An AI model is a trained program that turns an input into an output, and creators use four kinds: image, video, text and voice. On 13 October 2026, independent leaderboards from Artificial Analysis were led by GPT Image 2.5 Sunburst for images, Wan 3.0 for text to video, Claude Opus 5.5 on the Intelligence Index for text, and Eleven v4 Turbo for speech. Free tiers cover a first workflow, entry paid plans run from $4.99 to $20 a month, and Apache 2.0 open models such as Wan 2.2 and Z-Image run on your own hardware. This page reads the boards and the vendors' own pages; we did not test the models.
An AI model is a trained program that turns an input, such as a written prompt or a photo, into an output, such as text, an image, a video clip or speech. Creators use four kinds, and on 13 October 2026 the independent leaderboards we read were led by GPT Image 2.5 for images, Wan 3.0 for video, Claude Opus 5.5 for text and Eleven v4 Turbo for voice. Those leaders change often, so the useful skill is knowing the kinds, the price of a way in, and how to read a ranking.
We did not test these models: this is a comparison by published leaderboards, specs and prices. Everything was read on 13 October 2026, on the Artificial Analysis boards and on each vendor's own pages, linked where they are used. Midjourney, FLUX and SDXL figures match our Midjourney vs FLUX and FLUX vs SDXL pages, read the same day. If you searched for "AI model" meaning a virtual persona, start with the AI influencers library.
Quick answer
- A model is the trained file. The app (ChatGPT, Midjourney, Kling) is a product built around it.
- Four kinds matter to creators: image, video, text (LLMs) and voice. Multimodal models mix inputs.
- Hosted models are rented by subscription or per use. Open-weight models are downloaded and run on your hardware.
- No model is best at everything. Check an independent leaderboard for the task, then the price and the license.
- Start free: Claude, ChatGPT and Gemini have free tiers, Google Flow gives 50 video credits a day, ElevenLabs 10k credits a month.
- A persona is a fictional adult. Never copy a real face or voice without written permission.
What is an AI model?
A model is the result of training: a large file of numbers, called weights, that encodes patterns found in its training data. Give it an input and it computes an output. A generative AI model produces new content, which is what separates it from a model that only sorts or scores things. A multimodal AI model works across more than one kind of data: Anthropic says all current Claude models take text and image input, and Google describes Gemma 4 as handling text and image input with text output.
Two distinctions explain most of the confusion. First, model versus product: Midjourney is a service whose current model is V8.2, and Google Flow is a product that runs the Veo 3.1 and Gemini Omni Flash models. Second, hosted versus open-weight. A hosted model runs on the company's servers and you pay for access. An open-weight model is published as a file, usually on Hugging Face, for you to run yourself under its license.
The four kinds of AI model creators use
| Kind | What it makes | First on an Artificial Analysis board, 13 October 2026 | A way in |
|---|---|---|---|
| Image | Pictures from a prompt, or edits to a picture | GPT Image 2.5 Sunburst, on the text-to-image board (Elo 1199) | Midjourney from $10 a month, or open models on your own GPU |
| Video | Short clips from a prompt or a still image | Wan 3.0, on the text-to-video board with audio (Elo 1158) | Google Flow: 50 free credits a day |
| Text (LLM) | Writing, code, analysis and planning | Claude Opus 5.5 at its highest effort setting, on the Intelligence Index (58) | Free tiers of Claude, ChatGPT and Gemini |
| Voice | Speech from text, in a stock or designed voice | Eleven v4 Turbo, on the text-to-speech arena (Elo 1330) | ElevenLabs Free: 10k credits a month |
Boards: text to image, text to video, LLM leaderboard and text to speech. Prices: Midjourney plans, Google Flow credits, Gemini plans and ElevenLabs pricing.
Image models
On the Artificial Analysis text-to-image board, ranked by blind human votes, the first six were GPT Image 2.5 Sunburst (1199), GPT Image 2.5 Flare (1192), GPT Image 2 (1172), Nano Banana 2.1 from Google (1159), Grok Imagine Image 2.0 (1155) and MAI-Image-2.6 (1152). FLUX 3 Image was ninth at 1115. Midjourney V8.2 is not on that board, and Midjourney offers no API.
What you pay depends on how the model is sold. Midjourney is a subscription: $10, $30, $60 or $120 a month. FLUX is billed per image through an API, at $0.048 for a 1K image on FLUX 3 Image. The open models cost nothing per image and ask for a graphics card instead. Editing has its own models, covered in our Qwen Image Edit guide, and the open families are explained in the Stable Diffusion glossary entry and the Z-Image guide.
Video models
A text-to-video AI model makes a clip from a prompt, and an image-to-video model animates a still, which is how most persona content is made because the still fixes the character's look. Wan 3.0 led the text board with audio and MiniMax H3 led the image board when read. Clips are short: 8 to 30 seconds per generation, depending on the model. The specs and prices of eight tools are in our AI video generator comparison, and what each free plan allows is in the free AI video generator guide.
Text models (LLMs)
A large language model writes, summarizes, plans and codes. The three vendors most creators meet publish their current lineups and API prices, in US dollars per million tokens of input and output:
| Model | Maker | Input / output per 1M tokens | Context window |
|---|---|---|---|
| Claude Fable 5.1 | Anthropic | $10 / $50 | 1M tokens |
| Claude Opus 5.5 | Anthropic | $4 / $20 | 1M tokens |
| Claude Sonnet 5.5 | Anthropic | $2 / $10 | 1M tokens |
| GPT-6 Astra | OpenAI | $10 / $50 | 1.05M tokens |
| GPT-6.1 Sol | OpenAI | $2 / $10 | 1.05M tokens |
| GPT-6 Luna | OpenAI | $0.10 / $0.50 | 1.05M tokens |
| Gemini 3.8 Flash | $0.75 / $3.75 through 31 December 2026, with a free tier | 1,048,576 input tokens | |
| Gemini 3.1 Pro Preview | $2 / $12 for prompts up to 200k tokens | 1,048,576 input tokens |
Sources: Anthropic models overview, OpenAI models and pricing, Gemini models and pricing. Anthropic also lists Claude Haiku 5.5 from $0.10 and $0.50. On the Artificial Analysis Intelligence Index, a composite of benchmarks, the top scores when read were Claude Opus 5.5 at its highest effort setting (58), Claude Sonnet 5.5 and Meta's Muse Wonder Full (56), then Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon (53). A score is not your task: a cheaper model is often enough for captions and outlines. The plans are compared in Claude vs GPT and Claude vs Gemini.
Voice models
A voice model turns text into speech. ElevenLabs' model list describes Eleven v4 as its most emotionally rich speech model and Eleven v4 Turbo as its real-time one, each in 90+ languages. The two held first and second place on the Artificial Analysis text-to-speech arena (1330 and 1326), ahead of Qwen-Audio-3.1-TTS-Plus (1298), Sonic 3.6 (1286) and Gemini 3.8 Flash TTS (1274).
On ElevenLabs' pricing page the Free plan includes 10k credits a month. Starter, at $6 a month, adds 30k credits, a commercial license and instant voice cloning, and Creator, at $22, adds professional voice cloning. First-month discounts were showing on both. Clone only your own voice, or one you hold written consent for: the ElevenLabs use policy lists "Deception, Manipulation and Impersonation" among prohibited uses. Other platforms are in our voice AI platforms guide.
New AI models, and why rankings move
- 27 Aug 2026Gemini Omni 1.1 Flash
Google adds scene extension up to 40 seconds and first and last frame control.
- Sep 2026GPT Image 2.5 and Eleven v4
Release month as listed on the Artificial Analysis image and speech boards.
- 28 Sep 2026Kling 4.0 announced
Kling says it will launch in October, with Kling 4.0 Flash in limited early access.
- Oct 2026FLUX 3 Image and Nano Banana 2.1
Both listed with an October release on the text-to-image board.
Source: Google blog, Kling release note and Artificial Analysis leaderboards, read 13 October 2026
Sources for the dated items: Google's Omni 1.1 Flash post and Kling's release note. Two kinds of AI model ranking exist, and they answer different questions. An arena, like the Artificial Analysis image, video and speech boards, shows voters two outputs without names and builds an Elo score from their choices, so it measures preference. A benchmark index runs fixed tests, so it measures what the tests cover. Neither tells you how a model handles your character, your brand voice or your budget. Use a ranking to make a shortlist of two or three, then try each on your own prompt.
What a creator stack costs to start
Source: Monthly billing in US dollars before promotions, from each vendor's plans page and our compare pages, read 13 October 2026. Kling shows $6.99 for the first month.
You do not need all six. A first workflow runs on free tiers: Claude, ChatGPT and Gemini each have a $0 plan, Google Flow gives 50 video credits a day, and ElevenLabs gives 10k credits a month. Pay for one tool when its free limit blocks real work, such as Kling Standard for 1080p video, and add the next only when it does the same. Free plans often withhold commercial rights, so check that before client work.
Open-source and local AI models
"Open source" is used loosely. What you get is open weights plus a license, and the license decides what you may do. A local AI model is simply one of these running on your own computer, usually through ComfyUI for images and video. The hardware figures below are each publisher's own statement.
| Model | Kind | License | Publisher's hardware statement |
|---|---|---|---|
| FLUX.2 [klein] 4B | Image | Apache 2.0 | About 13GB of VRAM (Black Forest Labs) |
| Z-Image-Turbo | Image | Apache 2.0 | "Fits comfortably within 16G VRAM consumer devices" (Z-Image team) |
| Stable Diffusion XL 1.0 | Image | CreativeML Open RAIL++-M | 8GB of VRAM (Stability AI) |
| Wan 2.2 TI2V-5B | Video | Apache 2.0 | At least 24GB for the team's own script; ComfyUI says it should fit on 8GB with offloading |
| gpt-oss-120b | Text | Apache 2.0 | A single H100 GPU (OpenAI) |
| Gemma 4 | Text | Apache 2.0 | Not stated on the page read |
Sources: our compare pages and Z-Image guide for the image models, the Wan 2.2 repository, OpenAI's gpt-oss-120b page and Google's Gemma 4 model card. Not every downloadable model is free to use commercially. Several FLUX [dev] models carry a non-commercial license, and MiniMax H3's license excludes the United States, the EU, the UK and South Korea. Several widely used open releases come from Chinese labs: Wan and Z-Image from Alibaba, and H3 from MiniMax.
How to choose a model
- 1Name the output
A still, a clip, a script or a voiceover. That picks the kind of model.
- 2Shortlist from an independent board
Take two or three near the top for that task, not only the first.
- 3Read the license or plan terms
Commercial use, watermarks and territory limits differ between plans and between open licenses.
- 4Price your real volume
Count the clips, images or minutes you publish in a month, then compare plans by that number.
- 5Test on your own prompt
Run the same prompt through each on a free tier before you pay.
AI models: FAQ
What is an AI model?
An AI model is a trained program: a file of learned numbers, called weights, that turns an input into an output. A text model turns a prompt into writing or code, an image model turns a prompt into a picture, and a voice model turns text into speech. The app you type into, such as ChatGPT or Midjourney, is a product built around one or more models.
What types of AI models are there?
Creators meet four generative kinds: image models, video models, text models (large language models, or LLMs) and voice models. A multimodal model handles more than one kind of input or output, such as text and images in and text out. Models are also either hosted, where you pay a company to run them, or open-weight, where you download the file and run it yourself.
What is the best AI model right now?
It depends on the job, and the order changes monthly. On 13 October 2026, independent boards from Artificial Analysis were led by GPT Image 2.5 Sunburst for text to image, Wan 3.0 for text to video with audio, Claude Opus 5.5 on the Intelligence Index for text, and Eleven v4 Turbo for text to speech. We read the boards; we did not test the models.
Are open-source AI models free to use commercially?
Only when the license says so. Apache 2.0 models such as Wan 2.2, Z-Image, FLUX.2 [klein] 4B, gpt-oss-120b and Gemma 4 allow commercial use. Others are downloadable but restricted: several FLUX [dev] models carry a non-commercial license, and the MiniMax H3 license excludes the United States, the EU, the UK and South Korea. Read the license file before you publish paid work.
Can I run an AI model locally?
Yes, if your hardware fits the model. Publishers give their own figures: Black Forest Labs says FLUX.2 [klein] 4B needs about 13GB of VRAM, the Z-Image team says Turbo fits within 16G consumer devices, and OpenAI says gpt-oss-120b fits a single H100 GPU. ComfyUI is the usual program for image and video models on your own machine.
Which AI models does a creator need?
Usually one of each kind you publish in: a text model for scripts and captions, an image model for stills, a video model to animate them, and a voice model if your content speaks. Start on the free tiers, then pay for one tool at a time when a free limit blocks real work.
Can I use an AI model to copy a real person's face or voice?
Not without their written permission. An AI persona should be a fictional adult that you designed, and a cloned voice should be your own or one you have written consent to use. Model licenses and platform policies do not override that: ElevenLabs lists impersonation among its prohibited uses and runs a report form for unauthorized use of a voice.