The May 2026 frontier closed-model field has split decisively on multimodal video. xAI's Grok 4.3 ships full native video input, exposed via API on April 30. Google's Gemini 4.0 has multimodal capability including video understanding plus the Gemini File Search system that handles image and text retrieval natively. Anthropic's Claude Sonnet 4.6 ships image-only — no native video, no video-frame ingestion through the model interface. Three labs, three positions, and the gap between them is now wide enough that buyer-side workload routing has to take it seriously rather than treat multimodal as a uniform feature.

This desk read the release docs, the API specs, and the production reports across the three labs. The question is narrow: where does native video input materially change a workflow, and where is frame-extraction-plus-vision-model still the right architecture? The answer is more bounded than the marketing implies.

What "Native Video Input" Actually Means

The phrase has been used loosely. Three architectures get bundled under "video input" and they have different operational properties.

Architecture 1: Frame-extraction plus vision model. The video is decomposed externally into still frames at some sampling rate; the frames are fed into a vision-capable model as a sequence of images. This is what most production systems shipped through 2024-2025. It works, but loses temporal continuity, audio, and motion semantics. Frame-extraction is what Claude Sonnet 4.6 users do today when they need to process video — externally, before the model sees anything.

Architecture 2: Native video tokenization with audio dropped. The video is processed by the model directly as video, with motion semantics preserved across frames, but the audio track is not understood. Some early multimodal models took this approach. It captures action and visual narrative but misses spoken dialogue and audio cues.

Architecture 3: Full native multimodal video. The video, including audio, is processed end-to-end by the model. Spoken dialogue is transcribed and understood; visual content is understood; the model can answer questions that require integrating both modalities ("what did the person on the left say while the chart was on screen?"). This is the operational ceiling, and it is where Grok 4.3 and Gemini 4.0 are converging in May 2026.

The marketing label "native video" is used inconsistently across labs. Buyers should establish which architecture is in play before relying on the feature for production workloads.

Grok 4.3 — Full Native Video, API Access April 30

Grok 4.3 Beta launched April 17 with native video input in the chat interface and gained API exposure on April 30. The xAI release positioning treats video as a first-class input modality alongside text and images. Operational details that matter for buyers: video duration limits per request, per-second pricing in the API tier, and supported codec/container formats are documented in the release notes and should be checked against the workload before integration.

Where Grok 4.3's video shows in concrete tasks: instructional video QA (asking the model to explain steps shown in a tutorial), surveillance review (querying recorded footage for specific events), product-demo analysis, and meeting summarization where the recording carries both visual context and dialogue. The Grok Computer integration on top of native video opens a class of workflows where the model both watches a screen-recording and acts on it through the desktop automation layer.

The gap to be honest about: production reliability data on video reasoning lags release dates by a typical 4-8 weeks for independent benchmarks. The April 30 API release is the integration starting point, not the validated production state.

Gemini 4.0 ships multimodal capability across its frontier tier and pairs it with the Gemini API File Search update — image and text file search that became generally available in May 2026 with multimodal extensions. The architecture is different from Grok's — Google leans on Gemini's existing video understanding capability plus a retrieval-aware multimodal pipeline that integrates with Drive, Workspace, and the broader Google data surface.

The differentiator for Gemini in the multimodal space is integration with the Google ecosystem. For Workspace-anchored teams, video assets that already live in Drive get analyzed without integration work. For non-Google ecosystems, the multimodal capability is comparable to Grok's at the model level but lacks the ecosystem leverage.

The video understanding has been part of Gemini's positioning longer than Grok's, and independent benchmarks at the multimodal layer have more publication history. For risk-averse buyers, the longer track record is itself a feature.

Claude Sonnet 4.6 — Image Only, No Native Video Yet

Claude Sonnet 4.6 ships image input, no native video. The architectural decision is consistent with Anthropic's release cadence pattern — features ship after they are reliable rather than at parity with the field. The absence of native video is not a hidden capability but an explicit current state.

For workloads where Claude is already the production model on text and code reasoning, the multimodal gap creates a routing problem rather than a switching problem. Most production agentic systems can route video-input tasks to a different model (Gemini, Grok 4.3) and keep Claude for the rest of the workflow. Multi-model orchestration is a more mature buyer pattern than single-model migration in May 2026.

The Claude track on multimodal video is unspoken in public communications. A reasonable expectation is video input arriving in a future Sonnet/Opus release, but there is no publicly committed timeline. Buyers planning around video input should not assume a near-term Claude release.

Production Use Cases Where Native Video Matters

The cases where full native video changes outcomes versus frame extraction are narrower than the marketing implies but real where they apply.

Case 1: Audio-visual integrated reasoning. Meeting recordings, instructional content, podcast video, screencasts where the question requires connecting what was said to what was shown. Frame extraction misses dialogue; transcription-only misses visual context. Only full native video integrates both.

Case 2: Temporal motion semantics. Sports analysis, surveillance event detection, action recognition in robotics or manufacturing footage. Frame extraction at low FPS loses motion; high-FPS frame extraction blows up token costs. Native video tokenization preserves motion at lower per-second cost.

Case 3: Long-form continuous content. Hour-plus video analysis where frame extraction would generate thousands of images and exceed reasonable context budgets. Native video processing handles longer continuous content within practical context windows.

Production Use Cases Where Frame-Extraction Is Enough

The cases where frame extraction stays the right architecture are larger than the lab marketing suggests.

Case 1: Static-scene analysis. Product photography QA, document video where each frame is essentially a slide, security camera stills with motion-triggered sampling. Frame extraction at low rates captures everything that matters.

Case 2: Existing pipeline lock-in. Production systems already running frame-extraction with vision models, where the migration cost to native video exceeds the marginal accuracy gain. Most production multimodal systems shipped before May 2026 fall here.

Case 3: Cost-sensitive workflows. Native video pricing per second is meaningfully higher than image pricing per frame at the equivalent sampling rate. For high-volume workflows where accuracy is already adequate at frame extraction, the cost differential rules out native video at the margin.

Routing Architecture Implications for Multi-Model Production Systems

The May 2026 multimodal split changes how production systems should be architected, not just which single model gets selected.

Element 1: Multimodal capability is now a routing axis. Production systems running on a single frontier model are leaving capability on the table. A workflow that needs Claude's code reasoning and Grok's video understanding cannot be served by either alone. The pragmatic architecture is task-typed routing where the input modality and reasoning class together determine the model. This is more orchestration overhead than a single-model deployment but the capability gap is now wide enough that the overhead earns its place.

Element 2: Frame extraction is not deprecated. Native video is the future for the workloads that need it; frame extraction is the present for the workloads that do not. Production systems should not refactor working frame-extraction pipelines toward native video without a workload-specific gain calculation. Replacing a working architecture for marketing parity is a common failure mode in multimodal adoption.

Element 3: Vendor lock-in on multimodal is heavier than on text. Switching a text-only workflow between frontier labs is increasingly straightforward — APIs converge, prompt structures translate. Switching a multimodal workflow requires re-handling video duration limits, codec compatibility, per-second pricing math, and modality-specific prompting patterns that vary by lab. Buyers should plan for multimodal vendor stickiness as a structural reality rather than a temporary integration cost.

What This Desk Tracks Through Q2-Q3 2026

Three datapoints anchor ongoing tracking. First, independent benchmark publication on Grok 4.3 and Gemini 4.0 native video — the 4-8 week post-release lag should resolve into measurable comparative reliability data by late Q2. Second, Claude's multimodal video positioning. Whether Anthropic ships native video in a Sonnet 4.7 or Opus 5.0 cycle, and on what timeline, will reshape the routing calculus. Third, native video pricing trajectory. The April 30 Grok API cut suggests video pricing pressure; whether Gemini and a future Claude release land at comparable per-second pricing or stay at premium ceilings will determine whether native video becomes a default versus a specialist routing.

Honest Limits

This piece is a comparative read on publicly disclosed multimodal capabilities and pricing as of early May 2026. Direct production benchmarking of the three native-video paths under matched workloads has not been performed by this desk; the analysis relies on vendor documentation, third-party reviews, and the limited independent benchmark coverage available within the release window. Specific pricing figures should be verified against each vendor's primary documentation before procurement decisions. The Grok 4.3 API was released April 30 and the production-reliability profile at scale is not yet established. Gemini 4.0 multimodal video has more track record but the May 2026 File Search update is recent. Claude Sonnet 4.6 multimodal positioning may shift on a future release; this analysis reflects the current state. Routing recommendations are framework-level reads, not workload-specific recommendations — every multimodal production decision benefits from a small-volume pilot against the actual data before committing.

Sources