SOTA Models

The best-overall picks in this collection are — Image: FLUX.2 [pro]; Video: Veo 3.1; Audio: Fish Audio S2.1 Pro; Text · LLM: GPT-5.4 Pro. Each group also lists runner picks for workloads where the leader isn't the right fit.

Frontier picks cost several multiples per call of their fast-tier counterparts across every modality: FLUX.2 [pro] vs Nano Banana Pro on image, Veo 3.1 vs Seedance Fast on video, GPT-5.4 Pro vs Gemini 3.1 Pro on reasoning. If you are shipping a consumer app at scale, a two-tier strategy (fast model in the hot path, frontier on demand) almost always wins.

Sub-second latency comes at a fidelity cost in every modality. Nano Banana Pro vs FLUX.2 [pro] on image quality, Seedance Fast vs Veo 3.1 on motion coherence, streaming TTS chunks vs full synthesis on voice, and faster reasoning models vs GPT-5.4 Pro on multi-step problems. The fast tier is the right call for iteration, preview, and realtime UX. Do not ship it as final assets in production creative.

Most top picks here are closed-weight. If portability, on-prem deployment, or audit requirements matter to you, the Best Open Models collection curates the strongest open-weight options across image, video, audio, and text — FLUX.2 [klein], Z-Image, LTX, ACE-Step, Qwen, DeepSeek, and GLM families. They typically lag the frontier on the hardest prompts but are close enough for many production workloads.

All picks here are cleared for commercial use through Runware, but underlying terms vary by modality. Recraft and Veo are the safest defaults for brand creative on image and video. Voice-cloning workflows carry consent obligations that you must surface to your users. LLM output terms vary across providers on fine-tuning, distillation, and reuse as training data; check the model page before any pipeline that re-uses generations downstream.

Wire the fast pick into preview and iteration paths: Nano Banana Pro for image, Seedance Fast for video, a Flash-tier model for early chat responses, streaming TTS chunks for first-token voice playback. Reserve frontier picks for the export, final-render, and the hardest reasoning requests. Median spend drops 5–8× without hurting perceived quality.

Every model in this collection sits behind a Runware-managed endpoint with SLA, but no single upstream is bullet-proof. Define a secondary pick per modality and route to it on 5xx or queue-depth thresholds. The runners-up here are deliberately picked so this is one-line config.

License terms shift between model versions and they differ across modalities. Re-check the model page before any external launch. Recraft and Veo are the safest defaults for brand-led visuals; voice-cloning workflows carry consent obligations; LLM outputs may carry training-data and distillation restrictions that affect downstream pipelines. Anything destined for paid media or external distribution deserves a dedicated license review.

Pin model versions explicitly in your API calls. Do not rely on default routing. When a model is replaced in this collection, you want to opt in to the upgrade, not have it happen to you mid-campaign.

The model landscape moves fast. Every week a new image, video, audio, or language model claims the top spot on some benchmark, and most teams do not have the time to test them all against their own workload. This collection is the answer to a single question we hear constantly from developers and product teams on Runware: if you were starting a production AI feature today, which models would you ship with? It is intentionally small. It spans every major modality. It is biased toward models that are production-ready right now, with strong prompt adherence, predictable latency, sane pricing, and a license you can actually ship behind. It is not a leaderboard.

Each modality has a small fixed eval set that our team built and extends slowly, with prompts that span photoreal, illustration, motion coherence, voice naturalness, coding, tool use, and long-context reasoning. Every new model that launches on Runware runs that set, with results logged against the previous champion. Picks are not pure benchmark winners. A model only ships into this collection if it clears three additional bars: predictable latency under production load, transparent pricing, and a license that lets you actually use the output. Where the leader is too expensive or too slow for a given job, we add a runner-up labelled fastest or best price so you can pick the right one for your workload.

Every pick on this page links to its model page, where you will find the full API reference, parameter ranges, and pricing examples. Many models also ship with their own model-specific guides covering common workflows. For broader topics that span models (inpainting and character consistency on the image side, image-to-video and motion control on video, voice cloning and dubbing on audio, RAG and tool-calling patterns on text), the Learning Center is the authoritative reference.

New models run our eval set the week they land on Runware. When a new entrant clears the bar against the current champion across quality, latency, pricing, and licensing, it replaces the previous pick and we publish a short note explaining why. Models that fall out of the picks remain in the broader catalog. There is no fixed cadence; the collection moves when the field moves.

Every pick on this page runs on Runware's inference layer: a unified API with predictable latency, transparent per-call pricing, and a single billing surface across image, video, audio, and text. You can hit each model directly through its original provider if you prefer, but routing through Runware means consistent error handling, SLA-backed uptime, version pinning, and the ability to switch between picks without rewriting your integration.