Updated November 11, 2026. Fifteen new AI models are on OpenRouter right now, from Google, Anthropic, OpenAI, Meta, NVIDIA, Qwen and DeepSeek. This guide puts their real prices, release dates and abilities side by side, in plain English, so you can pick without reading fifteen pages of specs.
The quick answer: which AI model should you pick?
- You want the lowest bill: DeepSeek V4.1 Flash 1017. About $0.07 for 100 typical jobs. Claude Haiku 5.5 and DeepSeek V4.1 Flash are close behind at about $0.09 each.
- You want the highest published score: Claude Fable 5.2 (score 64.6), but it is also the most expensive at about $6.93 for 100 jobs.
- You want a middle path with a score: Gemini 4 Carbon (score 60, about $1.73 for 100 jobs) or GPT-6.1 Terra (score 51.3, about $0.87).
- You need it to understand audio and video: Gemini 4 Carbon, Gemini 4 Argon or Meta Muse Wonder.
- You need a standard everyday assistant: GPT-6.1 Terra, Claude Sonnet 5.5 or Gemini 4 Carbon sit in the middle on price.
Scores come from Artificial Analysis and are shown on OpenRouter for only three of the fifteen models, so treat them as one clue, not the full story.
How to read the prices
AI models charge by the token, a small chunk of text of about three-quarters of a word. Prices are quoted per one million tokens (about 750,000 words), and models charge one price for the text you send in and a higher price for the text they write back.
To make that easy to feel, this guide uses a typical job: the model reads about 1,500 words and writes about 1,000 words. The column “Cost of 100 typical jobs” is simply what that would cost 100 times. It is an estimate from the headline prices, not a bill.
All 15 models, cheapest first
| Model | Released | Price per 1 million tokens | Cost of 100 typical jobs | Can take in |
|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 DeepSeek | Oct 17, 2026 | $0.006 in $0.50 out | $0.07 | text, images |
| DeepSeek V4.1 Flash DeepSeek | Sep 10, 2026 | $0.051 in $0.57 out | $0.09 | text, images |
| Claude Haiku 5.5 Anthropic | Oct 7, 2026 | $0.10 in $0.50 out | $0.09 | text, images, files |
| Nemotron 3.5 Super VL NVIDIA | Oct 15, 2026 | $0.15 in $0.47 out | $0.09 | text, images, video |
| GPT-6.1 Terra Pro OpenAI | Oct 26, 2026 | $1 in $5 out | $0.87 | text, images, files |
| GPT-6.1 Terra OpenAI | Oct 26, 2026 | $1 in $5 out | $0.87 | text, images, files |
| Gemini 4 Carbon | Nov 10, 2026 | $2 in $10 out | $1.73 | text, images, files, audio, video |
| Gemini 4 Argon | Oct 13, 2026 | $2 in $10 out | $1.73 | text, images, files, audio, video |
| Claude Sonnet 5.5 Anthropic | Sep 28, 2026 | $2 in $10 out | $1.73 | text, images, files |
| GPT-6.1 Sol Pro OpenAI | Sep 29, 2026 | $2 in $10 out | $1.73 | text, images, files |
| GPT-6.1 Sol OpenAI | Sep 29, 2026 | $2 in $10 out | $1.73 | text, images, files |
| Qwen3.8 Max Prime Alibaba Qwen | Sep 23, 2026 | $4 in $12 out | $2.40 | text, images, video |
| Claude Opus 5.5 Anthropic | Sep 22, 2026 | $4 in $20 out | $3.47 | text, images, files |
| Meta Muse Wonder Meta | Oct 12, 2026 | $4 in $20 out | $3.47 | text, images, files, audio, video |
| Claude Fable 5.2 Anthropic | Nov 9, 2026 | $8 in $40 out | $6.93 | text, images, files |
Prices are the headline rates OpenRouter lists on November 11, 2026. Some models are sold by several providers at different rates, so the price you pay can be higher or lower.
What each model is built for
These descriptions come from OpenRouter’s own model pages, shortened.
- Gemini 4 Carbon (Google): Coding, agentic work and complex reasoning. Takes text, images, audio, video and files.
- Gemini 4 Argon (Google): Google's flagship reasoning model: hard software engineering, deep math and science, and multi-step agents.
- Claude Fable 5.2 (Anthropic): Anthropic's most capable model, aimed at the most demanding technical work. Better than Fable 5.1 at coding, reasoning and vision.
- Claude Haiku 5.5 (Anthropic): A small, fast model for high-volume, cost-sensitive jobs such as summaries, helper agents and browser use.
- Claude Sonnet 5.5 (Anthropic): Everyday, well-defined work such as building features and fixing bugs. A direct upgrade on Sonnet 5.
- Claude Opus 5.5 (Anthropic): Anthropic's flagship for demanding reasoning, coding and long, multi-step agent work in large codebases.
- Nemotron 3.5 Super VL (NVIDIA): An open multimodal model (12B active parameters out of 120B). Reads text, images, documents, screenshots and video.
- Qwen3.8 Max Prime (Alibaba Qwen): A higher-throughput version of Qwen3.8 Max, sold as a separate option at a higher price. Takes text, images and video.
- DeepSeek V4.1 Flash (DeepSeek): A sparse mixture-of-experts model, the first from DeepSeek built on its new encoder-decoder design.
- DeepSeek V4.1 Flash 1017 (DeepSeek): The October 17, 2026 snapshot of DeepSeek V4.1 Flash, served under the name deepseek-flash.
- Meta Muse Wonder (Meta): Meta's frontier reasoning model for complex personal and professional agent tasks. Takes text, images, video, audio and PDFs.
- GPT-6.1 Terra Pro (OpenAI): The same model as Terra, run in a higher-quality reasoning mode. OpenRouter warns that it spends far more tokens, so real cost is higher.
- GPT-6.1 Terra (OpenAI): The mid-tier GPT-6.1 model, priced below Sol. Suited to everyday agent, coding and writing work, and good at reading screenshots.
- GPT-6.1 Sol Pro (OpenAI): The same model as Sol, run in a higher-quality reasoning mode. OpenRouter warns that it spends far more tokens, so real cost is higher.
- GPT-6.1 Sol (OpenAI): An upgrade to GPT-6 Sol, sitting below OpenAI's GPT-6 flagship. Suited to agentic coding, computer use and document-heavy professional work.
Four things most people miss
1. Context size will not decide it
Every model here takes about one million tokens at once, roughly 750,000 words. That is a lot, and it is the same across the board, so look at price and abilities instead.
2. “Pro” costs more than the price tag says
GPT-6.1 Terra Pro and Sol Pro list the same price as their normal versions. OpenRouter notes that Pro mode spends far more tokens because the model thinks longer, so plan for a bigger bill.
3. Only three models carry a score
Claude Fable 5.2 (64.6), Gemini 4 Carbon (60) and GPT-6.1 Terra (51.3) have an Artificial Analysis Intelligence Index on OpenRouter. The other twelve have no score listed, so this guide does not rank them on quality. Test them on your own work before you commit.
4. Not every model can watch video
Only Gemini 4 Carbon, Gemini 4 Argon and Meta Muse Wonder take audio and video. If you want that, your list is short.
Frequently asked questions
Which AI model is cheapest right now?
On OpenRouter's listed prices on November 11, 2026, DeepSeek V4.1 Flash 1017 is the cheapest of these 15. It costs $0.006 per million input tokens and $0.50 per million output tokens, which works out to about $0.07 for 100 typical jobs.
Which of these AI models has the highest published score?
Of the 15, only three show an Artificial Analysis Intelligence Index score on OpenRouter: Claude Fable 5.2 (64.6), Gemini 4 Carbon (60) and GPT-6.1 Terra (51.3). Fable 5.2 scores highest and also costs the most. The other 12 models have no score listed there, so they cannot be ranked this way.
What does 'per million tokens' mean?
A token is a small chunk of text, roughly three-quarters of a word. One million tokens is about 750,000 words. Models charge separately for the text you send in and the text they write back.
Which models can handle audio and video?
Gemini 4 Carbon, Gemini 4 Argon and Meta Muse Wonder accept text, images, files, audio and video. NVIDIA Nemotron 3.5 Super VL and Qwen3.8 Max Prime accept video but not audio. The rest take text and images, and most of them also take files.
Do the Pro versions cost the same as the normal ones?
The listed price is the same, but Pro mode makes the model think much longer, so it uses far more tokens. OpenRouter's own note says Pro mode spends far more than the normal version, so your real bill will be higher.
Is a bigger context window a reason to choose one model over another?
Not here. Every model in this list handles about one million tokens at once, which is roughly 750,000 words, so context size does not separate them.
Conclusion
There is no single best model among these 15. The right pick depends on what you need most, so decide that first and the choice gets short.
- Keep the bill small: start with DeepSeek V4.1 Flash 1017, about $0.07 for 100 typical jobs.
- Want the highest published score: Claude Fable 5.2 leads at 64.6, and it also costs the most, about $6.93 for 100 jobs.
- Want a balance of price and score: Gemini 4 Carbon (60, about $1.73) or GPT-6.1 Terra (51.3, about $0.87).
- Need audio or video: only Gemini 4 Carbon, Gemini 4 Argon and Meta Muse Wonder take both.
Two things to remember before you commit. Only three of the 15 models have a published score, so the other twelve are not ranked on quality here. And the Pro versions list the same price as the normal ones but use far more tokens, so they cost more in practice. Pick one or two models from the list that fit your need, run them on a few of your own real tasks, and keep the one that gives the best result for the money.
All prices and dates are from OpenRouter on November 11, 2026 and can change, so check the model page before you buy.
Where this data comes from
Every price, date and input type above comes from OpenRouter’s public model list, checked on November 11, 2026. Scores are Artificial Analysis Intelligence Index values shown on the same pages. Prices change often, so check the model page before you buy. Sources:
- Gemini 4 Carbon on OpenRouter
- Gemini 4 Argon on OpenRouter
- Claude Fable 5.2 on OpenRouter
- Claude Haiku 5.5 on OpenRouter
- Claude Sonnet 5.5 on OpenRouter
- Claude Opus 5.5 on OpenRouter
- Nemotron 3.5 Super VL on OpenRouter
- Qwen3.8 Max Prime on OpenRouter
- DeepSeek V4.1 Flash on OpenRouter
- DeepSeek V4.1 Flash 1017 on OpenRouter
- Meta Muse Wonder on OpenRouter
- GPT-6.1 Terra Pro on OpenRouter
- GPT-6.1 Terra on OpenRouter
- GPT-6.1 Sol Pro on OpenRouter
- GPT-6.1 Sol on OpenRouter
I publish short, plain-English guides like this one on my site. My AI assistant SIGNAL OS drafts and posts them for me so I can spend my time on other work.


