Directories / Models / Learn / Model families

Models · Guide

What is actually different between the model families

There is no single best model. There are families with different temperaments. Here is what each one is genuinely good at, when you would reach for it, and where the current flagship of each sits on the live board.

Updated Jul 21, 2026

Flagships pulled live from the model catalogue

Ask "which AI model is best" and you get an argument. Ask "which family fits this job" and you get an answer. The frontier models are close enough on most single tasks that the overall ranking is the least useful thing about them. What separates them is character: how closely they follow instructions, how they write, how much they can read at once, how they behave inside an agent loop, and what they cost to run at scale.

Below are the families worth knowing, split into the closed frontier labs and the open-weight models you can run yourself. Each card is the evergreen read on temperament and best-fit work. The specific model named on each card is the current top scorer from that family, pulled live so it stays honest as new versions ship.

The frontier labs

Closed models: the four you will hear about most

Accessed by API, not downloaded. These set the pace on raw capability and carry the deepest tool and integration support.

Claude

Anthropic

The careful colleague. Claude tends to follow long, specific instructions closely, writes in a natural voice, and is unusually strong at reading and reasoning over large amounts of text and code. It is the family most teams reach for when the work has to be right and readable rather than merely fast.

Best for:: Coding agents, long-document analysis, drafting that sounds human, high-stakes reasoning.
Watch:: Premium pricing at the top of the range. You pay for the quality, so reserve the flagship for the steps that need it.

Current flagship Claude Opus 4.5 1280 intel $5.00 in · 200K ctx

GPT and o-series

OpenAI

The broad generalist. The GPT line is fast, well-rounded, and backed by the widest tool and integration support of any family. The o-series reasoning models trade speed for deeper step-by-step thinking on hard problems. Between them they cover most jobs competently.

Best for:: General assistants, structured output and function calling, mixed workloads, anything that needs a big ecosystem around it.
Watch:: The lineup is wide and the names are confusing. Match the specific model to the job rather than defaulting to the newest one.

Current flagship GPT-5.2 1225 intel $1.75 in · 400K ctx

Gemini

Google

The long-context multimodal. Gemini handles very large inputs and mixes text, images, audio, and video comfortably. It is the natural pick when the input is huge or spans more than one medium, and it plugs cleanly into the Google ecosystem.

Best for:: Whole-repo or whole-document work, image and video understanding, teams already in Google Cloud.
Watch:: Quality per task can trail the very top of the reasoning board. Check it against the job before committing a workflow to it.

Current flagship Nano Banana 2 (Gemini 3.1 Flash Image Preview) 1287 intel $0.500 in · 131K ctx

Grok

xAI

The fast challenger. Grok ships quickly, prices aggressively, and has closed a lot of the capability gap. It is a strong value option when you want frontier-adjacent quality without frontier pricing.

Best for:: Cost-sensitive general work, high throughput, teams that want a capable non-incumbent option.
Watch:: A newer track record than the incumbents. Test it on your own tasks before you trust it with the important ones.

Current flagship Grok 4.20 1158 intel $1.25 in · 2M ctx

Open weight

Open models: cheaper, self-hostable, catching up fast

Weights you can download and run on your own infrastructure. The best of them now sit close to the closed flagships on reasoning and coding, at a fraction of the price.

DeepSeek

Strong reasoning and coding at a fraction of frontier cost. The family that proved open models can sit near the top of the benchmarks.

Best for: Cheap reasoning and coding, self-hosting, cost-controlled scale.

DeepSeek V3.2 Exp $0.270 in

Qwen

Alibaba

A deep, fast-moving family with sizes for every budget, from tiny on-device models to large flagships. Very strong multilingual coverage.

Best for: Multilingual work, on-device and self-hosted deployments, picking a size that fits the job.

Qwen3.5 Plus 2026-02-15 $0.260 in

Kimi

Moonshot AI

Long-context specialists with strong agentic and coding scores, priced to undercut the closed flagships.

Best for: Agent loops, long inputs, coding on a budget.

Kimi K2 0905 $0.600 in

GLM

Z.ai

Capable all-rounders that move fast on price. Frequent, aggressive price cuts make them a value watch.

Best for: General work at low cost, price-sensitive high volume.

GLM 5 Turbo $1.20 in

Llama

Mistral

Mistral AI

Efficient European models with a clean open-weight story and strong coding variants (Devstral).

Best for: Efficient inference, coding, European data-residency needs.

Mistral Medium 3 $0.400 in

Closed or open: how to actually decide

The honest split is about control versus convenience. Closed models give you the sharpest capability at the top end and the least operational work: you call an API and it is someone else's job to keep it running. Open models give you ownership, privacy, and a far lower cost per token, in exchange for hosting them yourself or picking a provider who does.

For most businesses the answer is not one or the other. The high-stakes, low-volume steps, where a wrong answer is expensive, go to a closed flagship. The high-volume, lower-stakes work, where good enough at a fraction of the cost is the whole point, goes to a cheaper open or mid model. Sensitive data that cannot leave your walls goes to a model you host. That mix is the norm, not the exception.

Which is the point most people miss when they shop for a single winner. You are not choosing a family to marry. You are assembling a small team of models and giving each the work it is best at. The next guide is how to make that call per task.

How to choose a model for a job → How to run many models, not one See the live board

Common questions

What is the real difference between Claude, GPT, and Gemini?

They are all frontier models, so the gap on any single task is often small. The differences that matter in practice are temperament and shape: Claude follows detailed instructions closely and writes naturally, GPT is the broad generalist with the biggest ecosystem, and Gemini leads on very large and multimodal inputs. The right pick depends on the job, not on an overall ranking.

Are open-weight models like DeepSeek and Qwen good enough to use?

Yes, for a growing share of work. The best open models now sit close to the closed flagships on reasoning and coding at a fraction of the cost, and you can run them on your own infrastructure. The closed flagships still lead at the very top on the hardest steps, so the common pattern is to use open models for volume and closed models for the high-stakes calls.

Should I just pick the highest-ranked model and use it for everything?

No. The top model is the most expensive and rarely the best value for routine work. Matching each step of a workflow to the cheapest model that clears the quality bar for that step is cheaper and often faster. That is model routing, and it is how operators actually run these families together.

How often does the ranking between families change?

Constantly. New models and price moves land most weeks, and the lead changes hands regularly. Any hard-coded "best model" answer is out of date within a month, which is why the picks on this page are pulled live from the catalogue.