# API Models

> API Models (apimodels.app) is an AI model aggregation API that lets developers call image, video, audio, and LLM models from many providers through one unified API and a single balance — usually more affordable than going direct, and usable from China and worldwide.

## Key Pages

- [Models](https://apimodels.app/models): Browse 157 image / video / audio / LLM models, each with pricing and a one-line API call.
- [Pricing](https://apimodels.app/pricing): Per-call and per-token pricing for every model.
- [Docs](https://apimodels.app/docs): API reference, code samples (cURL / Python / Node), and CLI integration guides (OpenAI Codex, Claude Code).
- [Voice library](https://apimodels.app/voices): Browse MiniMax, ElevenLabs and Fish Audio text-to-speech voices with copyable voice IDs: filter by language, gender and source, play previews (ElevenLabs and Fish Audio), and open a model page with the voice preselected. Covers every MiniMax system voice, the ElevenLabs official presets and full shared voice library, Fish Audio licensed voices and a large set of community voices; refreshed daily.
- [Unified API](https://apimodels.app/unified-api): One key and one endpoint for image, video, and LLM models.
- [Cost calculator](https://apimodels.app/tools): Estimate what a workload costs here versus calling each provider directly.
- [Per-request billing](https://apimodels.app/docs/records): every LLM response carries an x-apimodels-request-id header (non-streaming also x-apimodels-cost); GET /v1/records/{id} returns the exact USD amount deducted for that one call, for any LLM, image, video or audio request. Free.
- [MCP server](https://apimodels.app/docs/mcp): Install `apimodels-mcp` with one npx command to call image, video, chat and speech models from Claude Desktop, Cursor or Cherry Studio; includes a review_image tool so the assistant critiques and regenerates its own output. Also hosted at https://api.apimodels.app/mcp with browser sign-in (OAuth), no Node.js needed.
- [Cursor guide](https://apimodels.app/docs/cursor): Connect Cursor to the hosted apimodels MCP server (one click or one url line in ~/.cursor/mcp.json, then Authenticate) so Cursor's agent generates images, video and voice-over; screenshots of each step and the fix for the port 8787 sign-in timeout.

## Pricing Comparisons
- [Nano Banana 2.1 vs GPT Image 2.5](https://apimodels.app/blog/nano-banana-2-1-vs-gpt-image-2-5): five identical prompts side by side (bilingual poster, infographic, character consistency, object swap, logo merge), one run each, unretouched. Both spelled every string correctly; GPT Image 2.5 Sunburst rendered richer scenes and readable background signs, Nano Banana 2.1 kept edit inputs untouched. Per image on apimodels.app: Nano Banana 2.1 $0.022 / $0.036 / $0.06, Sunburst high $0.03 / $0.04 / $0.16, Flare medium $0.02 / $0.03 / $0.045. 8:1 and 1:4 only on Nano Banana 2.1; transparent PNG only on GPT Image 2.5.
- [Nano Banana 2.1 vs Nano Banana Pro](https://apimodels.app/compare/nano-banana-2-1-vs-pro): Google list price (2.1 $0.0336 / $0.0504 / $0.113, Pro $0.134 / $0.134 / $0.24), two same-prompt tests at 2K, a week of production timings, and which jobs still need Pro.

Official list price next to what API Models charges, verified against the vendor's own pricing page on the date shown on each page, with measured success rates from production traffic.

- [GPT-6 API pricing](https://apimodels.app/access/gpt-6-api-pricing): Astra, Sol and Luna. Astra $3.40 / $17 per 1M tokens (66% below the $10 / $50 list), Sol $1 / $5 (50% below), Luna $0.09 / $0.45 (10% below).
- [GPT-6 vs GPT-6.1 (Sol 6.1)](https://apimodels.app/access/gpt-6-vs-gpt-6-1): what changed in GPT-6.1 Sol (29 Sep 2026) — same $2 / $10 list, cached input $0.20 → $0.10, DeepSWE v1.1 75.2% vs 68.8%, knowledge cutoff 30 April 2026, no reasoning_effort none; both $1 / $5 on APIMODELS, gpt-6 alias stays on GPT-6 Sol.
- [DeepSeek V4.1 Flash API pricing](https://apimodels.app/access/deepseek-api-pricing): official $0.30 / $1.20 per 1M tokens at peak (weekday Beijing working hours) and $0.15 / $0.60 off-peak; API Models $0.27 / $1.08 and $0.135 / $0.54, 10% below at every hour; Together, Fireworks and most OpenRouter providers charge $0.30 / $1.20 all day. 99.6% success over 20,011 calls.
- [Seedance API pricing compared](https://apimodels.app/access/seedance-api-pricing): per-second prices for Seedance 2.5, 2.0, 2.0 Fast and 2.0 Mini at 480p/720p/1080p on API Models next to fal, Kie.ai and EvoLink: Seedance 2.5 from $0.120/s, Seedance 2.0 from $0.092/s, and Dreamina Seedance 2.0 Mini from $0.0225/s at 480p, 44% below the BytePlus list price for Mini.
- [Grok Imagine Video API pricing](https://apimodels.app/access/grok-imagine-video-api-pricing): Grok Imagine Video 1.5 at $0.045 / $0.0529 / $0.0882 per second (480p / 720p / 1080p) on API Models against $0.08 / $0.14 / $0.25 at xAI, fal and OpenRouter.
- [MiniMax H3 (Hailuo 3) API pricing](https://apimodels.app/access/minimax-h3-api-pricing): 480P and 768P per-second prices from the official API, fal, Kie.ai, EvoLink and Replicate; API Models H3 Lite runs H3 at $0.01/s (480P) and $0.02/s (768P), the lowest in the table, as a draft tier that breaks on fast live-action. Native 2K and reference video are on minimax-h3.
- [GPT-5.6 API pricing](https://apimodels.app/access/gpt-5-6-api-pricing): Sol, Terra and Luna against OpenAI list, including the 272K long-context step.
- [AI car driving video prompt](https://apimodels.app/access/ai-car-driving-video-prompt): the viral September 2026 trend — replace the driver of a luxury-car clip with your photo, car and camera unchanged; four copy-ready prompts, clip and photo tips, 9 s 720P ≈ $1.62 with Wan 3.0.
- [Character reference sheet prompt](https://apimodels.app/access/character-reference-sheet-prompt): a copy-ready template for one landscape character sheet — large front portrait on the left half, full-body side and back views on the right, neutral grey background, no text — with photoreal, anime and 3D animated style presets and API code for gpt-image-2.5-sunburst ($0.04 per 2K sheet, about a minute).
- [Replace a person in a video (Wan 3.0)](https://apimodels.app/access/wan-3-0-replace-person-in-video): swap the driver, actor or presenter from a reference photo while the scene, camera and timing stay; prompt template, multi-angle references, duration -1, $0.045-$0.18/s standard billed on input + output seconds (9 s 720P ≈ $1.62).
- [Claude Opus 5.5 vs Sonnet 5.5](https://apimodels.app/access/claude-opus-5-5-vs-sonnet-5-5): Anthropic launch numbers side by side (Terminal-Bench 4.0 66.4% vs 70.6%, GDPval-AA 1846 vs 1844, Chartography without tools 64.4% vs 61.6%), list $4/$20 vs $2/$10, on apimodels.app $2.40/$12 vs $1.20/$6, Sonnet 5 vs 5.5, which to use.
- [Claude Opus 5.5 vs Fable 5.1](https://apimodels.app/access/claude-opus-5-5-vs-fable-5-1): Opus 5.5 ahead on every shared launch benchmark at less than half the list price; default effort medium vs high; when Fable is still worth it.
- [Claude Opus 5.5 vs GPT-6](https://apimodels.app/access/claude-opus-5-5-vs-gpt-6): Opus 5.5 against GPT-6 Sol and Astra by tier, list and apimodels prices, how to compare on your own prompts.
- [Claude Sonnet 5.5 pricing](https://apimodels.app/access/claude-sonnet-5-5-pricing): $2 / $10 per 1M tokens list, cache read $0.20 / write $2.50 (same as Sonnet 5); $1.20 / $6 on apimodels.app (40% below), cost per task, Haiku 5.5 status (released 7 Oct 2026; $0.09 / $0.45 on apimodels.app).
- [Making videos with Claude Opus 5.5](https://apimodels.app/access/claude-opus-5-5-video): Opus 5.5 writes a frame-rendering program, not a video file; our tested run: one call, $0.368, 12 s 1280x720 promo.
- [Is Claude Opus 5.5 free?](https://apimodels.app/access/claude-opus-5-5-free): not on the free claude.ai plan; Pro and above, finite API credits, or $2.40 / $12 pay-as-you-go.
- [Claude Code with Opus 5.5](https://apimodels.app/access/claude-code-opus-5-5): three environment variables, $2.40/$12 per million, cache prices, effort settings, whether Opus 5.5 is free.
- [Claude API pricing](https://apimodels.app/access/claude-api-pricing): Fable 5.1, Opus 5.5, Sonnet 5.5, Opus 5, Sonnet 5, Haiku 4.5 and Haiku 5.5 against Anthropic list, including cache-read prices.
- [Nano Banana API pricing](https://apimodels.app/access/nano-banana-api-pricing): Gemini 3 Pro Image and Gemini 3.1 Flash Image per image against Google list, and which model id to pick.
- [GPT Image 2.5 API pricing](https://apimodels.app/access/gpt-image-2-5-api-pricing): per-image price by resolution and quality against OpenAI token billing.
- [GPT Image 2 API pricing](https://apimodels.app/access/gpt-image-2-api-pricing): per-image price at 1K, 2K and 4K.

## Platform Alternatives

Head-to-head pages against other multi-model API platforms. Every competitor price was read from that platform's own pages on 2026-10-06, and each page says when the other platform is the better choice.

- [fal.ai alternative](https://apimodels.app/access/fal-ai-alternative): GPT Image 2 1K medium $0.025 vs $0.053, Nano Banana Pro $0.08 vs $0.15, Seedance 2.0 720p $0.197/s vs $0.303/s, Veo 3.1 8 s with audio $1.00 vs $3.20, Grok Imagine Video 1.5 lower here; fal is the choice for serverless GPUs, LoRA training, 4K Veo and its 1,000+ model long tail.
- [OpenRouter alternative](https://apimodels.app/access/openrouter-alternative): Claude Sonnet 5.5 $1.20 / $6 vs $2 / $10 per 1M tokens, GPT-6.1 Sol $1 / $5 vs $2 / $10, Gemini 3.1 Pro Preview, Nano Banana Pro, Veo 3.1 and Grok Imagine video lower here, no 5.5% top-up fee; OpenRouter is the choice for provider routing, zero data retention, BYOK, free models and enterprise SSO / SLAs.
- [Replicate alternative](https://apimodels.app/access/replicate-alternative): Nano Banana Pro $0.08 vs $0.15, Veo 3.1 8 s $1.00 vs $3.20, Kling 3.0 pro with audio $0.24/s vs $0.336/s, GPT Image 2 at 1K / 2K lower here, results kept 30 days vs 1 hour; Replicate is the choice for Cog, private models, Deployments, fine-tuning and its 50,000+ community models.
- [Kie.ai alternative](https://apimodels.app/access/kie-ai-alternative): Veo 3.1 Fast 1080p $0.088 vs $0.325 per clip, Veo 3.1 $1.00 vs $1.275, GPT Image 2 $0.025 / $0.03 / $0.05 vs $0.03 / $0.05 / $0.08, Nano Banana Pro $0.08 vs $0.09 lower here; Kie.ai has the fuller Suno V6 and Runway toolset and signed webhooks.

## Key Facts

- Category: AI model aggregator / unified AI API
- Audience: developers, indie hackers, and startups building AI features
- Main use cases: image generation & editing, video generation, text-to-speech / audio, LLM chat & agents (OpenAI / Claude / Gemini / Grok / GLM compatible)
- Models: 157 across image (34), video (46), LLM (51), audio (24) and embeddings (2)
- Billing: a single USD balance, pay-per-use; prices typically below official; $0.10 of trial credit for new accounts that sign up with Google using a personal mailbox or with Alipay (once per person; one new account per IP address in 48 hours); failed calls are never charged
- Access: one API key for all models; works from China and globally
- Endpoints: LLM POST /api/v1/messages (+ /chat/completions, /responses); Image POST /api/v1/images/generations; Video POST /api/v1/video/generations; Audio POST /api/v1/tts/stream; Realtime voice (speech-to-speech) WebSocket wss://api.apimodels.app/v1/realtime?model=grok-voice-3.0 (api subdomain only); Embeddings POST /api/v1/embeddings; Social media download (synchronous) POST /api/v1/media/extract, /playlist, /subtitles, /youtube-download
- Model list: GET https://api.apimodels.app/v1/models (no key needed) — each entry carries `modality` (chat / image / video / audio / embedding) and the `endpoint` to call it on
- Voice IDs: GET https://api.apimodels.app/v1/voices/catalog (public, no key, CORS enabled) lists TTS voices with id, name, languages, gender, accent, description, preview_url; filter with provider (minimax / elevenlabs / fish), source, language (ISO 639-1, Cantonese = yue), gender, q, sort (popular / name), page and page_size (up to 100). GET https://api.apimodels.app/v1/voices/catalog/manifest returns counts per provider, source and language. Pass an item's id as voice_id (MiniMax, ElevenLabs) or reference_id (fish-s2.1-pro, fish-drama-3).
- Result files are kept 30 days, then deleted — download or re-host anything you need to keep
- Last updated: 2026-10-30

## Newest Models

- [Gemini 4 Argon](https://apimodels.app/models/gemini-4-argon): Google's flagship reasoning model (announced 30 Sep 2026, on the Gemini API from 13 Oct), added 2026-10-14. $1.10 input / $5.50 output / $0.055 cached input per 1M tokens through 30 Nov 2026 — 45% below Google's $2 / $10 introductory list. From 00:00 UTC on 1 Dec 2026, when Google raises its list to $4 / $20, APIMODELS moves to $2.20 / $11.00 / $0.11 cached automatically, still 45% below. Benchmarks per Google: DeepSWE v1.1 77.9%, LVBench 91.7%, CWE-bench v1 68% (tied first), AutomationBench 51.3% (first). 1M context shared by input and output, up to 65,536 output tokens per response, text / image / video / audio / PDF input, tool calling; thinking tokens bill as output; implicit caching only. Native Gemini /v1beta/models/gemini-4-argon:generateContent and OpenAI-compatible /v1/chat/completions.
- [Claude Sonnet 5.5](https://apimodels.app/models/claude-sonnet-5-5): second of Anthropic's Claude 5.5 family (28 Sep 2026), added 2026-10-06. $1.20 input / $6.00 output per 1M tokens, cache reads $0.12, cache writes $1.50 — 40% below the $2 / $10 list. Per Anthropic: 30%+ faster output than Sonnet 5, up to 30% less per task, Terminal-Bench 4.0 70.6% (Sonnet 5: 10.3%). 1M context, 128K output, adaptive thinking (default effort high), knowledge cutoff June 2026. /v1/messages and /v1/chat/completions.
- [Claude Haiku 5.5](https://apimodels.app/models/claude-haiku-5-5): Haiku tier of Anthropic's Claude 5.5 family (7 Oct 2026), added 2026-10-09. $0.09 input / $0.45 output per 1M tokens, cache reads $0.009, cache writes $0.1125 — 10% below the $0.10 / $0.50 list; prompts over 100K tokens bill every rate at 5×. Benchmarks per Anthropic (max effort): GDPval-AA 1620 (Haiku 4.5: 735), OSWorld 2.1 72.4%, Terminal-Bench 4.0 39.2%; vs GPT-6 Luna 1620 vs 1437, 72.4% vs 48.9%, 39.2% vs 16.4% (Luna scores from Anthropic and third parties, not OpenAI). 1M context, 128K output, adaptive thinking (default effort medium). /v1/messages and /v1/chat/completions.
- [Claude Opus 5.5](https://apimodels.app/models/claude-opus-5-5): first of Anthropic's Claude 5.5 family (22 Sep 2026). $2.40 input / $12.00 output per 1M tokens, cache reads $0.12 — 40% below the $4 / $20 list. 1M context, 128K output, adaptive thinking always on, knowledge cutoff June 2026. Anthropic: Fable 5.1-level results on most work at 40% lower running cost than Opus 5.
- [GPT-6.1 Sol](https://apimodels.app/models/gpt-6.1-sol): OpenAI's upgraded Sol (29 Sep 2026), ahead of GPT-6 Sol on DeepSWE v1.1 (75.2% vs 68.8%) at the same list price. $1.00 input / $5.00 output / $0.05 cached per 1M tokens — half the $2 / $10 / $0.10 list. 1.05M context, 128K output, reasoning_effort low–max (no none); >272K input bills input 2x and output 1.5x. The bare alias gpt-6.1 routes here.
- [GPT-6.1 Terra](https://apimodels.app/models/gpt-6.1-terra): OpenAI's GPT-6.1 Terra (Oct 2026), listed at half the price of GPT-6.1 Sol; OpenAI reports 50.3% on Terminal-Bench 4.0 at max effort vs 21.5% for GPT-5.6 Terra. $0.50 input / $2.50 output / $0.025 cached per 1M tokens — half the $1 / $5 / $0.05 list. 1.05M context, 128K output, reasoning_effort none–max; >272K input bills input 2x and output 1.5x. Call it as gpt-6.1-terra (the bare gpt-6.1 alias routes to Sol).
- [GPT-6 Sol](https://apimodels.app/models/gpt-6-sol): OpenAI's GPT-6 workhorse tier (Sep 2026). $1.00 input / $5.00 output / $0.10 cached per 1M tokens — half the $2 / $10 list and below GPT-5.6 Sol. 1.05M context, 128K output, reasoning_effort none–max; >272K input bills input 2x and output 1.5x. The bare alias gpt-6 routes here.
- [GPT-6 Luna](https://apimodels.app/models/gpt-6-luna): OpenAI's most efficient GPT-6 model for focused, high-volume tasks (Sep 2026). $0.09 input / $0.45 output / $0.009 cached per 1M tokens — 90% of the $0.10 / $0.50 list. 1.05M context, 128K output, reasoning_effort none–max (reasoning tokens bill as output and count against max_tokens); >272K input bills input 2x and output 1.5x.
- [Jev](https://apimodels.app/models/jev): TypeSafe's System One decision model, not a chat model. POST /v1/systemone with a state and typed questions (choice / score / noul); answers come back with probabilities and confidence. $0.05 per 1M input tokens, output free, per-token billing, failed calls free. Same request and response as TypeSafe, so the official SDK works with only base_url and key changed. Docs: https://apimodels.app/docs/jev
September 2026 flagships. Prices are USD per 1M tokens; reasoning tokens count as output.

- [GPT-6 Astra](https://apimodels.app/models/gpt-6-astra): OpenAI's September 2026 flagship, added 2026-09-05. $3.40 input / $17.00 output / $0.34 cached input — 66% below OpenAI's own $10 / $50 list. Through this API: function calling, built-in web search and code interpreter (the computer-use tool is not available). Strongest on long-horizon agent and terminal work: 57.7% on Terminal-Bench 4.0 and 88.0% on SRE-Bench, against GPT-5.6 Sol's 37.3% and 55.9%. 1.05M context, 128K output, text + vision in. Requests above 272,000 input tokens are billed at 2x input / 1.5x output.
- [Claude Fable 5.1](https://apimodels.app/models/claude-fable-5-1): Anthropic's September 2026 flagship, added 2026-09-03. $5.00 input / $25.00 output / $0.22 cached read — exactly half Anthropic's $10 / $50 list. More than doubles Fable 5 on agentic science (52.6% vs 24.7% on Terminal-Bench-Science). 1M context, served on the native Anthropic Messages API, so Claude Code, the Anthropic SDK and Cursor work unchanged — only the base URL and key change.
- [Gemini 3.8 Flash](https://apimodels.app/models/gemini-3.8-flash): Google's September 2026 Flash, added 2026-09-03. $0.450 input / $2.250 output / $0.172 cached input — 40% below Google's own $0.75 / $3.75 list. Tops DeepSWE v1.1 and scores 89.4% on Terminal-Bench 2.1; beats Claude Opus 5 on HLE-Verified, Vals Finance Agent v2 and Harvey's Legal Agent. 1M context, built for coding and agents.
- [GLM-5.3](https://apimodels.app/models/glm-5.3): Z.ai's newest reasoning model, added 2026-08-26. $1.33 input / $4.18 output per 1M tokens. 1M-token context, strong function calling, built for long-horizon agent work and complex software engineering. OpenAI-compatible chat-completions endpoint.

The GPT-5.6 family remains the volume tier, callable on both /api/v1/chat/completions and /api/v1/responses. Reasoning depth is a model-name suffix (-low / -medium / -high / -xhigh / -max / -ultra), all at the same price.

- [GPT-5.6 Sol](https://apimodels.app/models/gpt-5-6-sol): frontier tier. $1.324 input / $6.618 output / $0.132 cached input. The `gpt-5.6` alias routes here. About 67% below OpenAI's own $4 / $20 list (OpenAI cut GPT-5.6 list prices on 2026-08-21).
- [GPT-5.6 Terra](https://apimodels.app/models/gpt-5-6-terra): balanced mini tier, exactly half of Sol. $0.551 input / $3.309 output / $0.055 cached input — 72% below OpenAI's own $2 / $12 list.
- [GPT-5.6 Luna](https://apimodels.app/models/gpt-5-6-luna): most affordable nano tier for high-volume work. $0.18 input / $0.96 output / $0.018 cached input — below OpenAI's own $0.20 / $1.20 list (10% on input, 20% on output). Call it as `gpt-5.6-luna-max`.
- [Grok 4.6](https://apimodels.app/models/grok-4.6): xAI's frontier LLM, built for long-running agents. $1.50 input / $4.50 output / $0.375 cached input — 25% below xAI's own $2 / $6. 500K context, text and image input, four reasoning-effort levels (low / medium / high / xhigh), default is high. Two things worth knowing: billed output tokens include reasoning tokens, and the default effort measures ~53s, so set client timeouts above 60s. Live web search works on /api/v1/responses, not on /chat/completions.
- [Qwen3.8 Max](https://apimodels.app/models/qwen3.8-max): Alibaba's newest flagship. $2.00 input / $6.00 output / $0.25 cached input per 1M tokens — the same as Alibaba's own list price, so the reason to call it here is the single key and balance, not a discount. OpenAI-format /api/v1/chat/completions, function calling, JSON output and streaming. Reachable directly from mainland China with no Alibaba Cloud account.
- [Qwen3.8 Flash](https://apimodels.app/models/qwen3.8-flash): the high-volume tier of the same generation. $0.15 input / $0.47 output / $0.016 cached input per 1M tokens, also at Alibaba's list price.
- [Qwen3.7 Max](https://apimodels.app/models/qwen3.7-max): previous flagship, 1M-token context and hybrid thinking (returns reasoning_content). $2.25 input / $6.75 output / $0.45 cached input — 10% below Alibaba's $2.5 / $7.5. Thinking is on by default and its tokens are billed inside completion_tokens at the output rate; send enable_thinking:false for simple work.
- [Qwen3.7 Plus](https://apimodels.app/models/qwen3.7-plus): the value tier of that generation — reasoning close to Max at roughly a sixth of the price. $0.36 input / $1.44 output / $0.072 cached input, 10% below Alibaba's list. Above 256K input tokens the whole request bills at triple rate ($1.08 / $4.32), a provider tier we pass through without a markup. Strongest on Chinese and long documents.

## Social Media Download

- [Media Downloader API](https://apimodels.app/models/media-downloader): turns a public post link from YouTube, TikTok, Instagram, Facebook, X, Xiaohongshu, Bilibili, Weibo, Douyin or any of 1000+ sites into structured JSON with direct links to its videos (every available quality), images and audio, added 2026-10-18, model id `media-downloader`. Four synchronous endpoints on https://api.apimodels.app with the same API key: POST /v1/media/extract (one post, $0.004 per request; short links such as youtu.be and vt.tiktok.com accepted), POST /v1/media/playlist (a channel, profile, playlist or hashtag, one page per call with next_cursor paging, $0.02 per page), POST /v1/media/subtitles (YouTube srt / vtt / json3 and Bilibili json subtitle tracks, $0.004 per request) and POST /v1/media/youtube-download (YouTube links up to 1080p through a temporary proxy, valid up to 1 hour and resumable with HTTP Range, $0.012 per request). Results return in the same HTTP response; download links are short-lived, and when a quality has separate video and audio tracks you download both and merge them with ffmpeg. Failed requests are not charged. Only download content you own or have permission to use, and respect each platform's terms and copyright. Docs: https://apimodels.app/docs/media-downloader

## Newest Video Models

- [Muse Video](https://apimodels.app/models/muse-video): Meta's Muse Video model, added 2026-10-30, model id `muse-video` on /api/v1/video/generations. Text-to-video from a `prompt` (up to 4,000 characters), image-to-video from one first frame (`image_url`; `first_frame_url`, `image` and `input_reference` also work), or up to 4 reference images (`images`; `image_urls` and `reference_image_urls` also work) that keep the same person or subject without starting from them — one or the other, not both; public URL or base64 data URI, PNG / JPEG / WebP. In our test, four pictures of one woman gave a 5 s beach clip that kept her face and hair. 720p only: `aspect_ratio` 16:9 (default, 1280×720) or 9:16 (720×1280); `size` 1280x720 / 720x1280 also works; square, 1080p and other sizes return 400 before a task is created. `duration` is whole seconds from 4 to 30 (`seconds` also accepted, default 5); 5, 6, 8, 10, 15, 20, 25 and 30 s come out exactly that long, other values round up to the next of those (4 → 5 s, 7 → 8 s), so a clip is never shorter than what was paid for. H.264 MP4 at 24 fps; Meta describes sound generated together with the picture. Async create-then-poll or `callback_url`; clips up to 10 s usually take 1-2 minutes, 15-30 s 3-7 minutes. $0.037 per requested second (5 s $0.185, 10 s $0.37, 30 s $1.11). Not every request succeeds (refused content, temporary errors, clips discarded for the wrong orientation or length); failed tasks are not charged. Not available (400): a first frame together with reference images, more than 4 reference images, a last frame, reference video or audio input, 1080p or any ratio other than 16:9 / 9:16, negative_prompt, seed, turning sound off; sound is not guaranteed on every clip. Docs: https://apimodels.app/docs/muse-video
- [OmniHuman 2.0](https://apimodels.app/models/omnihuman-2.0): ByteDance's audio-driven talking-avatar model, added 2026-10-29, model id `omnihuman-2.0` (aliases `omnihuman-2-0`, `omnihuman-2`, `omnihuman2`) on /api/v1/video/generations. One portrait (`image_url`; `image` or a one-item `image_urls` also work; JPEG / PNG / WebP / BMP / TIFF / GIF / HEIC / HEIF, under 30 MB, each side 300-6000 px, aspect ratio 0.4-2.5) plus one driving audio (`audio_url`, WAV or MP3, 4-300 s, under 15 MB) becomes an MP4 in which the person speaks or sings it, with lip sync, facial expressions and gestures; optional `prompt` up to 300 characters (Chinese, English, Japanese, Korean, Mexican Spanish, Indonesian) guides motion and camera; `enable_watermark` defaults to false. The video is as long as the audio and follows the portrait's aspect ratio at roughly 720p; one test on 2026-10-29: 5.36 s of English speech + a 1792×2390 portrait → 834×1112, 24 fps, 6.0 s, no watermark, about two minutes. ByteDance describes singing, multi-person scenes and prompt-guided motion. $0.091 per second of driving audio, rounded up to the whole second and measured at task creation (5.36 s → 6 s → $0.546; minimum 4 s $0.364, maximum 300 s $27.30); unreadable or out-of-range audio returns 400 before any charge. Docs: https://apimodels.app/docs/omnihuman-2-0
- [Kling 4.0](https://apimodels.app/models/kling-4.0): Kuaishou's flagship Kling video model, added 2026-10-22, model id `kling-4.0` (aliases `kling-4-0`, `kling-v4`) on /api/v1/video/generations, with sound generated on every clip (there is no silent mode). Two routes picked from the inputs: reference-to-video, which is a prompt alone or a prompt with up to 10 reference images (`image_urls`, mentioned as `@image_1`), 5 reference videos (`video_urls`, `@video_1`, 2-30 s each, 30 s in total), 7 named subjects (`elements`, `@Name`) and 3 named voices (`voices`), 15 materials at most, with `keep_original_sound` for the reference videos' own audio; and keyframe-to-video from `first_frame_url`, up to 8 `keyframe_urls` and `last_frame_url` (`@first_frame`, `@keyframe_1`, `@last_frame`; total frames at most min(10, duration + 1); the clip follows the first frame). Frames and references cannot be mixed. 3-30 s, 720p or 1080p, aspect ratio auto / 16:9 / 9:16 / 1:1 / 21:9, prompt up to 10,000 characters. $0.157/s (720p) and $0.235/s (1080p) without video input; with a reference video or video subject $0.094 / $0.141 per billed second, where billed seconds = output + max(input video, output): 10 s at 720p with a 5 s reference video is 20 billed s, $1.88. Not available: 4K or HDR, negative prompts, multi-shot storyboards, video extension. Docs: https://apimodels.app/docs/kling-4-0
- [Vidu Q4](https://apimodels.app/models/vidu-q4): the newest Vidu video model from ShengShu (Preview, upstream viduq4-preview), added 2026-10-22, with dialogue, sound effects and ambience generated together with the picture. Two modes on one model id: image-to-video from `first_frame_url` (the clip follows the image's aspect ratio, `is_rec` toggles prompt enhancement), and reference-to-video from `reference_image_urls` (1-15 images) plus optional `reference_audio_urls` (0-3 MP3 clips of 3-12 s) with a required prompt and `aspect_ratio` 16:9 / 9:16 / 1:1 / 4:3 / 3:4. 3-16 seconds, 540p / 720p / 1080p / 2K / 4K, `generate_audio` on by default, no watermark. No text-only mode, no last frame, no reference video. Measured via the API on 2026-10-22 at 540p · 3 s: image-to-video 102 s, reference-to-video 173 s. $0.042/s (540p), $0.0887/s (720p), $0.112/s (1080p), $0.1773/s (2K), $0.3639/s (4K), audio on or off the same price. Model id `vidu-q4` on /api/v1/video/generations. Docs: https://apimodels.app/docs/vidu-q4
- [FLUX 3 Max](https://apimodels.app/models/flux-3-max): a post-trained, latency-optimized variant of FLUX 3 Video from the FLUX family by Black Forest Labs, added 2026-10-20, model id `flux-3-max` on /api/v1/video/generations. Three modes behind one id, chosen by the input: text-to-video (prompt only), image-to-video (`first_frame_url`, optional `last_frame_url`) and keyframes-to-video (`keyframes`: 1–10 of {image_url, frame_index} at 24 fps, or {image_url, time} in seconds). 5–20 s, 720p or 1080p, aspect ratio auto / 21:9 / 2:1 / 16:9 / 4:3 / 1:1 / 3:4 / 9:16, generated audio on by default (`generate_audio`), `reasoning_level` low or high. In our tests through the API (20 Oct 2026, 720p, 5 s): text-to-video 9 s, image-to-video 15 s, keyframes 19 s to a finished file. $0.221/s at 720p and $0.377/s at 1080p, billed per second, same price for all modes; special price $0.085/s and $0.145/s until 7 Nov 2026. Guide with pricing, parameters, code and comparison with Seedance 2.5: https://apimodels.app/flux-3-max-api
Video is billed per output second unless stated otherwise. Failed generations are never charged.

- [Wan 2.7 Spicy](https://apimodels.app/models/wan-2.7-i2v-spicy): image-to-video with a generated stereo audio track, added 2026-09-29. First frame + prompt → 2–15 s at 720p or 1080p; pass `audio_url` (https WAV/MP3, 2–30 s) to drive the clip with your own sound, `audio: false` for a silent file. Relaxed moderation, anime-style and Asian-face scenes, free prompt rewriting; no last frame. $0.169/s at 720p, $0.26/s at 1080p, billed per requested second; a 1080p 2-second test clip returned in 45 s. Model id `wan-2.7-i2v-spicy` on /api/v1/video/generations. Docs: https://apimodels.app/docs/wan-2-7-i2v-spicy
- [Wan 2.2 Animate](https://apimodels.app/models/wan-2.2-animate): character image + reference video, added 2026-09-29. `mode: "animate"` (default) drives your character with the motion and expressions in the video; `mode: "replace"` keeps the video's scene and motion and swaps the person for your character. Output length equals the reference video (up to 120 s), 480p or 720p, audio kept. Billed per second of the reference video rounded up: $0.0742/s at 480p, $0.154/s at 720p, fixed when the task is created. Model id `wan-2.2-animate` on /api/v1/video/generations. Docs: https://apimodels.app/docs/wan-2-2-animate
- [Gemini Omni 1.1 Flash](https://apimodels.app/models/gemini-omni-1.1-flash): Google's next Omni video model, added 2026-09-04. Four modes behind one model id — text-to-video, first/last-frame image-to-video, reference-to-video (up to 7 reference images, reusable voices and characters) and video editing from a source clip. 720p $0.07/s, 1080p $0.10/s, 4K $0.20/s — at least 30% below the official rate; generations that take a video as input are a flat $0.70 (720p/1080p) or $1.05 (4K). 4 / 6 / 8 / 10 seconds, 16:9 or 9:16, synced audio, no visible watermark.
- [Seedance 2.5 Lite](https://apimodels.app/models/seedance-2.5-lite): the budget tier of ByteDance Seedance 2.5, added 2026-10-22, for text-to-video and video guided by up to 9 JPG or PNG reference images (references, not a first frame). 15 or 30 seconds at 720p only, six aspect ratios (16:9 default, 9:16, 1:1, 4:3, 3:4, 21:9), audio always generated in the same pass. A flat $2.50 per clip for either length, charged only on success; the same 30-second 720p clip on seedance-2.5 is about $8.10. No first/last frame, no reference video or audio, no silent output, no other resolution (each is a 400). Some jobs fail on this budget tier and are refunded. In our test on 2026-10-22 a 30-second clip took 401 s. Model id `seedance-2.5-lite` on /api/v1/video/generations; the old name `seedance-2.5-global` still works and now runs on Seedance 2.5 Lite (defaults to 30 seconds). Docs: https://apimodels.app/docs/seedance-2-5-lite
- [Wan 3.0 Video](https://apimodels.app/models/wan-3.0-video): Alibaba's all-in-one video model, added 2026-08-24. One model name covers text-to-video, first/last-frame, and omni-reference (up to 10 images + 5 videos + 5 audio clips, or a document / web page). Native 30-second single-shot output at 30fps with a synced audio track. Pass `mode:"prime"` for the high-speed tier — measured 58s versus 356s on the same job, for 50% more per second. Current per-second rates are on the model page.
- [MiniMax H3 RH Enhanced](https://apimodels.app/models/minimax-h3-rh-enhanced): an enhanced build of the open-source MiniMax H3 video model, added 2026-10-10. One model id, three modes picked from the inputs — text-to-video, first/last-frame image-to-video, and multi-reference (up to 9 images, 3 reference videos, 3 voice references) — with audio-driven lip sync and an audio_mode switch (lock_source keeps your track for music videos). 480P $0.05/s, 768P $0.07/s, 1080P $0.10/s, 4-15 seconds; references cost nothing extra.
- [MiniMax H3 Max Turbo](https://apimodels.app/models/minimax-h3-max-turbo): a post-trained, throughput-optimized build of MiniMax H3, added 2026-09-03 — clips in seconds instead of minutes, with stronger prompt adherence. Any whole length from 5 to 15 seconds, text-to-video in six aspect ratios, or image-to-video from a first frame with an optional last frame. No multi-image references, reference video or audio. Current per-second rates for 480P and 768P are on the model page.
- [Grok Imagine Video 1.5](https://apimodels.app/models/grok-imagine-video-1.5): xAI's video model, added 2026-08-22. Text-to-video and image-to-video, billed per output second from $0.045/s.
- [Digital Human](https://apimodels.app/models/digital-human): photo-plus-audio talking-avatar video at $0.01 per output second, added 2026-08-18.

## Newest Image Models

- [Muse Image](https://apimodels.app/models/muse-image): Meta's Muse Image 1.0 model for text-to-image and image editing, added 2026-10-30, model id `muse-image` (alias `muse-image-1.0`) on /api/v1/images/generations. One image of about 2.5 megapixels per request (`n` must be 1); `aspect_ratio` auto (default, the prompt decides; an unsized test came back 1920×1280) or one of 1:1, 3:2, 2:3, 4:3, 3:4, 5:4, 4:5, 16:9, 9:16, 2:1, 1:2, 21:9, 9:21, or an OpenAI-style `size` that sets the proportions (about 1600×1600 at 1:1, 2048×1152 at 16:9; exact pixels are chosen by the model; there is no 4K, and sizes above about 2.6 MP such as 3840x2160 or 2048x2048 return 400). Delivered as JPEG by default, `output_format` png or webp on request; about 35-70 s per image in our tests. Prompt up to 4,000 characters; Meta describes careful composition and text rendering. Image editing on the same request: add 1-4 reference images (`image` or `images`; also `image_url` / `image_urls`; public URL or base64 data URI, PNG / JPEG / WebP, up to 10 MB each and 24 MB in total), the first is the main image, the prompt can refer to the others; same price. Masks (inpainting), more than 4 reference images, a transparent background, negative_prompt and seed return 400 instead of being ignored. Not every request succeeds (refused content, temporary errors); failed requests are not charged. A flat $0.02 per image, charged only on success. Docs: https://apimodels.app/docs/muse-image
- [Nano Banana 2.1](https://apimodels.app/models/nano-banana-2-1): Google's high-efficiency image model for generation and editing, added 2026-10-07, model id `nano-banana-2-1` on /api/v1/images/generations. Priced per image by resolution: $0.022 at 1K (the default), $0.036 at 2K, $0.06 at 4K — a lower per-image price than Nano Banana 2 at every tier here ($0.06 at 1K/2K and $0.10 at 4K on gemini-3.1-flash-image). Text-to-image, or editing with up to 10 reference images (`image` for one, `image_urls` for several; reference images cost nothing extra). 15 aspect ratios including 1:4, 4:1, 1:8 and 8:1, prompts up to 20,000 characters, async create-then-poll or callback_url, charged only on success. First test (16:9, 1K): about 20 s, delivered 1376×768 with the sign text spelled correctly. Docs: https://apimodels.app/docs/nano-banana-2-1 The model page walks through nine use cases (character consistency, bilingual poster, infographic, 8:1 banner, 1:4 story, object swap, circle-to-edit, product shot with a logo merged in, three-round edit), each one real request with its verbatim prompt.
- [Doubao Seedream 5.0 Flash](https://apimodels.app/models/doubao-seedream-5-0-flash): ByteDance's fast, cost-efficient Seedream 5.0 tier from Volcano Ark's official endpoint. $0.03 per image at 1K or 2K (no 4K), about 10–15 seconds per image, text-to-image and editing with up to 10 references, edits that keep an existing PNG's transparency, and layer decomposition (layer splitting) into a base image plus up to 16 transparent PNG layers with names and positions (billed $0.03 per output image).
- [Grok Imagine Image 2.0](https://apimodels.app/models/grok-imagine-image-2): xAI's newest image model, model id `grok-imagine-image-2.0` on /api/v1/images/generations. Priced by resolution × quality: 1K low $0.03, 2K low $0.04, 1K medium $0.045, 2K medium $0.06 per image — all four 25% below xAI list, and 2K low 33% below. `quality` takes low or medium and defaults to medium. Fuses up to 3 reference images in one call (a 4th is refused explicitly, not dropped silently) and reference images cost nothing extra. 14 aspect ratios including ultra-wide 19.5:9, 9:19.5, 20:9 and 9:20. Strong at in-image typography — an exact headline, date line and small-print URL all came back character-correct on a 2K poster. Two verified limits: editing re-renders the frame rather than editing pixels in place, so crop and camera angle shift between passes; and it cannot produce transparent backgrounds — asked for alpha it paints the checkerboard as ordinary pixels and returns an RGB PNG.
- [FLUX.2 Klein 4B](https://apimodels.app/models/flux-2-klein-4b): fast (1-2s) photoreal image generation, added 2026-08-21. Text-to-image and image-to-image with up to 3 reference images, fixed 2MP output, 12 aspect ratios. From $0.006 per image, plus $0.0015 per reference image.
- [Z-Image Spicy](https://apimodels.app/models/z-image-spicy): text-to-image at any size from 256 to 1536 px per side (width/height, an OpenAI-style size or an aspect_ratio), added 2026-09-19. Seed for repeatable output, optional prompt rewriting (prompt_extend), about 15 s per image. Flat $0.015 per image, one image per request, no reference images. Model id `z-image-spicy` on /api/v1/images/generations.
- [Qwen Image Edit Spicy](https://apimodels.app/models/qwen-image-edit-spicy): relaxed-moderation image editing, model id `qwen-image-edit-spicy` on /api/v1/images/generations, added 2026-10-08. Send the image to edit plus a text instruction and up to 2 optional reference images; the result keeps the exact pixel size of the image you send. Typically 8–20 s per image. $0.05 with the main image only, $0.06 with 1 reference, $0.07 with 2, charged only on success.
- [Z-Image Spicy Pro](https://apimodels.app/models/z-image-spicy-pro): the higher-resolution tier of Z-Image Spicy, added 2026-09-20. Same request shape (prompt, seed, prompt_extend), sides from 256 to 2560 px (width/height, OpenAI-style size or aspect_ratio), 12–21 s per image in our tests, prompt rewriting off by default. Flat $0.03 per image, one image per request, no reference images. Model id `z-image-spicy-pro` on /api/v1/images/generations.
- [Qwen Image 3.0](https://apimodels.app/models/qwen3-image): Alibaba's text-to-image and image-editing model in one model id, on /api/v1/images/generations. A flat $0.035 per image at both 1K and 2K, plus $0.004 per reference image. Prompts up to about 4,500 tokens, dense in-image text layout, small text readable down to 10px, 12 languages and 20+ fonts.
- [Qwen Image 3.0 Pro](https://apimodels.app/models/qwen3-image-pro): the higher-fidelity tier of the same model — micro-expressions, pores and individual hair strands close to photographic. 1K $0.037, 2K $0.075, plus $0.004 per reference image. Same 4,500-token prompts, 10px small text, 12 languages and 20+ fonts; built for posters, storyboards and menus where the typesetting has to survive.
- [GPT-Image-2](https://apimodels.app/models/gpt-image-2): OpenAI's image model without an OpenAI organization verification step. Priced by native resolution: $0.025 at 1K, $0.03 at 2K, $0.05 at 4K per image. [GPT-Image-2 All](https://apimodels.app/models/gpt-image-2-all) puts text-to-image and editing behind one model id.
- [GPT-Image-2.5 Flare](https://apimodels.app/models/gpt-image-2.5-flare): OpenAI's GPT Image 2.5, default tier (medium quality by default) — sharper output, region edits that stay local, native transparent PNG. Priced per image by resolution × quality across fifteen cells, $0.01–$0.30: 1K $0.01–$0.08, 2K $0.02–$0.15, 4K $0.02–$0.30 (low → max); medium $0.02/$0.03/$0.045 and high $0.03/$0.04/$0.08 at 1K/2K/4K. [GPT-Image-2.5 Sunburst](https://apimodels.app/models/gpt-image-2.5-sunburst) is the highest-fidelity tier (high quality by default): the same 1K / 2K cells, its own 4K row at $0.02–$0.47.
- [Seedream 5.0 Pro Spicy](https://apimodels.app/models/seedream-5.0-pro-spicy): the relaxed-moderation variant of Seedream 5.0 Pro, added 2026-10-06, model id `seedream-5.0-pro-spicy` on /api/v1/images/generations. Same photorealistic quality and request shape as Seedream 5.0 Pro: text-to-image, or image-to-image with up to 10 reference images. 1K $0.05 / 2K $0.10 per image (default 2K), no 4K, JPEG output, typically 40–80 s, charged only on success. We add no filter of our own; the model's own safety systems still apply upstream.

## Newest Audio Models

- [ElevenLabs Sound Effects v3](https://apimodels.app/docs/audio): ElevenLabs' sound-effect model (SFX v3, released 29 Oct 2026), added 2026-10-30, model id `eleven-sfx-v3` on POST /api/v1/audio/generations. Send a prompt: several sounds described in sequence in one prompt are rendered in order, music and vocals work, and `loop: true` returns a seamless loop. 0.5–30 s per clip (`duration_seconds`, 5 s if omitted), $0.003 per second of the requested duration.
- [Grok Voice 3.0](https://apimodels.app/models/grok-voice-3.0): xAI's speech-to-speech model for realtime voice agents (released 14 Oct 2026), added 2026-10-16, model id `grok-voice-3.0`. xAI cites Artificial Analysis (13 Oct 2026): #1 on the Speech-to-Speech Index with 86.9 (Grok Voice 2.0: 81.3). WebSocket only, on the API subdomain: wss://api.apimodels.app/v1/realtime?model=grok-voice-3.0 with Authorization: Bearer <key>; browsers get a 10–600 s token from POST https://api.apimodels.app/v1/realtime/client_secrets and connect with the subprotocol xai-client-secret.<value>. Events follow xAI's Voice Agent API (largely OpenAI Realtime-compatible; input transcription is cumulative). 16-bit PCM 24 kHz by default, 28 built-in voices that each speak every language, 20+ languages including Chinese, function calling (no server-side tools yet). $0.104 per minute of connected session, billed per second (90 s = $0.156), plus $0.0052 per input_text message; 60 min per session, 5 concurrent sessions per account. Setup guide in Chinese and English: [Grok Voice 3.0 API tutorial](https://apimodels.app/grok-voice-3-api); docs: [/docs/grok-voice](https://apimodels.app/docs/grok-voice).
- [Fish Audio S2.1 Pro](https://apimodels.app/models/fish-s2.1-pro): text-to-speech across 83 languages with twenty free voice clones, added 2026-09-20. Send a ten-second sample and a voice id comes back in under twenty seconds; pass it as reference_id afterwards. Past twenty stored voices each extra slot is $1, and deleting one frees it again. Billed per 1,000 UTF-8 bytes rather than per character, so a Chinese character costs three times a Latin one. Two tiers run the same model on /api/v1/audio/generations: `fish-s2.1-pro` at $0.03 per 1,000 bytes, and `fish-s2.1-pro-free` at $0.005, which is best-effort (41s against 17s on the same passage), carries no SLA, and may have its requests used by the provider to improve their models. Output as mp3, wav, pcm or opus. Fish Audio's licensed voices (licensed: true in the voice library) cost twice the rate on `fish-s2.1-pro` ($0.06 per 1,000 bytes) and are not available on `fish-s2.1-pro-free`; community voices and your own clones are billed normally.
- [Fish Audio Drama 3](https://apimodels.app/models/fish-drama-3): text-to-speech for performed speech, added 2026-10-18. Write stage directions into the text — a bracketed note such as [barely holding it together] before a line, or paired tags like <whisper>…</whisper> around a few words — and voice up to 16 speakers in one request by passing reference_id as an array and switching with <|speaker:N|>. Phoneme tags pin the pronunciation of names and terms. 83 languages. Billed per 1,000 UTF-8 bytes like S2.1 Pro: `fish-drama-3` on /api/v1/audio/generations at $0.065 per 1,000 bytes, with a special price of $0.025 until 2026-11-12; Fish licensed voices cost 2.5 times that ($0.1625, or $0.0625 during the special). Docs: https://apimodels.app/docs/fish-drama-3

## In Chinese / 中文

apimodels.app 提供完整中文站,内容与英文站等价,URL 加 `/zh` 前缀:

- [模型目录](https://apimodels.app/zh/models):157 个图片 / 视频 / 音频 / 大语言模型,每个都带价格和一行调用示例。
- [价格](https://apimodels.app/zh/pricing):按次和按 token 的完整价目。
- [音色库](https://apimodels.app/zh/voices):MiniMax、ElevenLabs、Fish Audio 的 TTS 音色,一键复制音色 ID、在线试听,按语言、性别、来源筛选;同一份数据可用公开接口 GET https://api.apimodels.app/v1/voices/catalog 查询,不需要 Key。
- [文档](https://apimodels.app/zh/docs):API 参考、cURL / Python / Node 示例,以及 Claude Code、OpenAI Codex 等命令行工具的接入方式。
- 特点:一把 API Key 调用所有模型,统一美元余额,按量付费,国内可直连,调用失败不计费,注册赠送 $0.10 试用额度。生成结果文件保留 30 天。

## Contact

- Website: https://apimodels.app
- Brand: API Models