AI models
AI models are the engines behind chatbots, coding assistants and business tools. This grid compares the current large language models from the major makers side by side, so you can weigh price, context size, input types, reasoning, tool use, privacy defaults and whether you can run the model yourself. The real trade-off is quality against cost and control: top models are the most capable but cost many times more per token than small ones, and open-weights models give you control at the price of doing the hosting.
Last updated
How to choose
Start with the job, not the leaderboard. A model that is best at hard reasoning may be wasteful for summarizing emails.
1. Decide what you need
- What goes in: plain text, images, audio or video? Check the Image input, Audio input and Video input features.
- How much text per request? Compare Context window and Max output tokens against your longest document or code file.
- Will it call your tools or return data your software reads? You need Tool calling and Structured outputs.
2. Rule out deal-breakers
- Sensitive data: check Trains on API data and whether the model is on the cloud your company already uses (Major cloud availability).
- Must run on your own servers or offline: only Open weights models qualify.
3. Compare full cost
- Cost is input plus output. Chat and agents are usually dominated by output price; document analysis by input price.
- Look at Cached input price and Batch discount if you repeat prompts or can wait for results.
- Reasoning modes can generate many hidden tokens that are billed as output, so real cost can be well above the headline price.
- If you only want a chat app, compare the Cheapest paid plan with access and whether there is free access.
4. Judge quality claims
- Benchmarks are the makers' own marketing unless independent. The Independent intelligence score is one outside measure; try two or three finalists on your own real prompts before committing.
5. Before you commit
- Models are retired on a schedule, so check the maker's deprecation policy and avoid hard-coding a single model.
- Build so you can swap models: prices and rankings change every few months.
Answer
Browse the table. Filter to what matters.
Use the filters below, then sign in if you want to shortlist and decide with friends.
Sign in to vote, hide services and features, and reorder.
Verified: hover a cell and click its green dot for the source Estimate: no cited source Not stated: the company doesn't publish it
| Feature | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ | ⠿ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
⠿
Maker type
|
Closed
|
Closed
|
Closed
|
Closed
|
Closed
|
Closed
|
Closed
|
Closed
|
Closed
|
Closed
|
Open weights
|
Closed
|
Closed
|
Closed
|
Closed
|
Closed
|
||||||||||||||||||||||
|
⠿
Open weights
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
||||||||||||||||||||||
|
⠿
Tier
|
Flagship
|
Mid-range
|
Mid-range
|
Mid-range
|
Flagship
|
Mid-range
|
Flagship
|
Mid-range
|
Small and fast
|
|||||||||||||||||||||||||||||
|
⠿
Release date
|
Aug 2026
|
Sep 2026
|
||||||||||||||||||||||||||||||||||||
|
⠿
Knowledge cutoff
|
||||||||||||||||||||||||||||||||||||||
|
⠿
API model ID
|
||||||||||||||||||||||||||||||||||||||
|
⠿
Context window
|
||||||||||||||||||||||||||||||||||||||
|
⠿
Max output tokens
|
Doesn't apply
|
|||||||||||||||||||||||||||||||||||||
|
⠿
Input price
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
|||||||||||||||||||||||||||||||||
|
⠿
Output price
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
|||||||||||||||||||||||||||||||||
|
⠿
Cached input price
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
$0.20/1M tokens
|
$0.015/1M tokens
|
Doesn't apply
|
Doesn't apply
|
|||||||||||||||||||||||||||||||
|
⠿
Batch discount
|
Doesn't apply
|
None
|
None
|
Doesn't apply
|
Doesn't apply
|
|||||||||||||||||||||||||||||||||
|
⠿
Image input
|
||||||||||||||||||||||||||||||||||||||
|
⠿
Audio input
|
||||||||||||||||||||||||||||||||||||||
|
⠿
Video input
|
||||||||||||||||||||||||||||||||||||||
|
⠿
Reasoning mode
|
|
|
||||||||||||||||||||||||||||||||||||
|
⠿
Tool calling
|
||||||||||||||||||||||||||||||||||||||
|
⠿
Structured outputs
|
Doesn't apply
|
|||||||||||||||||||||||||||||||||||||
|
⠿
Built in web search
|
|
|
|
|
|
|||||||||||||||||||||||||||||||||
|
⠿
Built in code execution
|
|
|
|
|
|
|||||||||||||||||||||||||||||||||
|
⠿
Fine tuning
|
No
|
No
|
No
|
No
|
No
|
No
|
Weights allow it
|
No
|
No
|
No
|
No
|
No
|
||||||||||||||||||||||||||
|
⠿
Consumer app free access
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
|||||||||||||||||
|
⠿
Cheapest paid plan with access
|
$19.99/month
|
Doesn't apply
|
$20/month
|
$20/month
|
$20/month
|
$8/month
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
Doesn't apply
|
||||||||||||||||||||||||
|
⠿
Major cloud availability
|
Azure
|
None
|
Azure
|
None
|
Azure, Google Vertex AI
|
Azure, Google Vertex AI
|
Azure, Google Vertex AI
|
Azure, Google Vertex AI
|
||||||||||||||||||||||||||||||
|
⠿
Trains on API data
|
Doesn't apply
|
Doesn't apply
|
||||||||||||||||||||||||||||||||||||
|
⠿
Independent intelligence score
|
14 points
|
12 points
|
9 points
|
44 points
|
Why these features matter
-
Maker type
Open weights can be self-hosted and fine-tuned privately; closed models are usually stronger at the top end and easier to start with.
-
Open weights
Decides whether you can run it on your own hardware, keep data in-house, or fine-tune it freely.
-
Tier
A flagship costs several times more than a small model and is only worth it for hard tasks.
-
Release date
Newer models usually know more recent events and score higher, but older ones are often cheaper and more settled.
-
Knowledge cutoff
Past the cutoff the model knows nothing unless it can search the web.
-
API model ID
Needed to call the model and to tell versions apart.
-
Context window (tokens)
Sets how much text, code or documents fit in one request.
-
Max output tokens (tokens)
Limits long reports, big code files and long structured outputs.
-
Input price ($/1M tokens)
The biggest driver of cost for document-heavy or long-prompt use.
-
Output price ($/1M tokens)
Output is usually 3 to 8 times pricier than input and dominates chat and agent costs.
-
Cached input price ($/1M tokens)
Repeated prompts, system instructions and agent loops get much cheaper with caching.
-
Batch discount
Halves the bill for work that can wait hours, like bulk classification.
-
Image input
Needed for screenshots, charts, documents and photos.
-
Audio input
Enables voice assistants and meeting analysis without extra steps.
-
Video input
Matters for analysing recordings, lectures and clips.
-
Reasoning mode
Improves accuracy on math, code and planning at the cost of speed and tokens.
-
Tool calling
Required to build agents that use your software and data.
-
Structured outputs
Makes outputs reliably machine readable.
-
Built in web search
Gives current information without building your own search.
-
Built in code execution
Lets the model compute and analyse data instead of guessing numbers.
-
Fine tuning
Lets you teach the model a style or task, if you need that.
-
Consumer app free access
Free access is the easiest way to try it.
-
Cheapest paid plan with access ($/month)
The monthly cost for a regular person to use it heavily.
-
Major cloud availability
Companies often must buy through the cloud they already use.
-
Trains on API data
A key privacy and compliance question for business data.
-
Independent intelligence score (points)
A single independent test score, useful to compare quality between makers.
About the services
-
Claude Fable 5.2 · Anthropic
Anthropic's latest flagship for demanding reasoning, deep research and long-horizon agentic work. It keeps a 1M-token context and 128K maximum output while cutting API prices 20% from Fable 5.1.
-
Claude Fable 5.1 · Anthropic · Superseded Nov 2026
Anthropic's previous flagship for demanding reasoning and long-horizon agentic work. Fable 5.2 superseded it on November 9, 2026 with the same context and output limits at lower prices.
-
Claude Opus 5.5 · Anthropic
Anthropic's recommended starting point for most workloads, built for long-running agentic coding and knowledge work. Thinking is always on and cannot be disabled. Scores highest of the Claude family on Artificial Analysis.
-
Claude Sonnet 5.5 · Anthropic
The balanced Claude model: fast, with near-Opus scores at half the Opus price. A good default for production apps and coding. Free in the Claude app.
-
Claude Haiku 5.5 · Anthropic
Anthropic's fastest and cheapest model, for classification, extraction and routing at volume. Prices rise 5x for prompts over 100,000 tokens, unlike the other three. Free in the Claude app.
-
Gemini 4 Argon · Google
Google's flagship reasoning model, generally available Oct 2026 on the paid Gemini API. Takes text, image, audio, video and PDF input and can output up to 1M tokens. Current $2/$10 price is introductory; sources report it rising Dec 1, 2026.
-
Gemini 3.8 Flash · Google
Google's top Flash model for coding and agent work, at a fraction of Argon's price. Multimodal input, 1M context, free API tier. Listed price holds through Dec 31, 2026.
-
Gemini 3.6 Flash · Google
Previous-generation Flash, still sold at the same price as 3.8 Flash; most buyers should pick 3.8 Flash instead unless pinned to this version.
-
Gemini 3.5 Flash-Lite · Google
Cheapest current Gemini tier for high-volume tasks, with 1M context and multimodal input. Scores far below the Flash and Argon models.
-
Gemma 4 31B · Google
Google's largest open-weights Gemma 4 dense model under Apache 2.0, for self-hosting and fine-tuning. Text and image input; no Google paid API (free tier only).
-
Mistral Large 3 · Mistral AI
Mistral's top generally available model: a 675B mixture-of-experts multimodal model with open Apache 2.0 weights and low API prices. Suits teams wanting an open, EU-based flagship; it trails the newest frontier models and Large 4 is still in public preview.
-
Mistral Medium 3.5 · Mistral AI
Mistral's mid-tier multimodal model tuned for agents and coding, released with open weights under a Modified MIT license. It costs more per token than Large 3, so pick it for coding agents rather than price.
-
Mistral Small 4 · Mistral AI
A 119B-parameter (6.5B active) hybrid model that combines instruct, reasoning and coding, under Apache 2.0. A cheap, fast option for high-volume work.
-
DeepSeek V4.1 Flash · DeepSeek
DeepSeek's fast, very cheap model with 1M context, native image understanding and thinking mode. Peak-hour pricing is double off-peak. Data is handled by a China-based provider, which matters for compliance.
-
DeepSeek V4 Pro · DeepSeek
DeepSeek's larger model with 1M context and thinking mode, text-only, at a fraction of Western flagship prices. Peak-hour pricing is double off-peak; lower independent score than V4.1 Flash.
-
GPT-6 Astra · OpenAI
OpenAI's most capable model, for the hardest professional and agentic work. Priced at 5x GPT-6.1 Sol for a small intelligence gain; text and image input only, no fine-tuning.
-
GPT-6.1 Sol · OpenAI
Near-Astra quality at one fifth the price, the sensible default for complex work and agents. 1M-token context, reasoning, text and image input.
-
GPT-6.1 Terra · OpenAI
OpenAI's middle tier: half Sol's price and faster, aimed at visual reasoning, computer use and writing. 1M-token context, reasoning, text and image input. In ChatGPT for paid plans only.
-
GPT-6 Luna · OpenAI
OpenAI's cheapest current API model, for high-volume, focused tasks. Also the model free ChatGPT users get. Much lower scores on hard reasoning than Sol or Astra.
-
gpt-oss-120b · OpenAI
Open-weight (Apache 2.0) reasoning model that runs on a single H100. Self-host or fine-tune privately; text only, and well behind the GPT-6 family on quality.
-
gpt-oss-20b · OpenAI
Smaller Apache 2.0 open-weight reasoning model for low-latency and local use. Text only; the weakest option here on quality, but free to run yourself.
-
Qwen3.8-Max · Alibaba Cloud
Alibaba's top proprietary Qwen model (Aug 2026) with a 1M-token window and thinking mode. Priced well below US frontier flagships; Alibaba Cloud hosting is mainly outside US regions for some features.
-
Qwen3.8-Flash · Alibaba Cloud
Cheap, fast Qwen model with downloadable weights (released Aug 2026), suited to high-volume tasks and self-hosting.
-
GLM-5.3 · Z.ai
Z.ai's flagship text-only reasoning model with 1M context and 128K output, aimed at coding and agents at about one-third of US flagship prices. No image input.
-
GLM-5.3-Flash · Z.ai
Z.ai's low-cost fast tier at $0.15 input and $0.50 output per 1M tokens; good for bulk and agent sub-tasks.
-
Kimi K3.1 Pro · Moonshot AI
Moonshot's current flagship (Oct 2026), a further-trained Kimi K3 with a 1M context, always-on thinking and image and video input, and cheaper than K3 at $2/$10 per 1M tokens. Weights are promised by Nov 4, 2026; until then it is API and app only.
-
Kimi K3 · Moonshot AI · Superseded
Moonshot open-weights model with 1M context, always-on thinking and native vision, now the previous-generation flagship after Kimi K3.1 Pro. Strong for long agentic coding; the first-party web search tool is currently not recommended.
-
Kimi K2.6 · Moonshot AI · Superseded
Previous-generation Kimi at a lower price with 256K context, thinking and non-thinking modes and image input. Moonshot closes new registrations for the K2 series on October 31, 2026 and plans to retire it by the end of December 2026.
-
Nova 2 Lite · Amazon
Amazon's current generally available Nova model: 1M context, text, image and video input, optional reasoning, built-in web grounding and code interpreter. Available through Bedrock only.
-
Nova Pro · Amazon
First-generation balanced Nova model with 300K context and image/video input; still active on Bedrock but older and without reasoning.
-
Nova Micro · Amazon
Text-only, very cheap and fast Nova model with 128K context for summarising, translating and classifying at scale.
-
Grok 4.7 · xAI
xAI's current flagship for coding, agents and knowledge work, with configurable reasoning and a 500k context. Priced at $2/$6 per 1M tokens (doubling above 200k prompt tokens); text and image input only.
-
Grok 4.6 · xAI
xAI's August 2026 model, between Grok 4.5 and 4.7, at the same $2/$6 per 1M tokens (doubling above 200k prompt tokens). Little reason to pick it over 4.7 unless you are already pinned to it; text and image input only.
-
Grok 4.5 · xAI
The previous Grok flagship from July 2026, same $2/$6 price as 4.7 with a cheaper cached rate. Little reason to pick it over 4.7 unless you are already pinned to it.
-
Grok 4.3 · xAI
The cheaper Grok with a 1M context window at $1.25/$2.50 per 1M tokens. Suits high-volume tool-calling work where top-end reasoning is not needed.
-
Muse Spark 1.3 · Meta
Meta's recommended hosted model for agentic and coding work, with 1M context and text, image, video and PDF input at $1.25/$4.25 per 1M tokens. Closed weights; Meta says audio support in 1.3 is not fully supported, and the API is still new.
-
Muse Spark Open · Meta
Open-weight release of Muse Spark (816B parameters, 42B active, 1M context) from Oct 2026. Needs very large GPU hardware to self-host or a third-party host; Meta does not serve it on its own API, and the Muse license has commercial conditions.
-
Muse Glimmer · Meta
Meta's small 30B Apache-2.0 open-weight model, distilled from Muse Spark, that runs on a single consumer GPU. Text and image input only, 131k context, and weaker on hard tasks than the larger models.
Suggest a change
Related comparisons
-
AI coding toolsAI coding tools now range from autocomplete in your current editor to agents that work on tasks in the cloud while you do something else. This grid compares popular tools side by side on the practical things that decide a purchase: where they run, what a plan really costs, which models and agent features you get, and how your code is handled. The real trade-off is power against predictability and control: the most capable agents cost more, often bill by usage, and need more care over privacy and review.
-
AI search enginesAI search tools read the web for you and answer in a few paragraphs with links, instead of giving you ten blue links. This grid compares popular AI answer engines, search engines that now add AI answers, and the developer APIs that power AI apps, on what decides a choice: what is free, what paid plans cost, whether answers cite sources, how deep the research goes, who runs the search index, and what happens to your queries. The real trade-offs are convenience against privacy, and speed against how much you can trust and verify an answer.
-
Browser-based AI writing toolsBrowser-based AI writing tools draft, rewrite, and polish text where you already write: in a web app or through a browser extension in email, documents, and web forms. The choice comes down to what you write (marketing copy, everyday email, or polished English), how much you pay per month, and whether your team needs brand voice and team administration.
-
Website translation toolsWebsite translation tools range from WordPress plugins to hosted services that translate pages across many site platforms. The main trade-offs are setup and platform fit, how much control reviewers have over machine translations, search-friendly language URLs, and how word and language limits shape the ongoing cost.
-
AI voice generatorsAI voice generators turn text into natural speech and can clone or design voices. The market splits three ways: developer platforms sold by the character, creator apps sold by the month, and open-weights models you run yourself for free. This grid compares popular products side by side on what decides a choice: price and allowance, languages and voices, cloning and emotion control, speed for live conversation, commercial rights and consent safeguards. The real trade-offs are quality and polish against cost, and convenience against control.
-
AI image generatorsAI image generators turn a text prompt or a reference picture into a finished image. This grid compares popular tools side by side on what decides a choice: what you can do for free, what the paid plans really include, which model powers each one, whether it edits and keeps characters consistent, and what rights you get to use the results commercially. The real trade-offs are quality and control against price, and ease of use against how open and legally safe the output is.