AI Model Comparison

How to read this table

Every column answers a different question about a model. Read across a row and you get a quick profile; sort by a column and you can rank the whole field on the thing you care about most.

• Contextis the size of the model’s short-term memory, in tokens. It caps how much text — your prompt, attached documents, the chat history and the reply — the model can hold at once. We show it as 128K or 1M for readability.
• Input and output $/1M are the approximate prices, in US dollars, for a million tokens sent to the model and a million tokens it generates. Output is almost always the pricier of the two, often several times over.
• Inputs lists the modalities a model natively accepts — text alone, or text plus images, audio, even video. This decides whether you can hand it a screenshot, a chart or a voice clip.
• Open vs closed. Rows tagged Open ship their weights publicly, so they show Self-hostinstead of a price. Closed models are only reachable through their maker’s API at the listed rate.

The big model families in 2026

Four families dominate most conversations, with a lively fringe of challengers around them. Each has a genuine personality, and the gaps between the leaders are smaller than the marketing suggests.

OpenAI’s GPT line is the versatile default. The flagship GPT-4o is a capable all-rounder with a vast ecosystem of tools, integrations and tutorials around it, while GPT-4o mini handles high-volume work cheaply. The o-series reasoning models trade speed for depth, thinking longer before they answer, which pays off on maths, logic and thorny code.

Anthropic’s Claude line has earned a strong following for coding, careful reasoning and long-form writing. Opus is the heavyweight for the hardest problems; Sonnet is the balanced workhorse most people reach for day to day; Haiku is the fast, cheap option for high-throughput tasks. All three share a generous 200K-token context window.

Google’s Gemini line leans into scale and multimodality. Its context windows are the largest here — a million tokens or more — and it natively handles images, audio and video, which makes it a natural fit for large documents and mixed media. Flash gives you that huge window at a budget price.

Open-weight models— Meta’s Llama, DeepSeek and others — are the wildcard. Because their weights are public, you can run them on your own hardware, fine-tune them freely, and avoid sending data to a third party. The quality gap with closed frontier models has narrowed sharply, though you take on the cost and effort of hosting. Mistral, a European provider, and xAI’s Grok round out a field that is more crowded and more competitive than ever.

Context window — bigger isn’t automatically better

A million-token context window sounds like an obvious win, but it is a ceiling, not a promise. Three caveats matter. First, models often attend less reliably to information buried in the middle of a very long context — the so-called “lost in the middle” effect — so stuffing everything in can actually dilute the signal. Second, you pay for every token in the window on every call, so a bloated context is a bloated bill. Third, most real tasks simply do not need it: a focused prompt with the right few pages usually beats a giant dump of loosely related text.

The practical move is to fit what you need and no more. To translate a model’s token limit into something concrete — how many pages, words or chat turns it really holds — run the numbers through our context window calculator.

Pricing — what actually drives your bill

The per-token prices in the table are only the starting point. What you actually pay is tokens × price, summed over every call — and both halves move more than people expect. Output tokens usually cost several times more than input tokens, so verbose, open-ended answers get expensive fast. A long system prompt resent on every request quietly multiplies your input cost. And reasoning models can burn large numbers of hidden “thinking” tokens before they reply, which is why their headline price looks steep.

The cheapest model is rarely the cheapest solution, either: if a budget model needs three attempts and a long prompt to get a usable answer, a pricier model that nails it first time can cost less overall. To price a specific prompt across several models and see the monthly total, feed your token counts into our LLM cost calculator.

How to choose

Work from the task backwards, not from the brand forwards. A short checklist covers most decisions:

• How hard is the reasoning?For genuinely difficult logic, maths or code, a frontier or reasoning model earns its price. For routine drafting, summarising and Q&A, a mid-tier or mini model is plenty.
• How much text must fit? Whole codebases or long documents push you toward the largest context windows; short chats do not.
• What inputs do you have? If you need to pass images, audio or video, filter to models that accept them.
• How price-sensitive are you? At high volume, the gap between a flagship and a mini tier is enormous — test whether the cheaper model is good enough before paying for the flagship.
• Do you need to self-host? For privacy, control or fine-tuning, an open-weight model may matter more than a few points of benchmark quality.

If you would rather answer a few quick questions and get a specific recommendation instead of weighing every column yourself, our Which AI should you use tool walks you through it. And remember the whole field moves in months, not years, so treat any single comparison — including this one — as a snapshot rather than a verdict.