Veo 3.1 Historical API Capabilities: First/Last Frames, Three References and Native Audio

Why an obsolete model still deserves a precise reference page

Google announced Veo 3.1 and Veo 3.1 Fast on October 15, 2025. Its technical reference documented short video generation with native audio, portrait and landscape options, extension, and both initial/last frames. Those historically important details explain what a stored pipeline attempted. They do not justify presenting now-closed Gemini API Preview identifiers as usable on October 25, 2026.

A reference image is not the same as the first frame

The starting image is intended to animate from its contents, whereas an asset reference tells the model what a person, object or style should look like. For Veo 3.1 the historical Gemini API documentation describes up to three reference images, distinct from first/last-frame interpolation. An editor expecting a literal first frame should not replace it with an ingredient reference and assume equal composition. Specify which source controls exact geometry and which merely guides identity, then check whether each option exists in the destination platform.

Duration and resolution cannot be separated

The documentation identifies 24 frames per second. For 3.1 main/Fast, durations of four, six and eight seconds were available with conditions; reference-image requests and 1080p or 4K selected an eight-second output. Extension was restricted to 720p in the documented Preview profile. A matrix cell saying merely “supports 4K” conceals these dependencies. Store an operation tuple of model ID, mode, duration, resolution, ratio and audio instead of treating each capability as globally available.

Build a defensible archive schema

Each stored request should include the exact deployed model identifier, request date, operation type, reference assets and their hashes, the generator’s response ID and any editing that followed. Keep an immutable copy of the downloaded result. That preserves auditability after an API retires and makes later migration decisions less speculative. A versioned matrix can accurately say a capability existed historically without promising that an enterprise product or aggregator supports it today.

Storyboard example

A designer prepares a six-shot product narrative with consistent clothing. For each shot, identify whether the requirement is an exact opening camera angle, a matching ending pose, or merely the same shirt. These correspond to different controls in the historic Veo reference. Ask for a brief confirmation clip before paying for a larger production run, and have an editor check fingers, labels and background shifts. If migrating, replace each mode only after verifying equivalent parameters on the new service.

License and access caveats

The 2025 release note and historical technical guide are Google-first-party evidence, not independent video evaluations. Public Preview no longer functions as of Google’s October 22, 2026 notice. Whether Vertex AI or an enterprise platform offers a GA model with similar branding must be checked separately, with dated project entitlement and region. Preview usage charges cannot establish today’s price in any other account.

Alternatives

An editor can enforce initial/ending frames through deterministic traditional animation, or use an API that officially documents frame interpolation. For continuity-sensitive shots, composition and retiming may be more reliable than a long generative take. This is a historical capability explainer, not a ranking or a recommendation to recreate discontinued calls.

Sources & Further Reading

• Google — Veo 3.1 API model parameters and restrictions (first-party source; checked 2026-10-25)
• Google Developers Blog — Veo 3.1 release announcement, 2025-10-15 (first-party source; checked 2026-10-25)
• Google — Video generation in Gemini API (first-party source; checked 2026-10-25)