Gemini 3.7 Flash is now generally available, alongside new video controls in Gemini Omni 1.1 Flash

Google released Gemini 3.7 Flash to the Gemini API on 13 August 2026, just three weeks after Gemini 3.6 Flash shipped. The short turnaround is deliberate: Google says the update came directly from developer feedback combined with algorithmic improvements, and the benchmarks back that up. Then, on 27 August 2026, Google put Gemini Omni 1.1 Flash into public preview, bringing a meaningful set of video production controls to the API. Two separate releases, two separate audiences, but both worth understanding if you build with Google’s models.

What changed in Gemini 3.7 Flash

The headline improvements are in software engineering, web development, and multi-step agentic work. On the FrontierCode 1.1 benchmark, 3.7 Flash scores 43.6% against 34.4% for its predecessor. On Code Arena for web development it reaches 1588 Elo. For agentic tasks, it scores 65.3% on DeepSWE v1.1, up from 49.0%. These are not marginal gains.

The model supports a 1 million token context window, up to 64k output tokens, and accepts text, images, video, audio, and PDF as inputs. Thinking depth is configurable via a thinking_level parameter set to LOW, MEDIUM (the default), or HIGH. The knowledge cutoff is March 2026.

The pricing window you should not ignore

Google is offering Gemini 3.7 Flash at introductory pricing through 31 December 2026: $0.75 per million input tokens and $3.75 per million output tokens. On 1 January 2027, those rates double to $1.50 and $7.50 respectively. Context caching sits at $0.075 per million tokens during the introductory period, with cache storage at $0.50 per million tokens per hour (rising to $1.00 on 1 January 2027).

The same introductory pricing applies to Gemini 3.6 Flash across Google AI Studio and the Gemini Enterprise Agent Platform. If you are running batch jobs or building cost models for production agents, the end-of-year cliff is worth building into your planning now rather than discovering it in January.

What this means for you if you build agents or write code with AI

The more meaningful story here is not a single headline capability but a collection of improvements that reduce friction across the full development loop: better first-pass code accuracy, stronger adherence to UI design specs, more reliable instruction following, and fewer retries in multi-step workflows. If you have been using 3.6 Flash and found yourself catching and correcting errors more than you would like, 3.7 Flash is worth testing against your actual workloads rather than relying on benchmarks alone.

For agentic pipelines, the jump from 49.0% to 65.3% on DeepSWE is the number to pay attention to. That benchmark reflects realistic software engineering tasks across multiple steps, which is much closer to what production agents actually do than a single code-completion score.

Migrating from an earlier model

If you are upgrading from Gemini 3.5 Flash, Gemini 3 Flash (Preview), or Gemini 3.1 Pro, there are two breaking changes to address. The sampling parameters temperature, top_p, and top_k are now deprecated and must be removed. The thinking_budget parameter is also gone; replace it with the thinking_level enum. Prefilled model turns are no longer supported either. The Gemini API changelog has the full migration detail.

One other note: gemini-robotics-er-1.6-preview is being shut down on 31 August 2026, so if anything in your stack references that model string, update it before then.

Gemini Omni 1.1 Flash and what it adds for video

Gemini Omni 1.1 Flash entered public preview on 27 August 2026 on the Gemini Enterprise Agent Platform, with availability extending to Google AI Studio, Flow, and the Gemini app. It is positioned as a multimodal model rather than a dedicated text-to-video tool, but the new video controls are the practical focus of this release.

The key additions:

Scene extension. The model can now analyse up to 10 seconds of prior video context when generating new content (earlier versions looked at only the final second). You can extend clips in 10-second increments up to a cumulative 40 seconds. Outside the API, scene extension is also available in the Gemini app for Google AI Plus, Pro, and Ultra subscribers.

First-and-last-frame interpolation. Specify two keyframes and the model generates continuous motion between them. This is designed for camera orbits, zoom transitions, and looping content where you need control over both the start and end state.

Video reference input. Provide up to three seconds of existing video as a reference and the model uses it to maintain consistency in characters and environments across subsequent generations.

360p draft mode. A new lower-resolution output option that renders up to 60% faster and at one-third the cost of standard 720p. Run your tests at 360p, then upscale the clips that work.

4K upscaling. Both 1080p and 4K outputs are upscaled from the 720p generation, not rendered natively. The default output remains 720p.

API pricing for video output sits at $17.50 per million tokens, which works out to roughly $0.10 per second of 720p footage. Audio is generated to match the visuals by default and can be directed via prompt, but the model does not accept audio files as input and ignores audio tracks in reference videos.

What this means for you if you create or build with video

Adobe has already integrated Omni Flash into Firefly. Figma’s creative director described the model as moving “beyond generating videos to truly directing them.” That framing is useful: the new controls are about repeatability and iteration, not just raw generation. First-and-last-frame control in particular addresses one of the persistent frustrations with generative video, which is the inability to reliably specify where a sequence ends up.

For developers building video features into products, the 360p draft mode is the practical starting point. Generate quickly and cheaply, identify the prompts and references that produce consistent results, then commit to full-resolution output. That workflow mirrors how professional iterative design already works, and the cost differential (one-third the price at draft resolution) makes it viable to explore properly before committing budget.

Both Gemini 3.7 Flash and Gemini Omni 1.1 Flash are available to explore in Google AI Studio now.