How large language models differ, and what to compare

Large language models differ in five ways that matter once you pay for them: how much text they hold at once (the context window), how hard they think before answering (reasoning effort), whether they can call tools, which kinds of input they accept, and what they charge per token. Rankings change every few months; these five tell you which model fits a job. Every figure below comes from the vendors' own pages, checked on 23 October 2026.

That date matters more than usual. Anthropic's models overview lists four current Claude models and nine legacy ones still available, OpenAI's model catalog leads with three flagship models, and Google's models page runs to dozens of model codes across Gemini 4, 3 and 2.5. Any leaderboard you read is a snapshot. The dimensions underneath it hold still, so that is what this post explains.

Everything is measured in tokens, the chunks of text a model reads and writes. A token is not a word, and the ratio moves even inside one vendor: Anthropic says 1M tokens is roughly 555,000 words on its current tokenizer, while its models before Claude Opus 4.7 fit about 750,000 words in the same 1M tokens (Anthropic models overview). So compare token prices within a vendor freely, and across vendors with care. If you want the layer underneath all of this, start with how neural networks work.

The context window is everything the model can see in one request: your instructions, the conversation so far, any documents and tool results, and the answer it is writing. Anthropic's context window docs describe it as the model's working memory, separate from the data it was trained on.

On paper the three vendors now look alike at the top. All four current Claude models have a 1M token window and write up to 128K tokens per request (Anthropic). GPT-6 Astra lists a 1,050,000 token window, of which at most 922,000 can be input, and 128,000 output tokens (OpenAI). Gemini 4 Argon has a 1,048,576 token window shared between input, thinking and output, with 65,536 output tokens by default and up to the full window in its long decode mode (Google).

Filling that window is a different matter. The same Anthropic page warns that as the token count grows, accuracy and recall degrade, which it calls context rot. Google's long context guide makes a related point: models score well when asked to find one fact in a long document, but accuracy drops when they have to find several at once.

Price can change with length too. Claude Haiku 5.5 charges more once a prompt passes 100,000 tokens, and OpenAI prices GPT-6 Astra prompts over 272K input tokens at 2x the input rate and 1.5x the output rate for the whole request (Anthropic pricing, OpenAI). That is why a smaller, curated context often beats a full one, and why retrieval augmented generation is still worth knowing about even with a 1M token window. In coding tools the same effect is what cutting Claude Code token usage is about.

Current models can think before they answer, and all three vendors now let you set how much. Anthropic calls it effort, with levels from low to max; it applies to every output token, including tool calls and thinking, and Claude Opus 5.5 and Claude Haiku 5.5 default to medium while the other current models default to high (Anthropic effort docs). OpenAI exposes reasoning.effort, from low to max on GPT-6 Astra, with GPT-6 Luna also accepting none (Astra, Luna). Google lists thinking levels of low, medium and high for Gemini 4 Argon (Google).

The cost follows directly. Anthropic bills thinking tokens as output tokens, and Google's price table says its output price includes thinking tokens (Anthropic, Google pricing). Output is the expensive side of every price list below, so raising effort on a task that did not need it is one of the quickest ways to inflate a bill. Anthropic's own table suggests low for simpler tasks such as subagents and xhigh for long agentic coding runs.

Tool use is what lets a model act instead of only talk: it asks for a function by name, your code runs it, and the result goes back into the context. Every current Claude model supports text and image input, text output and tool use (Anthropic). OpenAI lists function calling plus hosted tools such as web search, file search, code interpreter, computer use and MCP for GPT-6 Astra (OpenAI). Google marks function calling, code execution, search grounding and computer use in preview as supported on Gemini 4 Argon (Google). This is the layer that AI agents are built on, and it is what plugging into your stack usually means in practice.

Inputs are where the vendors still differ most. GPT-6 Astra takes text and images and writes text. Gemini 4 Argon takes text, images, video, audio and PDF, and also writes text. If your work starts from call recordings or screen captures, that difference decides the shortlist before price does.

Prices are quoted per million tokens, with input and output priced separately. These are the standard API list prices on 23 October 2026; each one links to the page it came from.

The spread is the point. Claude Haiku 5.5 and GPT-6 Luna charge $0.10 per million input tokens, a hundredth of Claude Fable 5.1 and GPT-6 Astra at $10 (Anthropic, OpenAI). Discounts move the numbers again: Anthropic and OpenAI both price batch requests at half the standard rate, and repeated context read from a cache costs a fraction of fresh input (Anthropic models overview, OpenAI). Google has also published a price rise for Gemini 4 Argon on 1 December 2026, which is a good reason to date any figure you quote.

A worked example shows what that means per task. Take a job that reads 50,000 tokens and writes 5,000. On Claude Sonnet 5.5 at $2 and $10 per million, that is $0.10 of input plus $0.05 of output, about $0.15 a run (Anthropic prices). On Claude Haiku 5.5 the same run costs under a cent. Whether the cheaper model is good enough is the question only your own tests can answer.

Benchmarks will not tell you how a model handles your documents, your customers or your codebase. OpenAI's guide to building agents gives a method that works for any vendor: build the first version with the most capable model to set a baseline, set up evals, then swap in smaller models wherever they still hit your accuracy target.

In practice that usually ends in a mix rather than one winner. A small model sorts and extracts, a mid tier model does most of the writing, and the expensive one is kept for the steps where it measurably earns its price. Anthropic's own advice on its models page is to start with Claude Opus 5.5 for most workloads and move up to Claude Fable 5.1 only when evals at higher effort still fall short.

This post sits in the AI explained series alongside the agentic AI explainer. For agents that run on these models in production, see the AI agents category, and if you want help choosing and wiring the right mix for your own work, that is part of our custom AI agents service.