AI Provider Models

July 26, 2026 · View on GitHub

Single source of truth for AI model information used across VibeFrame.

Agent LLM Providers (7)

Used for natural language processing in Agent mode (vibe command).

Provider	Model	API Model ID	Env Key	CLI Option
OpenAI	GPT-5-mini	`gpt-5-mini`	`OPENAI_API_KEY`	`-p openai`
Claude	Sonnet 4.6	`claude-sonnet-4-6`	`ANTHROPIC_API_KEY`	`-p claude`
Gemini	2.5 Flash	`gemini-2.5-flash`	`GOOGLE_API_KEY`	`-p gemini`
xAI	Grok 4.1 Fast	`grok-4-1-fast-reasoning`	`XAI_API_KEY`	`-p xai`
OpenRouter	Auto (300+ models)	`openrouter/auto`	`OPENROUTER_API_KEY`	`-p openrouter`
Evolink	GPT-5.2 (via unified API)	`gpt-5.2`	`EVOLINK_API_KEY`	`-p evolink`
Ollama	Local models	user-configured	-	`-p ollama`

OpenAI model options:

gpt-5-mini is the default — 10× cheaper input / 5× cheaper output than GPT-4o, with better performance. Ideal for agentic loops (20+ calls per task). You can override per-session:

Model ID	Variant	Notes
`gpt-5-mini`	GPT-5 Mini	Default. Best cost-performance ratio. $0.25/M input, $2/M output
`gpt-5.4`	GPT-5.4	Frontier model, 1M context. $2.50/M input, $20/M output

gpt-5.4-pro is not available via Chat Completions (Responses API only) — not usable for Agent mode.

To use GPT-5.4 in agent mode: vibe agent -p openai --model gpt-5.4

Claude model options:

claude-sonnet-4-6 is the default — best balance of capability and cost for agentic loops. You can override per-session:

Model ID	Variant	Notes
`claude-sonnet-4-6`	Sonnet 4.6	Default. Best cost-performance for agent loops. $3/M input, $15/M output
`claude-opus-4-7`	Opus 4.7	Highest capability, step-change in agentic coding, 1M context. $5/M input, $25/M output
`claude-haiku-4-5-20251001`	Haiku 4.5	Fastest, lowest cost. $1/M input, $5/M output
`claude-opus-4-6`	Opus 4.6 (legacy)	Previous Opus tier — same price as 4.7. Still supported.

To use Opus in agent mode: vibe agent -p claude --model claude-opus-4-7

xAI model options:

grok-4-1-fast-reasoning is the default — optimized for tool calling with 2M context window, 15× cheaper input / 30× cheaper output than Grok 4. You can override per-session:

Model ID	Variant	Notes
`grok-4-1-fast-reasoning`	Grok 4.1 Fast (reasoning)	Default. Agent-optimized, 2M context. $0.20/M input, $0.50/M output
`grok-4`	Grok 4 (flagship)	Highest capability, 256K context. $3/M input, $15/M output
`grok-4-1-fast-non-reasoning`	Grok 4.1 Fast (non-reasoning)	Faster responses, no chain-of-thought. $0.20/M input, $0.50/M output

To use Grok 4 in agent mode: vibe agent -p xai --model grok-4

OpenRouter model options:

openrouter/auto is the default — automatically routes to the best available model. You can specify any model available on OpenRouter:

Model ID	Provider	Notes
`openrouter/auto`	Auto	Default. Automatically selects best model
`anthropic/claude-sonnet-4-6`	Anthropic	Claude Sonnet via OpenRouter
`openai/gpt-5-mini`	OpenAI	GPT-5 Mini via OpenRouter
`google/gemini-2.5-flash`	Google	Gemini Flash via OpenRouter
`meta-llama/llama-4-scout`	Meta	Llama 4 Scout (open-weight)
`deepseek/deepseek-r1`	DeepSeek	DeepSeek R1 reasoning model

To use a specific model: vibe agent -p openrouter --model anthropic/claude-sonnet-4-6

See openrouter.ai/models for the full list of 300+ available models.

Evolink model options:

gpt-5.2 is the default — reliable, fast, and cost-effective for agentic loops. You can specify any model available on Evolink:

Model ID	Provider	Notes
`gpt-5.2`	OpenAI	Default. GPT-5.2 via Evolink
`evolink-auto`	Auto	Auto-routing (may have spotty availability)
`claude`	Anthropic	Claude via Evolink
`gemini-2.5-pro`	Google	Gemini 2.5 Pro via Evolink
`deepseek-chat`	DeepSeek	DeepSeek V3 via Evolink
`doubao-seed-2.0-pro`	ByteDance	Doubao Seed 2.0 Pro via Evolink

To use a specific model: vibe agent -p evolink --model gpt-5.2

See docs.evolink.ai for the full model catalog.

Why Gemini 2.5 Flash for agent mode, not Gemini 3.5 Flash?

Agent mode runs an agentic loop — the LLM is called repeatedly (potentially dozens of times per task) to reason and call tools. For this use case:

Speed matters: Flash responds ~3–5× faster than Pro, keeping the interactive session snappy
Tool calling stability: gemini-2.5-flash has well-tested, stable function calling support; newer Gemini 3.x models may have stricter rate limits or different behavior in multi-turn tool calling loops
Cost: Flash is significantly cheaper per token — important when a single agent task may trigger 20+ LLM calls
Agent loop risk: newer frontier models can be better for one-shot reasoning, but agent mode values predictable repeated tool calls

For one-shot Gemini calls, VibeFrame now maps flash / latest to gemini-3.5-flash. That applies to media analysis, render review, storyboard generation, Gemini-backed silence-cut analysis, and Gemini composition. Agent mode keeps gemini-2.5-flash as its explicit default.

gemini-3.1-pro-preview remains available for motion graphics code generation (vibe generate motion -m gemini-3.1-pro) where its creative reasoning matters for a single generation call.

Gemini model options:

gemini-2.5-flash is the default — fastest and most cost-effective for agentic loops. You can override per-session:

Model ID	Variant	Notes
`gemini-2.5-flash`	2.5 Flash	Default. Fast, stable tool calling. Free tier available
`gemini-3.5-flash`	3.5 Flash	One-shot `flash` / `latest` default for analysis, review, storyboard, and Gemini composition
`gemini-3-flash-preview`	3 Flash Preview	Preview model available via `flash-3` or explicit model ID
`gemini-2.5-pro`	2.5 Pro	Higher reasoning capability, slower. $1.25/M input, $10/M output
`gemini-3.1-pro-preview`	3.1 Pro (preview)	Latest, preview — may have unstable tool calling

To opt into Gemini 3.5 Flash in agent mode: vibe agent -p gemini --model gemini-3.5-flash

Motion Graphics LLM (vibe generate motion)

Used for Remotion component code generation (vibe generate motion).

Alias	Model	Provider	Env Key	CLI Option	Notes
`sonnet`	`claude-sonnet-4-6`	Claude	`ANTHROPIC_API_KEY`	`-m sonnet`	Default
`opus`	`claude-opus-4-7`	Claude	`ANTHROPIC_API_KEY`	`-m opus`	Best quality (step-change agentic coding)
`opus-4-6`	`claude-opus-4-6`	Claude	`ANTHROPIC_API_KEY`	`-m opus-4-6`	Previous Opus tier (legacy)
`gemini`	`gemini-3.5-flash`	Gemini	`GOOGLE_API_KEY`	`-m gemini`	Gemini 3.5 Flash one-shot default
`gemini-2.5-pro`	`gemini-2.5-pro`	Gemini	`GOOGLE_API_KEY`	`-m gemini-2.5-pro`	Previous Gemini motion alias target
`gemini-3.1-pro`	`gemini-3.1-pro-preview`	Gemini	`GOOGLE_API_KEY`	`-m gemini-3.1-pro`	Gemini 3.1 Pro

Text-to-Image (3 providers, 7 models)

Provider	Model	Env Key	CLI Option	Notes
OpenAI	`gpt-image-2`	`OPENAI_API_KEY`	`-p openai`	Flagship OpenAI image model
OpenAI	`gpt-image-1.5`	`OPENAI_API_KEY`	`-p openai -m 1.5`	Previous default, still strong on editing (#1 editing leaderboard). Quality tiers: low ($0.009), medium ($0.035), high ($0.133)
Gemini	`gemini-2.5-flash-image`	`GOOGLE_API_KEY`	`-p gemini`	Nano Banana Flash - GA, fast. Auto-selected when only `GOOGLE_API_KEY` is set
Gemini	`gemini-3.1-flash-image-preview`	`GOOGLE_API_KEY`	`-p gemini -m 3.1-flash`	Nano Banana 2 - Image Search grounding, 512px
Gemini	`gemini-3-pro-image-preview`	`GOOGLE_API_KEY`	`-p gemini -m pro`	Nano Banana Pro - higher quality, up to 4K
xAI Grok	`grok-imagine-image`	`XAI_API_KEY`	`-p grok`	$0.02/image, standard quality
xAI Grok	`grok-imagine-image-pro`	`XAI_API_KEY`	`-p grok -m pro`	$0.07/image, higher quality

Image Aspect Ratios (Gemini)

All Gemini image models support 14 aspect ratios: 1:1, 1:4, 1:8, 2:3, 3:2, 3:4, 4:1, 4:3, 4:5, 5:4, 8:1, 9:16, 16:9, 21:9

Image Aspect Ratios (Grok)

Grok Imagine supports 14 aspect ratios: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 2:1, 1:2, 19.5:9, 9:19.5, 20:9, 9:20, auto

Image Editing (3 providers)

Provider	Model	Max Input Images	CLI Option	Features
Gemini	Flash	3	`-p gemini` (default)	Fast editing, 1K output
Gemini	3.1 Flash	3	`-p gemini -m 3.1-flash`	Image Search grounding, 512px-1K output
Gemini	Pro	14	`-p gemini -m pro`	Multi-image composition, up to 4K output
OpenAI	`gpt-image-1.5`	16	`-p openai`	Instruction-based editing, multipart upload
OpenAI	`gpt-image-2`	16	`-p openai -m 2`	Flagship editor, 2026-04-21 GA
xAI Grok	`grok-imagine-image`	1	`-p grok`	Single image editing, $0.02/edit

Text-to-Video (5 stable + 1 experimental)

Models marked Audio: Yes generate synchronized sound (dialogue, SFX, ambient). Silent models need separate vibe generate speech / vibe generate sound-effect.

Provider	Model	Duration	Audio	Env Key	CLI Option	Notes
Seedance via fal.ai	`seedance-2.0` (ByteDance)	4-15 sec	Yes	`FAL_API_KEY`	`-p seedance`	ByteDance Seedance through fal.ai
Seedance via fal.ai	`seedance-2.0-fast`	4-15 sec	Yes	`FAL_API_KEY`	`-p seedance --seedance-model fast`	Lower-latency / lower-cost variant of Seedance 2.0
xAI Grok	`grok-imagine-video`	1-15 sec	Yes	`XAI_API_KEY`	`-p grok`	Fallback default when `FAL_API_KEY` is unset. Best lip-sync/native audio. $0.07/s (720p)
Kling	`kling-v2-5-turbo`	5-10 sec	No	`KLING_API_KEY`	`-p kling`	Fast (~36s generation)
Kling	`kling-v2-6`	5-10 sec	No	`KLING_API_KEY`	`-p kling -m v2.6`	High quality
Kling	`kling-v3`	5-10 sec	No	`KLING_API_KEY`	`-p kling -m v3`	Higher quality, multi-shot, lip-sync
Kling	`kling-v3-omni`	3-15 sec	Yes	`KLING_API_KEY`	`-p kling -m v3-omni`	Native audio (multilingual), character consistency
Veo	`veo-3.1-fast-generate-preview`	4-8 sec	Yes	`GOOGLE_API_KEY`	`-p veo`	Native audio, fast
Veo	`veo-3.1-generate-preview`	4-8 sec	Yes	`GOOGLE_API_KEY`	`-p veo --veo-model 3.1`	Native audio, higher quality
Veo	`veo-3.0-generate-preview`	5-8 sec	Yes	`GOOGLE_API_KEY`	`-p veo --veo-model 3.0`	Native audio
Runway	`gen4.5`	2-10 sec	No	`RUNWAY_API_SECRET`	`-p runway`	Flagship, text+image-to-video (12 credits/sec)
Runway	`gen4_turbo`	5-10 sec	No	`RUNWAY_API_SECRET`	`-p runway --runway-model gen4_turbo`	Legacy, image-to-video only
Gemini Omni ⚠️	`gemini-omni-flash-preview`	preview	Yes	`GOOGLE_API_KEY`	`-p omni`	Experimental, opt-in only. New `/v1beta/interactions` endpoint. Never auto-selected. See below.

-p fal is a deprecated v0.x alias for -p seedance and will be removed at the 1.0 cut. Use -p seedance in new scripts.

⚠️ Seedance rejects image-to-video inputs showing a recognizable face. This is ByteDance's platform-wide likeness policy, not a fal.ai quirk: the API returns a deterministic HTTP 422 ("The images or videos provided may contain likenesses of real people") for keyframes where a person's face is clearly visible — including AI-generated photoreal faces. Hands-only or back-of-head shots pass. Retrying does not help. vibe build falls back to Runway automatically for such beats when RUNWAY_API_SECRET is configured; you can also pin a beat with a provider: runway cue, or run vibe generate video "<motion>" -p runway -i keyframe.png directly. Observed 2026-07-26 on seedance-2.0.

⚠️ Gemini Omni is experimental. It is not wired into default provider resolution — you must pass -p omni explicitly. The preview interactions schema may change; treat it as unstable.

Veo Advanced Options

Option	Values	Description
`--negative-prompt`	text	What to avoid in the generated video
`--resolution`	720p, 1080p, 4k	Video resolution
`--last-frame`	image path	Frame interpolation (first→last frame)
`--ref-images`	image paths (max 3)	Character consistency (Veo 3.1 only)
`--person`	allow_all, allow_adult	Person generation setting
`veo-extend`	operation-name	Extend a previously generated Veo video

Image-to-Video

All text-to-video providers also support image-to-video. Key differences per provider:

Provider	Model	I2V Support	Image Input	Notes
Seedance via fal.ai	`seedance-2.0`	Yes	URL only	Auto-uploads via ImgBB (`IMGBB_API_KEY`) for local image paths.
xAI Grok	`grok-imagine-video`	Yes	URL or data URI	Same pricing as T2V
Kling	all v2.5+ models	Yes	URL only	Auto-uploads via ImgBB (`IMGBB_API_KEY`)
Veo	all models	Yes	base64 (first frame)	Supports `--last-frame` for frame interpolation
Runway	`gen4.5`	Yes	URL or data URI	Text+image-to-video
Runway	`gen4_turbo`	I2V only	URL or data URI	Cannot do text-only generation
Gemini Omni ⚠️	`gemini-omni-flash-preview`	Yes (experimental)	base64 (first frame)	Sends `video_config.task: image_to_video`; opt-in `-p omni` only

Gemini Omni (experimental)

Google Gemini Omni (gemini-omni-flash-preview) is a preview video generation/editing model on the Generative Language API. VibeFrame ships an opt-in, experimental client for it — vibe generate video -p omni. It is not part of default provider resolution (Seedance / Veo / Kling / Runway / Grok remain the stable path), and the preview request/response schema may change.

Reference: https://ai.google.dev/gemini-api/docs/omni

Property	Value
Model ID	`gemini-omni-flash-preview`
Endpoint	`POST /v1beta/interactions` (new stateful endpoint — not Veo's `:predictLongRunning`)
Auth	`GOOGLE_API_KEY` (same key as Gemini / Veo — no new credential)
Tasks	`text_to_video`, `image_to_video`, `reference_to_video`, `edit` via `generation_config.video_config.task`
Aspect ratios	`16:9`, `9:16`
Watermark	SynthID

Preview limits (per Google's docs, subject to change):

Audio-reference upload is not supported.
Video references are limited to ≤3 seconds.
Editing uploaded video is restricted in the EEA, Switzerland, and the UK.
English-tested only; no system-instruction or temperature controls.
Large (>4MB) videos may be returned via an async delivery: uri reference.

VibeFrame's Omni client (packages/ai-providers/src/gemini/gemini-omni.ts) parses the interactions response defensively for the video URL, since the preview schema is not yet stable.

Audio (5)

Provider	Capability	Env Key	Notes
ElevenLabs	TTS, SFX, Music, Voice Clone	`ELEVENLABS_API_KEY`	Music: 3s-10min, model music_v1. TTS: eleven_v3
OpenAI TTS	TTS	`OPENAI_API_KEY`	gpt-4o-mini-tts (~$0.015/min of audio); voices incl. marin, alloy, nova
Kokoro	TTS (local, free)	—	Kokoro-82M (Apache 2.0); ~90MB model on first use; bundled in the Desktop extension
Whisper	Transcription	`OPENAI_API_KEY`	OpenAI API
Replicate	Music generation	`REPLICATE_API_TOKEN`	MusicGen, max 30s

Quick Reference

Environment Variables

# LLM Providers
export OPENAI_API_KEY="sk-..."        # GPT, Whisper, GPT Image 1.5
export ANTHROPIC_API_KEY="sk-ant-..." # Claude
export GOOGLE_API_KEY="AIza..."       # Gemini (image, Veo video)
export XAI_API_KEY="xai-..."          # xAI Grok
export OPENROUTER_API_KEY="sk-or-..." # OpenRouter (300+ models)
export EVOLINK_API_KEY="el-..."          # Evolink (GPT-5, Claude, Gemini, DeepSeek & more)

# Media Providers
export ELEVENLABS_API_KEY="..."       # TTS, SFX
export RUNWAY_API_SECRET="..."        # Runway Gen-4 Turbo
export KLING_API_KEY="..."            # Kling v2.x/v3
export REPLICATE_API_TOKEN="..."      # Replicate (music)

API Keys by Command

Command	Required API Key	Model
`vibe` (default)	`OPENAI_API_KEY`	GPT-5-mini (Agent LLM)
`vibe -p claude`	`ANTHROPIC_API_KEY`	Claude Sonnet 4.6 (Agent LLM)
`vibe -p gemini`	`GOOGLE_API_KEY`	Gemini 2.5 Flash (Agent LLM)
`vibe -p xai`	`XAI_API_KEY`	Grok 4.1 Fast (Agent LLM)
`vibe -p openrouter`	`OPENROUTER_API_KEY`	OpenRouter Auto (Agent LLM)
`vibe -p evolink`	`EVOLINK_API_KEY`	GPT-5.2 via Evolink (Agent LLM)
`vibe generate image -p openai`	`OPENAI_API_KEY`	OpenAI image generation
`vibe generate image -p gemini`	`GOOGLE_API_KEY`	Gemini image generation
`vibe edit image`	`GOOGLE_API_KEY`	Gemini Nano Banana
`vibe generate speech`	`ELEVENLABS_API_KEY`	ElevenLabs
`vibe generate music`	`ELEVENLABS_API_KEY`	ElevenLabs Music (default)
`vibe generate music -p replicate`	`REPLICATE_API_TOKEN`	Replicate MusicGen
`vibe generate video -p seedance`	`FAL_API_KEY`	Seedance via fal.ai
`vibe generate video -p grok`	`XAI_API_KEY`	Grok Imagine
`vibe generate video -p kling`	`KLING_API_KEY`	Kling v2.5-turbo
`vibe generate image -p grok`	`XAI_API_KEY`	Grok Imagine
`vibe generate video -p veo`	`GOOGLE_API_KEY`	Veo 3.1
`vibe generate video -p omni`	`GOOGLE_API_KEY`	Gemini Omni (experimental)

Provider Selection Notes

Provider defaults depend on which API keys are configured. For public docs and demos, prefer explicit provider flags so the required key is obvious:

vibe generate image "..." -p openai
vibe generate image "..." -p gemini
vibe generate video "..." -p seedance
vibe generate video "..." -p veo

Use command help as the runtime source of truth for supported flags:

vibe generate image --help
vibe generate video --help
vibe edit image --help