Final image
Nano Banana Pro when typography, fine detail and the final render matter. I wouldn't use it to crank out 40 rough drafts.
labs.google →Image · video · AI creation
Nano Banana, Sora, Gemini Omni: what should you pick?
You don't pick creative models the way you pick a chatbot. Here I split things by job: quick draft, final image, text inside the image, short video, synced audio, editing and social formats.
Quick decision
What would I pick today?
Quick variants
Nano Banana 2 to test compositions, styles and ideas fast before you settle on the best direction.
labs.google →Text inside the image
Ideogram is still a very handy option when the visual has to carry a slogan, a poster or a layout.
ideogram.ai →Creative style
Midjourney keeps a real edge on art direction, especially if the API isn't your main criterion.
midjourney.com →Control / open
FLUX and Stable Diffusion are the way to go if you want to customize, self-host or experiment.
blackforestlabs.ai →AI video
Gemini Omni for multimodal editing, Sora for narration, Veo for a short scene with audio.
See the video section ↓Top quality
Nano Banana Pro
- Best pick for a final image that matters.
- Better suited to complex scenes and to typography.
- Less natural for churning out lots of cheap variants.
Fast iteration
Nano Banana 2
- Best pick to explore several directions fast.
- More sensible for drafts, tests, styles and batches of images.
- Swap in Pro once the image has to be publishable as is.
Image vocabulary
The 4 words that change your pick.
You don't need to be a designer to get it. But these 4 terms show up in every doc and shape what you can actually do.
| Word | In plain English | What it changes for you |
|---|---|---|
| Prompt-to-image | You describe it in text, the AI builds an image from scratch. | The classic mode. Every image AI can do this. |
| Image-to-image | You give it a base image, the AI transforms it based on your prompt. | Essential for retouching, restyling or iterating on an existing visual. |
| Inpainting | You select a specific area and ask the AI to regenerate it. | Really useful for erasing an object, swapping a face, fixing a detail. |
| Reference / ControlNet | You give it a reference image for the style, the pose or the composition. | Key for consistency: the same character across several visuals. |
Image catalog
The image models worth knowing.
AI video · in real life
AI video, without the hype.
It's not an LLM with a camera.
AI video answers to different criteria than image or text: maximum length, synced audio or not, character consistency across shots, image-to-video to start from a photo, price per second.
In 2026, no single model does it all. You often pick based on the deliverable: an 8-second teaser for Instagram, a 30-second narrative scene, or a video reworked from existing images.
8–30stypical length of an AI video today
10×pricier than image, per second of render
Oftenaudio still has to be added in post
Rarereal long-form character consistency
In your situation...
Which video model for what?
Short narration
Sora is built to produce a sequence with a clear narrative intent — handy for a teaser, a spot or a 10-to-20-second piece of storytelling.
openai.com/sora →Synced audio
Veo 3.1 outputs a short video with a coherent soundtrack. The most natural trade-off when you want to skip adding sound afterward.
labs.google/fx (Flow) →Multimodal editing
Gemini Omni takes text, image, audio or video as input and edits your video in a conversation. The most interesting bet on the workflow side.
gemini.google.com →Creative production
Runway Gen-4 acts more like a studio than a chatbot. A better fit when you want a production tool with a timeline and shot-by-shot control.
runwayml.com →Quick social clip
Kling 2.5 Turbo is still a serious contender for short clips. Keep it on hand, but check your sources before any brand-sensitive use.
klingai.com →Testing a social format
Grok Imagine is for testing a format for X super fast or a throwaway teaser. Riskier ground: avoid it for a brand that has to protect its image.
grok.com →Premium narration
Sora
- Designed for sequences with story intent.
- Image-to-video and audio generation built in.
- Go-to when narrative quality matters more than speed.
- Stays inside the OpenAI ecosystem: handy if you're already there.
Native synced audio
Veo 3.1
- Outputs coherent video + audio on the first pass.
- Short but clean, ideal for a scene with dialogue or ambient sound.
- Available through Flow (Google) and the API.
- My pick when I want to skip a separate audio tool.
Comparison table
AI video: who does what?
| Model | Typical length | Native audio | Image-to-video | Access | Use case |
|---|---|---|---|---|---|
| Sora ↗ | 10–20 s | Yes (recent) | Yes | OpenAI chat · API | Short narration, teaser, storytelling |
| Veo 3.1 ↗ | 8–10 s | Yes — native | Yes | API · Flow | Scene with dialogue or ambient sound |
| Gemini Omni ↗ | Short | Possible (multimodal) | Yes | Gemini · Flow · API | Editing video in a conversation |
| Runway Gen-4 ↗ | 10 s + | No (post-production) | Yes | Web app · API | Studio production, creative timeline |
| Kling 2.5 Turbo ↗ | Short | Limited | Yes | Web app · API | Quick clip, social format |
| Grok Imagine ↗ | Very short | No | Limited | X · Grok app | Testing a social format, throwaway |
Audio · the third creative floor
Audio, the link everyone forgets.
A silent AI video tells no story. A voice agent with no voice of its own sounds like a robot. These 4 tools cover the real audio needs: clean narration, real-time voice agents, accurate transcription, open-weight audio.
Clean narration
Eleven v3 is still the reference when you want a voice that doesn't sound synthetic: video voiceover, podcast, audiobook. It's what I use for AI Wars.
elevenlabs.io →Real-time voice agent
Gemini Native Audio is built for agents that speak and listen — not just read out a script. Worth a look if you're building an interactive voice assistant.
ai.google.dev/live →Transcription
Whisper is still handy for turning audio into text. An open-weight model from OpenAI, it can run locally — ideal for volume with no per-minute cost.
github.com/openai/whisper →Open-weight audio
Voxtral from Mistral does transcription and audio understanding, open-weight (Apache 2.0). The right pick if you want multilingual without shipping data to a vendor.
mistral.ai/voxtral →In detail
What each model actually does.
Eleven v3
ElevenLabs · text-to-speech · proprietary
The most natural consumer voice model for long-form narration. Believable French voices, prosody control (pauses, intonation, emphasis) via tags, voice cloning from a 1-minute sample.
- Modalities
- Text-to-speech, voice cloning, dubbing
- Languages
- 32+ languages, including very natural French
- Price
- Subscription: Starter $5/mo · Creator $22/mo · Pro $99/mo (by characters/minutes)
- Access
- Web app, API, integrations (CapCut, Descript, Make...)
- Who it's for
- Podcast, video, audiobook and e-learning creators
- Limits
- No open-weight, platform lock-in, quality varies by voice
Gemini Native Audio
Google · real-time bidirectional audio · proprietary
Able to listen AND reply out loud, in real time, with natural interruptions. The building block for an agent that actually converses — not just a TTS reading a script.
- Modalities
- Audio in, audio out, real time, interruption
- Languages
- Multilingual (French included)
- Price
- API: billed per audio token (≈ $25/M audio input tokens)
- Access
- Gemini Live API, Vertex AI, Google AI Studio
- Who it's for
- Devs building a voice assistant, AI customer support
- Limits
- More dev than creative, latency to measure on your own use
Whisper
OpenAI · speech-to-text · MIT (open-weight)
The de facto standard for transcription. 99 languages detected automatically, robust on noisy audio, can run locally on a Mac. large-v3 is still the reference in 2026.
- Modalities
- Speech-to-text, translation (into English)
- Languages
- 99 languages auto-detected
- Price
- Free locally · OpenAI API: $0.006/minute
- Access
- GitHub (pip install), Whisper.cpp, OpenAI API, Hugging Face
- Who it's for
- Podcast transcription, meeting notes, subtitling
- Limits
- Doesn't generate voice, local latency depends on your machine
Voxtral
Mistral AI · audio understanding · Apache 2.0 (open-weight)
Mistral's open-weight audio model (2024-2025). Two sizes: Mini 3B (fast, edge) and Small 24B (production quality). Broader than Whisper: it transcribes AND understands (Q&A on the audio, summary, analysis).
- Modalities
- Speech-to-text + audio understanding (Q&A, summary)
- Languages
- Solid multilingual (FR, EN, ES, DE, IT, PT, NL, HI)
- Price
- Free locally · Mistral API: $0.001/min (Small) · $0.0004/min (Mini)
- Access
- Mistral API, Hugging Face, Le Chat, local deployment (vLLM)
- Who it's for
- Privacy, volume, multilingual, long context (40 min)
- Limits
- Doesn't generate voice, ecosystem less mature than OpenAI's
My stack — Jerwis Productions
How I produce the AI Wars podcast
No buzz, just the real thing. Here's the exact tool chain I use to produce a 15-20 minute episode, with synthetic voices, studio mastering, and zero humans in front of a mic. Copy it or use it as inspiration.
Narrator · Paul K
"Deep French Narrator" voice on Eleven v3. The calm, low tone that carries a 17-minute narration without wearing you out.
Sam Altman · Simon
French male Eleven voice, pitched a bit higher to match Altman's fast timbre.
Dario Amodei · Mathieu
French male Eleven voice, more measured. The contrast with Simon keeps the dialogue easy to follow.
Daniela Amodei · Camille Martin
French female Eleven voice, direct and professional register.
01 · Script
Written with Claude Sonnet from public sources (interviews, articles, books). Human-reviewed to avoid hallucinations.
02 · Voice synthesis
Each line assembled in Python via the ElevenLabs API. Prosody tags for pauses, emphasis, transitions.
03 · Mix & mastering
Imported into REAPER. Ambient music, light sound design, studio mastering at -16 LUFS for podcast standards.
Back to the full comparison
Image, video and audio are only part of the picture. The hub also lists LLMs, local models, embeddings and rerankers.