Google closes the multimodal loop with Gemini 3.5 Transcribe and the GA release of Omni 1.1 Flash for video
On 26 August 2026, Google made Gemini 3.5 Transcribe generally available, two dedicated speech-to-text models with diarization and custom vocabulary, and on 27 August it shipped Gemini Omni 1.1 Flash, its conversational video generation model with interpolation and 4K output. Transcription is no longer a feature of the generalist model — it is a standalone product. Here is what that changes for teams that transcribe or produce video.
26 August 2026. Google makes Gemini 3.5 Transcribe available, two dedicated speech-to-text models. 27 August 2026. Gemini Omni 1.1 Flash goes generally available for conversational video generation and editing. 30 September 2026. The old gemini-omni-flash-preview endpoint is deprecated. Net result: in two days, Google turns what used to be a feature of the generalist Gemini model into a family of standalone products — transcription, video, voice.
Transcription becomes a product, not a feature
For years, transcription was a capability of the main model. With Gemini 3.5 Transcribe, Google inverts the logic: it extracts its audio understanding from Gemini and turns it into two dedicated speech-to-text models.
The first, gemini-3.5-transcribe, targets high-accuracy, low-latency non-streaming transcription. It handles 85+ languages, utterance-based language detection, speaker diarization, word-level timestamps and custom vocabulary biasing of up to 1,000 terms. The second, gemini-3.5-transcribe-live, targets low-latency bidirectional streaming over WebSockets via the Live API, with interim and finalized events, a Smart transcription mode and multiple voice activity detection (VAD) strategies.
The positioning is unmistakable: Google is attacking the specialist transcription market head-on, where Whisper, AssemblyAI and Deepgram hold sway. The 1,000-term custom vocabulary is a direct pitch to jargon-heavy sectors — healthcare, legal, industrial — where generic models choke on proper nouns and technical terms.
Omni 1.1 Flash: conversational video hits GA
On 27 August, Google made Gemini Omni 1.1 Flash generally available under the identifier gemini-omni-1.1-flash. It is the stable release of the “conversational” video generation and editing model that had been running in preview. Three new capabilities come with the GA.
The first is video extension: extending an existing clip by generating a continuation at the end, via the extend task or directly with a prompt. The second is interpolation between a first and a last frame — the image_to_video task accepts up to two images and generates the transition. The third is resolution control: a new resolution parameter in video_config accepts 360p, 720p (default), 1080p and 4K, with 1080p and 4K produced via upscaling.
The detail that matters for developers: the gemini-omni-flash-preview endpoint will be deprecated on 30 September 2026. Anyone who integrated the preview must migrate to gemini-omni-1.1-flash before that date or face a broken integration.
A multimodal strategy that is consolidating
These two announcements are not isolated. They fit a rapid sequence: Gemini 3.7 Flash went GA on 13 August for coding and agents, Gemini 3.6 Flash and 3.5 Flash-Lite landed in July for efficiency and cost. Each release specializes one modality or one workload profile.
The through-line is specialization. Google no longer offers a single “can do everything” model: it splits its capabilities into dedicated models — transcription, video, code, agents — each with its own endpoint, price and support window. For an engineering team, that means an architecture decision now sits where a single API call used to suffice.
The consumer side follows the same logic with Gemini Live, the voice-first experience built into the Gemini app. Voice becomes a first-class mode of interaction, not a side feature.
The competition now turns on use cases
The transcription market is currently split between three families of players. On one side, Whisper and its open-source derivatives offer high accuracy with no licensing cost, but leave the team to host, orchestrate and tune it themselves. On the other, specialist APIs such as AssemblyAI and Deepgram sell integration simplicity and business-ready features. Google arrives third with an asset neither can easily replicate: the native audio understanding of Gemini, already trained on a massive multimodal corpus.
What now separates the field is no longer the generic error rate, but the features that pull transcription into a real product: diarization that assigns each sentence to a speaker, word-level timestamps that enable navigation, and custom vocabulary that pins down the spelling of domain terms. That is exactly the ground Gemini 3.5 Transcribe claims.
The timing matters. August 2026 has been one of the densest months in the history of model releases — more than ten launches in roughly twenty days, by several counts. In that flood, differentiation shifts from “who has the biggest model” to “who solves the specific use case”. Dedicated transcription is a direct answer to that pressure.
What it changes for teams that transcribe
For a team that processes audio, Gemini 3.5 Transcribe shifts the economics of transcription. The selection criteria are no longer raw accuracy alone, but diarization, word-level timestamps and custom vocabulary — the three levers that determine whether a transcript is actually useful in a product.
The 1,000-term vocabulary biasing is the most concrete point. If your product processes drug names, legal references or part identifiers, a model that spells them correctly on the first pass saves you an entire post-processing layer. That is exactly what this release promises.
The Live variant opens a different frontier: real-time assistants, live captioning, call centres. Interim events and VAD control are primitives teams previously had to assemble themselves.
And for those that produce video
On the video side, Omni 1.1 Flash in GA raises two concrete questions. The first is technical: interpolation between two images and clip extension change the nature of assisted editing. You no longer start from text to generate a scene — you start from an existing fragment to extend it or link it to another.
The second is budgetary: 4K output via upscaling and the resolution controls give you a clear cost ladder — generating in 720p and upscaling to 4K is often cheaper than generating natively at high resolution. That is a direct optimisation lever for production pipelines.
The 30 September date imposes discipline: preview integrations have a firm migration deadline. A team that leaves a gemini-omni-flash-preview integration lying around courts an announced outage.
Verdict
If you transcribe audio at scale, evaluate Gemini 3.5 Transcribe on your own recordings, measuring accuracy on your domain’s terms — the 1,000-term custom vocabulary is the differentiator, not the generic benchmark. The Live variant deserves a prototype if you are targeting real time.
If you produce generated video, migrate to gemini-omni-1.1-flash now and use interpolation and extension to cut the cost of your sequences. Do not let a preview integration reach 30 September: the deprecation date is firm, and the migration is trivial if done in time.