does not do: it has no SRT, VTT or transcript export, no translation, no audio-only upload, no link import and no API. If any of those is the job, AI Studio or a subtitle-file tool is the right pick. For speech that mixes Hindi and English, the Hinglish caption generator is where to start.
What to do next
If you need a subtitle file, try the workflow above on a one-minute clip first and check every timestamp before you trust it on a long video. If you need captions on the video itself, skip the file and use the auto subtitle generator. SRT vs VTT vs TXT covers which file a platform wants, and burned-in vs closed captions helps you decide which of the two you need.
