Model Catalog

State-of-the-art open-source background removal. BiRefNet's high-fidelity dichotomous image segmentation delivers excellent edge quality on hair, fur, and fine details.

Salesforce BLIP (173M+ runs) — image captioning, visual question answering, and image-text matching in one model. The classic choice for bulk captioning at 1 credit per image.

Change anyone's hairstyle and hair color from a single photo, powered by FLUX.1 Kontext [pro]. Choose from 90+ hairstyles and 30 hair colors — or let 'Random' surprise you — while keeping the face untouched.

The famous creative upscaler (30M+ runs). Instead of just enlarging, it re-imagines detail while upscaling — with controllable creativity, resemblance, prompt guidance, and tiled diffusion for high scale factors.

CLIP ViT-L/14 embeddings for text AND images (163M+ runs) — puts both in the same vector space for cross-modal search, image dedup, and zero-shot classification. 1 credit per run.

Robust face restoration for old photos and AI-generated faces (54M+ runs). Its signature fidelity dial balances restoration quality against staying true to the original face, with Real-ESRGAN background enhancement built in.

The classic sketch-to-image model (38M+ runs) — turn any scribble or line drawing into a detailed image guided by your prompt. Draw the composition, describe the content.

Dedicated edit endpoint for xAI's Grok Imagine Image 2.0. Precise, instruction-following edits on an existing image - ideal for ads, mockups, and asset touch-ups.

Dedicated edit endpoint for xAI's Grok Imagine image model. Pass a source image URL and a prompt describing the change - the fastest, lowest-cost editing tier in the family.

Edit an existing image with Meta's Muse Image model - pass a source image URL and a prompt describing the change. Up to 10 variants per request, webp/png/jpeg output.

Dedicated edit endpoint for Nano Banana, Google's Gemini 2.5 Flash-based image model. Pass input image URLs to perform conversational editing with character consistency and multi-image fusion.

Edit endpoint for Nano Banana 2, built on Gemini 3.1 Flash Image. Adds resolution control (1K/2K/4K), Google Search grounding, and thinking mode while preserving conversational editing and multi-image fusion.

Edit endpoint for Nano Banana Pro built on Gemini 3 Pro. Professional-grade controls, legible multilingual typography, real-time grounding via Google Search, and resolution up to 2K for editing.

ByteDance Seedream 5.0 Flash image editing via Fal.AI. Fast, low-cost editing and multi-reference composition from up to 10 input images (referenced as Figure 1, Figure 2, ...), at 1K-2K output.

Turn a face photo into 6 fun styles (15M+ runs) — 3D, Emoji, Video game, Pixels, Clay, or Toy. The viral avatar generator behind countless profile-picture apps.

Turn any face photo into a fun die-cut sticker (1.6M+ runs). InstantID keeps the likeness while IP-Adapter controls take the artwork from faithful caricature to loose cartoon — with an optional 2x upscale for print quality.

Microsoft's Florence-2 all-in-one vision model — captioning, object detection, phrase grounding, OCR, and segmentation in a single API. Pick a task, optionally add text input, done.

4-step distilled FLUX.2 [klein] 9B — fast, higher-capacity sibling of Klein 4B for text-to-image generation and multi-reference image editing (up to 5 input images), with selectable output megapixels.

Very fast image generation and editing model. 4-step distilled, sub-second inference for production and near real-time applications.

Maximum-quality FLUX model supporting up to 10 reference images and advanced typography. The most capable model for complex, multi-reference creative projects.

The highest fidelity image model from Black Forest Labs. Best-in-class prompt following and the most consistent editing in the FLUX.2 lineup — preserves colors, lighting, faces, text, and objects across edits with up to 8 reference images.

Professional-grade FLUX 2 with high-quality editing and up to 8 reference image support. Excellent balance of quality, speed, and creative control.

Development version of FLUX.2 with image editing capabilities and reference image support. Ideal for iterative design workflows and experimentation.

Fast high-quality image generation, an upgrade to FLUX.1 Pro with faster speed and improved quality. Perfect for production workloads requiring both speed and fidelity.

FLUX1.1 [pro] in ultra and raw modes. Images are up to 4 megapixels — the highest-resolution tier of the FLUX 1.1 Pro family. Use raw mode for realism.

A 12 billion parameter rectified flow transformer capable of generating images from text descriptions, tuned for open, high-quality experimentation.

A premium text-based image editing model that delivers maximum performance and improved typography generation for transforming images through natural language prompts.

State-of-the-art text-based image editing model that transforms images through natural language. Excellent for style transfer, object modification, text replacement, background changes, and character consistency.

Photorealistic image generation that specifically avoids the 'AI look', producing natural-looking images indistinguishable from real photographs.

PuLID identity customization on FLUX-dev: generate photorealistic portraits of a specific person from one face photo, with markedly higher fidelity than SDXL-based variants. Tune id_weight and start_step to balance likeness against prompt editability.

Ultra-fast image generation model optimized for speed. Generates high-quality images in just 1-2 seconds, perfect for real-time applications and rapid prototyping.

Runway's Gen-4 Image model with references: combine up to 3 reference images with @tag mentions in your prompt to keep characters, objects, and locations consistent across every angle and scene, at 720p or 1080p.

Gen-4 Image Turbo is 2.5x faster and cheaper than Gen-4 Image, with the same reference-driven API: use 1 to 3 reference images with @tag mentions for consistent characters and objects, at a flat price regardless of resolution.

Tencent ARC's face restoration model — 115M+ runs, the most-run model in the restoration collection. Restores old/blurry photos and fixes faces in AI-generated images.

OpenAI's fastest model for high-quality, everyday image generation. Generate images from text or edit existing images with strong instruction following and sharp text rendering — tuned for high-volume work where speed matters. For the most precise edits, use GPT Image 2.5 Sunburst.

OpenAI's most capable image model, built for workflows where editing precision matters most. Generate and edit images with close instruction following — it keeps identity, composition and lighting while changing only what you ask. Same pricing as GPT Image 2.5 Flare, which is faster.

OpenAI's state-of-the-art image generation and editing model with strong instruction following, sharp text rendering, and detailed editing. Quality-based pricing lets you trade off cost vs. fidelity.

OpenAI's latest image generation model with better instruction following and adherence to prompts, including sharp text rendering and detailed editing.

Cheaper, faster tier of OpenAI's GPT Image 1, served through the LLM gateway via the OpenAI Images API. Same token-billed generation flow at a fraction of the cost - ideal for drafts and high-volume image generation.

xAI's newest Grok Imagine image model, served over the direct xAI API. Strong at precise compositions - infographics, ads, game assets, UI/UX mockups, and storyboards.

Generate images using xAI's Grok Imagine model over the direct xAI API - fast, low-cost text-to-image generation. For editing an existing image, use xai/grok-imagine-image/edit.

Text-prompted object detection (39M+ runs) — describe what to find in natural language ('red car, person wearing a hat') and get bounding boxes with confidence scores plus an annotated image.

Tencent's flagship 3D generation — create high-polygon textured 3D models from a text prompt OR an image, with optional PBR (physically based rendering) materials.

A middle-ground tier in Ideogram's v4 family, balancing generation speed and output quality. Delivers strong typography and photorealism at a lower cost than the Quality tier.

The highest-fidelity tier of Ideogram's v4 model family, tuned for maximum detail, realism, and typography accuracy. Best suited for final production assets where quality matters more than speed.

The fastest and cheapest model in Ideogram's v4 family, built for rapid iteration while retaining Ideogram's signature text rendering and style consistency.

The fastest and cheapest Ideogram v3 tier. V3 creates images with stunning realism, creative designs, and consistent styles.

Generate consistent characters from a single reference image. Render the same character in many styles — realistic or fiction — insert them into existing photos with mask inpainting, and rely on Ideogram's signature text rendering for legible signs and typography.

MiniMax's first image generation model with character reference support: provide a single face photo via subject_reference and generate consistent images of that person across prompts, styles, and aspect ratios — up to 9 images per request.

Kuaishou's Kling Omni 3 (O3) image-to-image via Fal.AI. Edits and transforms up to 10 reference images addressed in the prompt as @Image1 / @Element1, at 1K/2K/4K, as single images or a related series of 2-9, while preserving visual consistency.

Professional-grade upscaling from Topaz Labs, the industry standard for photo enhancement. Five specialized enhance models, up to 6x upscale, subject detection, and optional face enhancement.

Bria's commercially-safe image upscaler (130K+ runs). Increase resolution 2x or 4x with a model trained exclusively on licensed data, preserving alpha transparency — built for enterprise pipelines that require full legal liability coverage.

Krea AI's flagship text-to-image model, focused on photorealistic output with strong prompt adherence. Supports style-reference and moodboard-guided generation for consistent visual direction.

A lower-cost variant of Krea 2 Large, trading some fidelity for faster and cheaper generation while keeping the same photorealistic style focus.

Microsoft's highest-fidelity image model for production-grade text-to-image generation via Fal.AI. Built for hero imagery, detailed compositions, precise text rendering, photorealism, stylized illustration, commercial design, and visually rich concept work.

Small but capable vision-language model (14M+ runs) — ask free-form questions about any image and get detailed answers. Efficient VQA for tagging, moderation prep, and rich alt-text.

Multilingual text embeddings (74M+ runs) — 1024-dimension vectors across 100 languages including Korean. Pairs perfectly with Core.Today customer databases' vector search (knn_vector).

Meta's Muse Image model over the Meta developer API - text-to-image with built-in web and image search grounding, up to 10 images per request, webp/png/jpeg output. To edit an existing image use meta/muse-image-1.0/edit.

Google's latest fast image generation model — the high-efficiency counterpart to Nano Banana Pro. Builds on Nano Banana 2 with sharper output, better text rendering, and real-time grounding while keeping Flash-family speed and cost. Features conversational editing, multi-image fusion with up to 14 reference images, Google Web / Image Search grounding, and resolutions up to 4K.

Google's fast image generation model built on Gemini 3.1 Flash Image. The high-efficiency counterpart to Nano Banana Pro — combining Pro-level visual quality with Flash-level speed and pricing. Features conversational editing, multi-image fusion, character consistency, accurate text rendering, and Google Search grounding. Supports up to 14 reference images and resolutions up to 4K.

Google's lightweight Nano Banana 2 variant built on Gemini 3.1 Flash Image, tuned for faster and cheaper generation. Retains conversational editing, multi-image fusion, and character consistency from the full Nano Banana 2.

Google Gemini 2.5 Flash-based image generation with multimodal editing capabilities. Fast and versatile for both creation and editing tasks.

Google's state-of-the-art image generation and editing model built on Gemini 3 Pro. Creates detailed visuals with legible text in multiple languages, connects to real-time information from Google Search, and provides professional-grade creative controls. Supports up to 14 reference images and resolutions up to 4K.

The standard NSFW image classifier (127M+ runs) — returns 'normal' or 'nsfw' for any image. An essential, ultra-cheap moderation gate for UGC platforms.

Kling O1 performs precise image edits under strong reference control. Supply up to 10 reference images and character elements, then address them directly in the prompt as @Image1 or @Element1 to transform subjects, styles, and local details while preserving visual consistency.

Pruna AI's distilled text-to-image model optimized for extremely low-cost, high-throughput generation. One of the most-run community models on Replicate (15.6M+ runs) thanks to its speed and price.

PrunaAI's high-end image upscaler (575K+ runs) reaching up to 128-megapixel output. Target-megapixel or factor-based control, optional detail/realism enhancement passes, and megapixel-tiered pricing that starts at just 15 credits.

Generate stylized photos of a person from 1-4 reference photos (9M+ runs). Ten styles including Cinematic, Disney Character, and Digital Art — keep the identity, change everything else.

The stylization-focused variant of PhotoMaker: turn 1-4 photos of a person into paintings, comics, 3D art, and more with stronger style transfer. Pairs with the base PhotoMaker — use this one when style matters more than photorealism.

Turn any single photo into a polished professional business headshot, powered by FLUX.1 Kontext [pro]. Pick a background — white, black, gray, neutral, or office — and get a LinkedIn-ready portrait in one step.

Popular anime-focused image model (12M+ runs) — high-quality anime and illustration styles with img2img and inpainting support. The go-to for anime avatars and webtoon-style art.

ByteDance PuLID: tuning-free identity customization on SDXL. Give it one face photo and a prompt to generate portraits in any scene or style — no training, 4-step fast sampling, and even two-identity blending. Extremely cost-effective at 2 credits per image.

The classic image upscaler (94M+ runs). Real-ESRGAN super-resolution up to 10x with optional GFPGAN face enhancement — the go-to default for cleanly enlarging photos and AI images.

Recraft's latest image generation model, built around design taste — strong prompt accuracy, art-directed composition and integrated text rendering at standard (~1MP) resolution.

Recraft V4.1 at ~2048px resolution — the same design taste and prompt accuracy as V4.1 with higher resolution for print-ready and large-scale work.

Generates detailed SVG vector graphics with Recraft V4.1 Pro — more geometric detail and finer paths than V4.1 SVG, with clean layers and editable output at ~2048px canvas.

Generates production-ready SVG vector images with Recraft V4.1's design taste — clean geometry, structured layers and editable paths that stay crisp at any size.

Recraft's next-generation text-to-image model, improving on V3's typography, style range, and prompt adherence for production-grade brand and design assets.

Generates vector graphics directly in SVG format using Recraft V4, ideal for logos, icons, and scalable illustrations that need to stay crisp at any size.

Recraft V3 (code-named red_panda) is a text-to-image model with the ability to generate long texts, and images in a wide list of styles. SOTA in image generation per the Artificial Analysis Text-to-Image Benchmark.

Fast, affordable upscaler from Recraft designed for sharp, crisp results — a single image input with no tuning needed. Great default for UI assets, illustrations, and product images.

AI-powered background removal tool for images. Clean, accurate cutouts for any subject with professional-quality edge detection.

One of the most-run background removers on Replicate (17M+ runs). A single-parameter API that strips the background and returns a transparent PNG — at just 1 credit per image, the cheapest cutout on the platform.

Stability AI's classic SDXL (85M+ runs) — the battle-tested text-to-image model with img2img, inpainting, refiner, and LoRA support. A dependable workhorse with a huge ecosystem.

ByteDance's 4-step SDXL Lightning — the most-run model on Replicate (1B+ runs). Near-instant 1024px image generation at one of the lowest prices in the catalog.

Seedream 5.0 lite: image generation with built-in reasoning, example-based editing, and deep domain knowledge. Supports multi-reference generation with up to 14 images and sequential batch generation.

ByteDance's flagship Seedream 5.0 Pro image generation model, with built-in reasoning, precise instruction following, and multi-reference support for up to 10 images. Offers 1K and 2K resolution output with higher fidelity than Seedream 5 Lite.

Upgraded ByteDance image model with stronger spatial understanding and world knowledge. Supports single/multi-reference image-to-image editing and sequential (multi-image) generation.

ByteDance's latest image generation model with exceptional prompt understanding and creative capabilities.

Simple, massively-used OCR (91M+ runs) — extracts text from an image with a single input and returns plain text. Great default for receipts, screenshots, and scanned documents at just 1 credit.

Kuaishou's Kling Omni 3 (O3) text-to-image via Fal.AI. High-consistency image generation at 1K/2K/4K, single images or a related series of 2-9, with optional character elements (@Element1) for face/identity control.

ByteDance Seedream 5.0 Flash text-to-image via Fal.AI. Fast, low-cost image generation built for workflows where speed and budget matter, with strong realism and typography, at 1K-2K output.

Microsoft's TRELLIS image-to-3D (836K+ runs — the most-used 3D model on Replicate). Turns one or more images into a textured GLB 3D asset, with turntable render videos and optional Gaussian PLY.

Tongyi-MAI's Z-Image Turbo, a 6B-parameter text-to-image model tuned for speed — a full image in as few as 8 steps. Priced per output megapixel, so smaller presets cost proportionally less, making it one of the cheapest options for high-volume generation.

Google's image upscaler (770K+ runs). Upscale images 2x or 4x using generative AI while preserving natural detail, with adjustable output compression — a simple, reliable enhancer at a flat price.

Pruna's ultra-fast Z-Image Turbo (48M+ runs). Megapixel-priced image generation up to 2048x2048 — pay exactly for the resolution you generate, with excellent price/performance for high-volume use.

Video understanding and captioning — ask free-form questions about a video and get detailed answers about actions, scenes, and content. Great for video search indexing and moderation prep.

Turn a Seedance 2.5 draft into the final 1080p video. Make a cheap 480p draft with bytedance/seedance-2.5 and draft: true, check it, then send that job's job_id here. The final reuses the draft's prompt, reference media, duration, aspect ratio, seed and audio setting, so what you approved is what you get.

Edit an existing video with xAI's Grok Imagine Video model. The output inherits the source video's duration, resolution (720p cap), and aspect ratio.

Kling O3 native 4K video editing via Fal.AI. Edits an existing video (cited as @Video1) from a text instruction — change the environment, restyle the scene, or swap a subject for an element image (@Element1) — keeping the original motion and timing, and outputs native 4K directly. Input clips of 3-15 seconds.

Black Forest Labs' multimodal FLUX 3 generates 5-20 second video with synchronized audio from a text prompt, from one or more images (first frame, first+last frame, or a storyboard of up to 10), or by continuing an existing clip — at 720p or 1080p, with a low-cost draft mode.

Google's fast multimodal video generation and editing model with native audio, using the Interactions API. Text-to-video, image-to-video, keyframe interpolation, reference-to-video, and video editing at 360p-4K.

Google Gemini Omni Flash text-to-video via Fal.AI. Generates a video directly from a descriptive text prompt — pacing and audio (dialogue, background music) are controlled in the prompt itself, in 16:9 or 9:16 at 3-10 second durations.

Runway's Gen-4.5 model, offering state-of-the-art video motion quality, prompt adherence, and visual fidelity for text-to-video and image-to-video generation.

Bria's chroma-key background removal via Fal.AI — removes the green-screen background from a video with automatic green-spill suppression, outputting on a transparent background (alpha-capable codecs) with clean, professional edges.

xAI's Grok Imagine Video 1.5 over the direct xAI API: multi-mode video creation - text-to-video and image-to-video with synchronized audio, up to native 1080p.

Generate videos with xAI's Grok Imagine Video model over the direct xAI API. Text-to-video and image-to-video with synchronized audio. For editing an existing clip, use xai/grok-imagine-video/edit.

MiniMax H3 multimodal video generation via Replicate — one endpoint for text-to-video, first/last-frame image-to-video, and reference-based generation with images, videos, and audio cited in the prompt. 4-15 second duration at 768P or 2K.

Realistic human motion video generation with advanced character consistency and natural movement.

Lower-latency version of Hailuo 2.3 optimized for faster generation while maintaining good quality for human motion videos.

Alibaba's Happy Horse 1.1 generates videos from text, animates a single image, or builds a video from multiple reference images. Supports 720p and 1080p, 3-15 second durations, and five aspect ratios.

FLUX.3 is Black Forest Labs' frontier video model. This endpoint animates a single still image into coherent, natural motion at 720p or 1080p for 5-20 seconds, with synchronized audio generated alongside the video and a tunable safety tolerance.

Google Gemini Omni Flash 1.1 image-to-video via Fal.AI. Animates a still image into 3-10 second video with synchronized audio, or interpolates between a first and an optional end frame — 16:9 or 9:16 at 360p to 4K.

Kling 3.0 Turbo Pro image-to-video via Fal.AI. Animates a first-frame image into high-quality 1080p video with native audio, improved lip-sync, and multi-shot storyboards, 3-15 seconds.

Kling 3.0 Turbo Standard image-to-video via Fal.AI. Animates a first-frame image into 720p video with native audio — quick, affordable image-driven motion with optional multi-shot storyboards, 3-15 seconds.

MiniMax H3 Max image-to-video via Fal.AI. fal's post-trained variant of MiniMax Hailuo-03 (H3), tuned for stronger prompt adherence and better aesthetics — animates a first-frame image (optional last frame for keyframe control) into 5-15 second clips with native audio at 480P/768P native or 1080P refined output.

MiniMax Hailuo-03 (H3) image-to-video via Fal.AI. 2K video generation from a first-frame image, 5-15 second duration, native audio, with optional first-to-last keyframe control via end_image_url.

ByteDance Seedance 2.5 image-to-video via Fal.AI. Animates a single still into a native clip of up to 30 seconds at 480p/720p without the drift or stitching of multi-clip workflows, with optional end-frame control and synchronized audio generated in the same pass.

Kling Video 3.0 Omni: a unified multimodal video model that generates and edits video from text, images, reference images, and existing video. Combines text-to-video, image-to-video, reference-based generation, and video editing with native audio and multi-shot control.

Kling Video 3.0: Kuaishou's flagship text/image-to-video model generating cinematic videos up to 15 seconds with multi-shot control, native audio, start/end frame images, and a dedicated 4K mode.

Kling 2.6 Pro: top-tier image-to-video with cinematic visuals, fluid motion, and native audio generation. Audio generation is enabled by default.

Cinematic-grade video generation with enhanced motion and scene coherence. Top-tier Kling model for professional output.

Kling v2.1 with 720p/1080p support and frame transition capabilities for smooth, high-quality video generation.

ByteDance's open-source lipsync — re-syncs a video's mouth movements to any audio track using latent diffusion. State-of-the-art open lipsync quality for dubbing and localization.

Add AI-generated sound to any video (5M+ runs) — synthesizes synchronized audio (ambience, effects, foley) from the video content and an optional text prompt. The perfect finisher for silent AI-generated clips.

Pruna's quality-focused successor to P-Video. Generates video from text, a starting image, or an audio track, with native-speech lip-sync, sharp close-ups and strong identity consistency, up to 1080p / 48 fps, plus a low-cost draft mode for fast iteration.

Pruna's highest-quality video model. Generates video from text or images with cost, speed and quality modes, first- and last-frame control, and durations from 5 to 15 seconds at 480p or 768p.

PrunaAI's fast video generator with a built-in draft mode for rapid creative iteration. Text-to-video, image-to-video, and audio-conditioned generation in a single endpoint, with clips up to 20 seconds — one of the longest durations on the platform.

Pruna AI's motion-transfer video model — animates a reference image with the motion and audio of a source video. Optimized for speed and cost (about 5.24s of generation per 1s of output video).

Pruna AI's talking-avatar video model — animates a single input image to speak provided text (voice_script) or an uploaded audio track, with selectable Gemini-family preset voices, language, and visual delivery prompt.

PixVerse's flagship video generation model. Generates cinematic videos with synchronized audio, multi-shot sequences with scene transitions, astonishing physics, and precise camera control at up to 1080p.

Advanced video generation with special effects capabilities and anime-optimized output, supporting multiple visual styles.

Luma's flagship Ray video model. Text-to-video and keyframe (start/end image) generation with optional HDR-encoded output and professional EXR export.

Luma's Ray Flash 2 generates 5 or 9 second 720p videos faster and cheaper than Ray 2. Supports keyframe control via start and end images, seamless loops, and a rich library of camera motion concepts — the first Luma model on the platform.

Video upscaling with Real-ESRGAN — enhance videos to FHD, 2K, or 4K frame by frame. The go-to open-source video upscaler for old footage and AI-generated clips.

Kling O3 native 4K video-to-video reference via Fal.AI. Generates a new 4K shot that follows a reference video (cited as @Video1) — its camera motion, pacing, and framing — while integrating characters/objects from element images (@Element1) and style images (@Image1). Outputs 4K directly with no upscaling pass, 3-15 seconds.

MiniMax Hailuo-03 (H3) reference-to-video via Fal.AI. Generates video from multimodal references — subject/style images, motion video clips, and audio clips, each cited in the prompt by order — keeping subjects consistent while following the referenced motion and audio.

ByteDance Seedance 2.5 reference-to-video via Fal.AI. Generates video from up to 50 multimodal references — up to 30 images, 10 videos, and 10 audio files — locking a character, set, and palette across a full 30-second take. Reference inputs are addressed from the prompt as @Image1, @Video1, @Audio1, and drive motion transfer, editing, extension, and lip-sync.

Alibaba Wan 3.0 Prime reference-to-video via Fal.AI — the higher-fidelity Wan 3.0 tier. Combines up to 10 reference images, 5 reference video clips, and 5 reference audio clips — each cited in the prompt by position (Image 1, Video 1, Audio 1) — into one coherent video that keeps character identity, style, motion, and sound cues. 2-30 second output at 480p/720p/1080p with optional generated audio.

Alibaba Wan 3.0 reference-to-video via Fal.AI. Combines up to 10 reference images, 5 reference video clips, and 5 reference audio clips — each cited in the prompt by position (Image 1, Video 1, Audio 1) — into one coherent video that keeps character identity, style, motion, and sound cues. 2-30 second output at 480p/720p/1080p with optional generated audio.

Talking-head video from a single photo and an audio track — animates the face with natural head motion and eye blinks, with optional GFPGAN face enhancement.

ByteDance's flagship multimodal video model with native audio, native 30-second single-pass generation, and large multimodal reference sets (up to 30 images, 10 videos, 10 audios). Supports text-to-video, image-to-video, first/last-frame control, video editing, extension, and lip-sync.

ByteDance's next-generation multimodal video model with native synchronized audio. Combines up to 9 reference images, 3 videos, and 3 audio files in a single generation for character-consistent, lip-synced video creation, editing, and extension.

A faster, cheaper variant of Seedance 2.0 for quicker video generation with multimodal reference inputs (up to 9 images, 3 videos, 3 audios) and native audio, at 480p or 720p.

Lighter, cheaper variant of ByteDance's Seedance 2.0. Native audio, multimodal reference inputs (images/videos/audio), text-to-video and image-to-video, capped at 720p (no 1080p/4K tier).

A lightweight ByteDance video generation model offering text-to-video and image-to-video support for 4-12 second videos at 480p, 720p, or 1080p resolution.

A pro version of Seedance that offers text-to-video and image-to-video support for 2-12 second videos, at 480p, 720p, and 1080p resolution.

ByteDance's cinematic video generation model with fast generation speed and professional output quality.

OpenAI's video generation model with realistic physics simulation and audio generation capabilities, producing highly coherent videos.

OpenAI's most advanced synced-audio video generation model. The premium tier of Sora 2 with higher fidelity, up to 1024p resolution, and image-to-video via an input reference frame.

Google Gemini Omni Flash 1.1 text-to-video via Fal.AI. Generates 3-10 second video with synchronized native audio from a text prompt, grounded in Gemini's real-world knowledge and physics understanding, with cinematic camera control in natural language — 16:9 or 9:16 at 360p to 4K.

Kling 3.0 Turbo Pro text-to-video via Fal.AI. Generates high-quality 1080p video with native audio from a text prompt, with improved lip-sync and multi-shot storyboards, 3-15 seconds.

MiniMax H3 Max text-to-video via Fal.AI. fal's post-trained variant of MiniMax Hailuo-03 (H3), tuned for stronger prompt adherence and better aesthetics — 5-15 second clips with native audio, six aspect ratios (21:9 through 9:16), and 480P/768P native or 1080P refined output.

MiniMax Hailuo-03 (H3) text-to-video via Fal.AI. State-of-the-art 2K video generation from a text prompt, 5-15 second duration, native audio, and wide aspect-ratio support (21:9 through 9:16).

ByteDance Seedance 2.5 text-to-video via Fal.AI. Generates a native single-shot clip of up to 30 seconds at 480p/720p from a text prompt alone, reasoning about the whole shot at once so motion, lighting, and subject identity stay coherent from first frame to last. Synchronized audio (dialogue, sound effects, music) is generated in the same pass.

Veed Lipsync v2 via Fal.AI. Replaces a source video's mouth movements to articulate a new audio track — takes any source video + a new audio track and produces a lip-synced output video.

Google's state-of-the-art video generation model with built-in audio generation, producing cinematic-quality videos with synchronized sound.

Fast version of Veo 3.1 with audio generation, optimized for speed while maintaining high quality output.

Google's cost-efficient Veo 3.1 variant for high-volume video generation — text-to-video, image-to-video and first/last-frame interpolation at 720p or 1080p, 4/6/8 seconds, with natively synchronized audio on every clip.

FFmpeg-powered video utilities (20M+ runs) — convert to mp4/gif, extract audio as mp3, or dump zipped frames, all with one task parameter. The Swiss-army knife for media pipelines.

Alibaba's Wan 3.0 generates video from a text prompt or a starting image, with cinematic motion and support for 480p, 720p, and 1080p output up to 30 seconds.

Alibaba's Wan 3.0 Prime generates video from a text prompt (or an optional starting image) with coherent motion and lighting, at 480p, 720p or 1080p and 2 to 30 seconds long.

Alibaba Wan 2.7 image-to-video model. Animates a first frame (and optional last frame or continuation clip) with audio synchronization, up to 15 seconds, at 720p or 1080p.

Alibaba Wan 2.7 text-to-video model. Supports up to 15 seconds, audio synchronization for voice/music, multilingual prompts, and prompt expansion, at 720p or 1080p.

Image-to-video model with lip sync support, animating still images into realistic videos with natural motion.

Fast image-to-video variant of Wan 2.5, optimized for rapid generation of animated videos from still images.

Text-to-video model with audio synchronization support, producing high-quality videos from text prompts with natural motion.

Fast text-to-video generation variant of Wan 2.5, optimized for speed with good quality output.

A very fast and cheap PrunaAI-optimized version of Alibaba's Wan 2.2 A14B image-to-video model. Turns a single still image plus a prompt into a short animated clip at 480p or 720p, with an optional frame-interpolation pass for smoother motion.

A very fast and cheap PrunaAI-optimized version of Alibaba's Wan 2.2 A14B text-to-video model. Generates short clips at 480p or 720p directly from a text prompt, with 30 FPS frame interpolation enabled by default — the text-to-video sibling of Wan 2.2 I2V Fast.

Suno's text-prompted generative audio model. Produces speech with nonverbal sounds like [laughs] and [sighs], plus music and sound effects, in 100+ speaker presets across 13 languages. Returns audio plus an optional .npz history file for voice continuity.

Resemble AI's production-grade open-source TTS with unique emotion exaggeration control and instant voice cloning from a short reference audio. MIT-licensed and benchmarked against leading closed-source systems.

Chatterbox open-source TTS in 23 languages with instant voice cloning and emotion exaggeration control. Max 300 characters per request.

Resemble AI's fastest open-source TTS without sacrificing quality. 20 pre-made voices, paralinguistic tags like [sigh] and [chuckle], and optional instant voice cloning from 5s+ reference audio. Max 500 characters per request.

ElevenLabs Eleven v4 text-to-speech via Fal.AI. Expressive speech with inline audio tags ([whispering], [excited]), IPA pronunciation, stability/similarity controls, up to 5,000 characters per request and optional character-level timestamps.

ElevenLabs Eleven v4 Turbo text-to-speech via Fal.AI. The Turbo tier of Eleven v4 at half the rate, with the same audio tags, IPA pronunciation, voice controls and optional character-level timestamps.

Google Gemini 3.8 Flash TTS via Fal.AI. Expressive single-speaker narration with 30 prebuilt voices, delivery directions given as separate style instructions, and inline vocal events such as <laugh> and <sigh>. Returns WAV audio.

Google Gemini 3.1 Flash native-audio TTS via Replicate. Style instructions (separate from the spoken text) plus 30 preset voices and 40+ language/locale codes, with inline markup tags for expressive delivery ([sigh], [whispering], [shouting], etc.).

OpenAI's gpt-transcribe speech recognition, the successor to whisper-1 - plain transcript text with detected languages, optional language hints and keyword spelling. No timestamps. Billed per second of audio (0.174375 credits/s, about 10.5 credits per minute), up to 1 hour / 25MB per request.

Whisper large-v3 optimized for speed (38M+ runs) — transcribes roughly 150 minutes of audio in under 100 seconds using batched inference. Chunk-level or word-level timestamps.

Mureka instrumental generation via fal.ai. Produces an instrumental track from a text description with the chosen Mureka model version (7.6 / 8 / 9 / 9.5). Returns an MP3 file.

Google's latest Lyria 3.5 music generation model via fal.ai. Generates full-length songs up to a few minutes with vocals and lyrics; control length and structure with duration hints or section timestamps in the prompt. Supports English, German, Spanish, French, Hindi, Japanese, Korean and Portuguese. Returns an MP3 file.

Google's Lyria 3 music generation model. Generates 30-second music clips with vocals, lyrics and multi-language support from a text prompt, optionally inspired by up to 10 reference images. Returns an MP3 file.

Google's most capable music generation model. Generates full-length songs up to 3 minutes with vocals, lyrics, song structure tags ([Verse], [Chorus], [Bridge]) and timestamp control from a text prompt, optionally inspired by up to 10 reference images. Returns an MP3 file.

Meta's Muse Voice Transcribe speech-to-text over the Meta developer API - 25 languages with code-switching (Korean included), speaker diarization, endpointing, keyword and language biasing. Billed per second of audio (0.11625 credits/s, about 7 credits per minute), up to 10 minutes per request.

Suno's latest model, with the highest audio quality and richest arrangements. Creates two complete songs from a text description, or from exact lyrics with style and title in custom mode. Returns audio, cover image, and metadata for each track.

Suno V5 music generation with improved audio quality and musicality. Creates two complete songs from a text description, or from exact lyrics with style and title in custom mode. Returns audio, cover image, and metadata for each track.

Suno V4.5 music generation. Creates two complete songs from a text description, or from exact lyrics with style and title in custom mode. Returns audio, cover image, and metadata for each track.

MiniMax Music 3 via Fal.AI. Generates complete songs with vocals up to five minutes long from a music description (genre, mood, BPM, key, vocals, arrangement) and structured lyrics. Returns 44.1 kHz stereo WAV.

Generate full-length songs or instrumentals (up to ~6 minutes) from a text prompt. 99%+ accurate BPM/key control, 14+ structure tags, auto-generated lyrics, and instrumental-only mode.

Generate full-length songs (up to ~5 minutes) with vocals, lyrics, and rich instrumentation. 14 structure tags, style-aware mixing, and an expanded instrument library including orchestral and traditional instruments.

MiniMax Music 1.5 via Replicate. Full-length songs up to 4 minutes with natural vocals and rich instrumentation from a style prompt and structured lyrics. The previous generation to Music 2.5 / 2.6 at a fifth of the price.

MiniMax Music-01 via Replicate. Quickly generates up to 1 minute of music with lyrics and vocals in the style of a reference track. Give a song, voice and/or instrumental reference file (WAV or MP3 longer than 15 seconds) and optional lyrics.

ElevenLabs Music via Replicate. Composes a song from a text prompt with a target length of 5 seconds to 5 minutes, optional vocals, and MP3 / WAV output. Billed per second of the requested length — a 10-second clip costs about 190 credits.

Reimagine any song in a different style. The model extracts the melodic structure from the input audio and regenerates the track — the melody and duration stay the same, but voice, instruments, genre, and arrangement can all change.

CassetteAI's fast, low-cost music generator via fal.ai. Produces 10 to 180 seconds of music from a text prompt at under one credit per second. Returns a WAV file.

Meta's MusicGen via Replicate. Generates music from a text prompt or a reference melody (stereo / melody / large variants), can continue an input clip, and outputs MP3 or WAV. Billed per second of requested duration.

Alibaba's unified TTS with three modes: preset speakers (custom_voice), instant voice cloning from reference audio (voice_clone), and creating a brand-new voice from a text description (voice_design).

Meta's Muse Voice Transcribe over a WebSocket - stream 16-bit PCM (24 kHz or 16 kHz mono) and receive partial and final transcripts while you speak, with endpointing and diarization. Billed per second of audio Meta processed (0.11625 credits/s, 35 credits per 5-minute block), sessions up to 20 minutes.

Inworld's most expressive TTS with natural-language steering — place bracketed instructions like [speak quickly] before the text they apply to. Real-time latency and 15+ language support.

Fish Audio's current production TTS: 83 languages, bracketed natural-language emotion and tone control, 1M+ Voice Library voices and voice cloning. 61% listening-test win rate over S2 Pro with ~70 ms time-to-first-audio upstream. Billed per UTF-8 byte.

Fish Audio S2-generation TTS: 80+ languages, bracketed natural-language emotion and tone tags, Voice Library voices. Same price as S2.1 Pro; kept for pipelines tuned on S2. Billed per UTF-8 byte.

Fish Audio S1 (legacy) TTS with the fixed parenthesis emotion tag set — (happy), (sad), (whispering) — and Voice Library voices. Same price as the S2 family; for prompts written against S1. Billed per UTF-8 byte.

Suno V5 sound effect generation. Creates two short sound effect variants from a text description, with optional loop mode, tempo, and musical key controls. Ideal for UI sounds, game audio, and video foley.

Mureka song generation via fal.ai. Two modes: lyrics-to-song (send your lyrics, the prompt steers the style) or prompt-to-song (prompt plus genre styles, the model writes the lyrics). Choose the Mureka model version. Returns an MP3 with timed lyric sections.

Ranked #1 on Artificial Analysis Speech Arena and Hugging Face TTS Arena. Broadcast-quality TTS with autoregressive Transformer + Flow-VAE decoder, 32+ languages, voice cloning, natural interjections, and emotion control.

Low-latency MiniMax Speech 2.8 Turbo with under 250ms latency, 40+ languages, voice cloning, natural interjections, and real-time pricing. Ideal for interactive and real-time applications.

Studio-quality multilingual text-to-speech with nuanced prosody, emotion control, and premium voices for professional applications.

Fast multilingual text-to-speech with emotional control, optimized for real-time applications with low latency.

Low-latency text-to-speech model with multilingual support, emotional voice control, and 300+ voice options.

Stability AI's Stable Audio 2.5. Generates high-quality music and sound up to 190 seconds from a text prompt with step count and CFG scale controls. Flat price per generation regardless of length. Returns an MP3 file.

Sonilo v1.1 text-to-music via Fal.AI. Generates up to 3 distinct music tracks (up to 600 seconds each) from a text description of the desired sound.

Inworld's highest-quality realtime TTS with under 200ms latency. Supports SSML break tags for pauses, emotion markups like [happy], and 15 languages.

NAVER Clova Voice Premium TTS with 108 voices across 6 languages. High-quality Korean voice synthesis with emotion control, Pro voices, and bilingual support.

High-quality, low-latency ElevenLabs text-to-speech in 32 languages. The same 26 premium voices as v3 at half the price, optimized for real-time and high-volume use.

ElevenLabs' most expressive text-to-speech model. 26 premium voices, inline audio tags like [laughs] and [whispers], fine-grained style and stability controls, and 70+ language support.

ElevenLabs Music v2.5 via fal.ai. Composes a song from a text prompt with a target length of 3 seconds to 10 minutes, optional vocals, and MP3 / Opus output. Billed per output minute, rounded up — the best value for songs of one minute or more.

ElevenLabs Music v2 via fal.ai. Composes a song from a text prompt with a target length of 3 seconds to 10 minutes, optional vocals, and MP3 / Opus output. Billed per output minute, rounded up — the best value for songs of one minute or more.

ElevenLabs Multilingual v2 text-to-speech in over 30 languages. Stable, proven voice quality with the same premium voice lineup and fine-grained voice settings.

Sonilo v1.1 video-to-sound-effects via Fal.AI. Adds AI-generated sound (ambience, effects, foley) to an input video — auto-captions the scene if no prompt is given, or accepts per-segment sound descriptions for finer control.

Separate a Suno-generated track into clean vocal and instrumental stems. Takes the taskId and audioId from a previous Suno music generation and returns separated vocalUrl/instrumentalUrl audio files.

Clone any voice from a 10-second to 5-minute audio sample. Returns a custom voice_id you can pass to MiniMax speech models, plus a preview clip synthesized with the cloned voice.

OpenAI's hosted whisper-1 speech recognition with the OpenAI API parameters - json or verbose_json output, segment and word timestamps (timestamp_granularities), language hint and prompt. Billed per second of audio (0.2325 credits/s, about 14 credits per minute), up to 1 hour / 25MB per request.

OpenAI's Whisper large-v3 speech recognition (144M+ runs) — the standard for transcription. Automatic language detection across ~100 languages, English translation, and plain text / SRT / VTT output formats.

Whisper transcription with speaker diarization (8M+ runs) — returns who said what, with per-segment speaker labels and timestamps. The go-to for meetings and interviews.

Coqui XTTS-v2 multilingual voice-cloning TTS. Clone a voice from a single short audio sample and speak in 16 languages — one of the most popular open-source voice cloning models.

Anthropic's Fable-tier model, positioned above Opus. 1M-token context window and 128K max output tokens. Priced at $10/$50 per M tokens with prompt caching (read $1.00/M, 5-minute write $12.50/M) and web search. Compatible with the Anthropic Messages and OpenAI Chat Completions formats.

Fast, cost-effective model for everyday tasks. Great balance of speed, intelligence, and cost for high-volume applications.

Anthropic's current Opus model (September 2026), cheaper than Opus 5 at $4/$20 per M tokens. 1M-token context window and 128K max output tokens, with prompt caching (read $0.20/M, i.e. 5% of input; 5-minute write $5.00/M) and web search. Compatible with the Anthropic Messages and OpenAI Chat Completions formats.

Anthropic's Opus 5 model (superseded by the cheaper Claude Opus 5.5; Claude Fable 5.1 is the top tier), with a 1M-token context window by default and 128K max output tokens. Same pricing as Opus 4.5–4.8 ($5/$25 per M tokens) with prompt caching (read $0.50/M, write $6.25/M) and web search. Compatible with the Anthropic Messages and OpenAI Chat Completions formats.

Anthropic's most capable Opus-tier model, with a 1M-token context window (200K on some surfaces), 128K max output tokens, and knowledge cutoff to January 2026. Builds on Opus 4.7 with stronger long-horizon agentic coding, better tool triggering, and adaptive thinking that reasons only when a turn needs it. Compatible with the Anthropic Messages and OpenAI Chat Completions formats.

Anthropic's latest flagship model with reliable knowledge cutoff to January 2026 and 128K max output tokens. Builds on Opus 4.6 with improved reasoning, coding, and instruction-following while staying compatible with the Anthropic Messages and OpenAI Chat Completions formats.

Anthropic's most capable model. Delivers breakthrough performance in reasoning, coding, and complex analysis with enhanced safety and instruction following.

Anthropic's most powerful model for highly complex tasks. Exceptional at research, analysis, and creative projects requiring deep expertise.

Anthropic's newest Sonnet model (September 2026), balancing speed, cost, and intelligence at $2/$10 per M tokens. 1M-token context window and 128K max output tokens, with prompt caching (read $0.10/M, 5-minute write $2.50/M) and web search. Compatible with the Anthropic Messages and OpenAI Chat Completions formats.

Anthropic's Sonnet 5 model, tuned for the best balance of speed, cost, and intelligence. Features a 1M-token context window, 64K max output tokens, and knowledge cutoff to August 2025. Supports adaptive thinking that reasons only when a turn needs it, vision, and tool use. Priced at $2/$10 per M tokens (cache read $0.20/M). Compatible with the Anthropic Messages and OpenAI Chat Completions formats.

The successor to Claude Sonnet 4.5. Anthropic's frontier Sonnet-tier model for complex reasoning, nuanced understanding, and coding tasks with exceptional instruction following, at the same token pricing as Sonnet 4.5.

Anthropic's most intelligent and capable Sonnet model. Best-in-class for complex reasoning, nuanced understanding, and coding tasks with exceptional instruction following.

DeepSeek's strongest V4 model (checkpoint V4-Pro-0813). Advanced reasoning with thinking mode on by default, a 1M-token context window, and discounted cache reads - still far cheaper than comparable frontier models.

DeepSeek's fast, low-cost model - currently DeepSeek V4.1 Flash, with text and image input. Thinking mode is on by default, with a 1M-token context window, up to 384K output tokens, and cache reads at ~98% off. $0.30 input / $1.20 output per million tokens.

Google's flagship reasoning model (GA 2026-10-13). A 1M-token context window with 65,536 max output tokens (thinking included), multimodal input, adjustable thinking levels and agentic tool use, priced at $4 input / $20 output per million tokens with implicitly cached inputs at $0.20/M.

Google's newest Flash-tier model (released 2026-09-02). A 1M-token context window with 65,536 max output tokens, multimodal input, thinking and tool use, priced at $1.50 input / $7.50 output per million tokens with cached inputs at $0.15/M.

Google's Gemini 3.7 Flash (released 2026-08-13), an earlier release in the Gemini 3.x Flash line - newer Flash generations (3.8) are available. A 1M-token context window with 65,536 max output tokens, multimodal input, thinking and tool use, priced at $1.50 input / $7.50 output per million tokens with cached inputs at $0.15/M.

Google's Gemini 3.6 Flash (released 2026-07-21), an earlier release in the Gemini 3.x Flash line - newer Flash generations (3.7 and 3.8) are available. A 1M-token context window with 65,536 max output tokens, multimodal input, thinking and tool use, priced at $1.50 input / $7.50 output per million tokens with cached inputs at $0.15/M.

Google's latest Flash-tier workhorse model. Flat pricing across all input modalities (text, image, video, audio, PDF) at $1.50/$9 per million tokens with no long-context premium, a 1M token context window, and 65,536 max output tokens. Cached inputs get a 90% discount.

Google's lightweight Gemini 3.5 model (released 2026-07-21) for high-throughput, budget-conscious workloads. $0.30/$2.50 per million tokens with cached input at $0.03/M. Successor to Gemini 3.1 Flash Lite, which Google retires on 2027-05-07.

Generally available (GA) Gemini 3.1 Flash with native image generation - the stable successor to gemini-3.1-flash-image-preview, at the same price. Can generate images directly in chat responses alongside text. Features separate pricing for text and image output tokens.

Gemini 3.1 Flash with native image generation capabilities. Can generate images directly in chat responses alongside text. Features separate pricing for text and image output tokens.

Stable release of the ultra-lightweight Gemini 3.1 Flash Lite. The most cost-effective Gemini model at $0.25/$1.50 per million tokens, with cached input (90% discount), audio input, and batch (50% discount) support. Ideal for high-throughput, budget-conscious applications.

Lightweight, lower-cost sibling of Gemini 3.1 Flash Image with native image generation. Generates images directly in chat responses alongside text, with separate pricing for text and image output tokens: $0.25 input, $1.50 text output and $30 image output per million tokens.

Ultra-lightweight variant of Gemini 3.1 Flash. The most cost-effective Gemini model with support for cached input and audio input. Ideal for high-throughput, budget-conscious applications.

Google's latest and most capable Gemini model in preview. Features dynamic pricing that adjusts based on context length, with enhanced pricing for inputs over 200K tokens.

Google's most advanced reasoning model with state-of-the-art multimodal understanding, PhD-level reasoning, and leading coding performance.

Generally available (GA) premium image generation model in the Gemini 3 Pro family - the stable successor to gemini-3-pro-image-preview, at the same price. Generates high-quality images directly in chat with the highest fidelity among Gemini image models. Image output tokens are priced at 10x text output tokens.

Google's premium image generation model within the Gemini 3 Pro family. Generates high-quality images directly in chat with the highest fidelity among Gemini image models. Image output tokens are priced at 10x text output tokens.

Google's fast and efficient model with built-in thinking capabilities. Great balance of speed, reasoning, and cost for high-volume applications.

The cheapest tier of the Gemini 2.5 family, optimized for high-volume, latency-sensitive workloads. Delivers 2.5-generation quality at a fraction of the cost, ideal for classification, extraction, and real-time chat at scale.

Google's most capable model with state-of-the-art reasoning and 1M token context. Excels at complex coding, math, and multi-document analysis.

Google's second-generation Gemini embedding model and the successor to gemini-embedding-001. Generates vector representations for semantic search, clustering, classification, and RAG at $0.20 per million text input tokens.

Google's text embedding model for generating vector representations. Optimized for semantic search, clustering, and similarity tasks.

Z.ai's GLM 5.3 served through the OpenRouter aggregator (listed 2026-08-18) - a large-scale text reasoning model built for complex software engineering and long-horizon agent tasks, and the successor to GLM 5.2. 1.31M-token context window, billed at OpenRouter's actual usage cost.

Z.ai's GLM 5.3 Flash served through the OpenRouter aggregator (listed 2026-08-26) - a low-cost native multimodal model (text, image, video input) for efficient coding and long-horizon agent tasks. Its hybrid sparse and linear attention keeps long-context behavior accurate across a 1.31M-token window. Billed at OpenRouter's actual usage cost.

Z.ai's GLM 5.2 served through the OpenRouter aggregator - a large-scale text reasoning model with a 1M-token context window, suited for long-horizon agent workflows and project-level software engineering. Billed at OpenRouter's actual usage cost.

Z.ai's GLM 4.7 served through the OpenRouter aggregator - a flagship text model with enhanced programming capabilities and more stable multi-step reasoning/execution, showing clear gains on complex agent tasks. 200K context, billed at OpenRouter's actual usage cost.

OpenAI's GPT-6.1 Sol — the Sol tier of the GPT-6.1 generation at $2/$10 per million tokens, with cached inputs at a 95% discount ($0.10/M, half of GPT-6 Sol's cached rate). 1.05M token context window (up to 922K input), 128K max output tokens, text and image input. Supports Fast mode and Ultrafast mode. Prompts above 272K input tokens are billed at 2x input / 1.5x output for the whole request.

OpenAI's GPT-6.1 Terra — a fast, efficient model of the GPT-6.1 generation for visual reasoning, precise computer use and clear, natural writing, at $1/$5 per million tokens with cached inputs at a 95% discount ($0.05/M). 1.05M token context window (up to 922K input), 128K max output tokens, text and image input. Supports Fast mode. Prompts above 272K input tokens are billed at 2x input / 1.5x output for the whole request.

OpenAI's most capable model (GPT-6 generation, September 2026), built for the hardest end-to-end work: complex reasoning, coding, computer use, research and document creation. 1M token context window, 128K max output tokens, cached inputs at a 90% discount. Prompts above 272K input tokens are billed at 2x input / 1.5x output for the whole request.

The smallest, lowest-cost tier of OpenAI's GPT-6 generation, at $0.10/$0.50 per million tokens — built for high-volume, latency-sensitive work such as classification, extraction, routing and short-form generation. Cached inputs at a 90% discount. Prompts above 272K input tokens are billed at 2x input / 1.5x output for the whole request.

The mid tier of OpenAI's GPT-6 generation (in the API since September 14, 2026) — strong reasoning, coding and multimodal performance at $2/$10 per million tokens, a fifth of GPT-6 Astra. 1.05M token context window, 128K max output tokens, cached inputs at a 90% discount. Prompts above 272K input tokens are billed at 2x input / 1.5x output for the whole request.

The fast, low-cost tier of OpenAI's GPT-5.6 family (GA July 2026). Luna is built for high-throughput workloads at $0.20/$1.20 per million tokens, with a 1M token context window, 128K max output tokens, and cached inputs at a 90% discount. Prompts above 272K input tokens are billed at 2x input / 1.5x output for the whole request.

The flagship tier of OpenAI's GPT-5.6 family (GA July 2026). Sol delivers the strongest reasoning, coding, and multimodal performance of the generation with a 1M token context window, 128K max output tokens, and cached inputs at a 90% discount. Prompts above 272K input tokens are billed at 2x input / 1.5x output for the whole request.

The balanced tier of OpenAI's GPT-5.6 family (GA July 2026). Terra offers near-flagship quality at $2/$12 per million tokens (40% of Sol), with a 1M token context window, 128K max output tokens, and cached inputs at a 90% discount. Prompts above 272K input tokens are billed at 2x input / 1.5x output for the whole request.

Maximum-capability tier of the GPT-5.5 family (dated snapshot). Uses more compute to think longer and deliver the most reliable answers on the hardest reasoning, coding and analysis tasks. Served only via the OpenAI Responses API. $30/$180 per million tokens (no cached-input discount); prompts above 272K input tokens bill the whole request at $60/$270.

OpenAI's newest flagship model with a 1.05M token context window and 128K max output tokens. Supports cached inputs at 10× discount and improved reasoning, coding, and multimodal performance over the GPT-5.4 series.

Maximum-capability tier of the GPT-5.4 family (dated snapshot). Uses more compute to think longer and deliver the most reliable answers on the hardest reasoning, coding and analysis tasks. Served only via the OpenAI Responses API. $30/$180 per million tokens (no cached-input discount); prompts above 272K input tokens are billed at 2x input / 1.5x output for the whole request ($60/$270 per million tokens).

OpenAI's newest flagship model with 1M context window and 128K output tokens. Delivers top-tier reasoning across all domains with adjustable reasoning effort levels from none to xhigh. Prompts above 272K input tokens are billed at 2x input / 1.5x output for the whole request ($5 / $22.50 per 1M tokens, cached $0.50).

Fast and cost-efficient variant of GPT-5.4 with 400K context window and 128K output tokens. Excellent balance of performance and affordability for everyday tasks.

Ultra-lightweight and fastest GPT-5.4 variant with 400K context and 128K output. Designed for high-throughput, low-latency applications at minimal cost. Supports MCP for tool integration.

Codex model on the GPT-5.3 base, optimized for agentic coding harnesses and served via the OpenAI Responses API. 400K token context window, 128K max output tokens, cached inputs at a 90% discount.

Maximum-capability tier of the GPT-5.2 family (dated snapshot). Thinks longer with more compute to deliver the most reliable answers on the hardest reasoning and coding tasks.

OpenAI's latest and most advanced GPT model. Delivers state-of-the-art performance across reasoning, coding, and creative tasks with enhanced capabilities.

Dated snapshot of GPT-5.1 for reproducible results. Supports cached input tokens for cost savings on repeated context. Ideal for production deployments requiring model version pinning.

OpenAI's latest flagship model. Delivers exceptional performance across reasoning, coding, and creative tasks with a massive 1M token context window and 32K output tokens. Supports vision, function calling, and JSON mode.

Fast and efficient variant of GPT-5. Delivers strong performance across reasoning, coding, and creative tasks with a 1M token context window and 32K output tokens, at a fraction of the cost of GPT-5.

Ultra-fast and lightweight variant of GPT-5. Designed for high-throughput, low-latency applications with a 1M token context window and 32K output tokens at minimal cost.

OpenAI's maximum-capability GPT-5 tier. Uses more compute to think longer and deliver the most reliable answers on the hardest reasoning, coding, and analysis tasks.

OpenAI's most capable model for coding and instruction following. Features a 1M token context window, 32K output tokens, and major improvements in coding, complex prompts, and long-context tasks. 20% cheaper than GPT-4o on output.

A significant leap in small model performance. Matches or exceeds GPT-4o in intelligence while reducing latency by nearly half and cost by 83%. Ideal balance of speed, quality, and affordability.