Large language models, compared

There is no single best large language model. The names people are comparing in late 2026 are GPT, Claude, Gemini, Mistral, Llama, DeepSeek and Qwen. Which one is useful depends on the task, the context window you actually need, the deployment constraint, and how you will judge the output.

This page is a comparison list, not a ranking and not a buying recommendation. Model names and capabilities move. Check the vendor page before you treat a row as current.

The model families are listed above the illustration. The sections below say how to choose among them for a company task. They do not say which model is largest, cheapest or best in the abstract.

How to choose one for a company

For the organisational decisions around a model choice, continue with the Enterprise AI guide. For Spain’s public multilingual work, read about ALIA and Salamandra.

Methodology

How to use this comparison

This is an editorial selection of model families, not a measured ranking. The official sources below establish available products and model documentation. The evaluation criteria are a framework for your own assessment.

Compare an exact model version rather than a brand. Record the date, prompt, settings, tools and scoring rules. Keep benchmark results separate from provider claims and from results measured in your own environment.

LLM comparison table

Use this table to organise a shortlist. Capabilities, context limits, prices and licences belong to the specific model version, so check its official documentation before running a comparison.

GPT (OpenAI)

The GPT family is part of OpenAI’s model catalogue. The catalogue distinguishes models by capabilities and intended tasks. A ChatGPT subscription and an API deployment are different ways of using the products; assess the one your team actually needs.

For a shortlist, check input types, supported tools and output limits for the exact model. Then measure whether it follows your instructions and produces work a reviewer can accept. Keep the cost of retries and review in the comparison.

Claude (Anthropic)

Anthropic’s model overview describes Claude versions and their capabilities. A family-level claim such as “best for coding” hides differences between versions, settings and workloads.

Test document interpretation, code changes or writing with the material you expect to use. Include incomplete information and contradictory instructions. Compare the resulting errors and review effort with the other candidates under the same conditions.

Gemini (Google)

Google’s Gemini catalogue includes models with different input and output capabilities and release states. Check whether a particular model is stable or in preview, and whether it supports the media your task requires.

For work involving images, audio or video, evaluate those inputs directly. Good performance on text questions does not establish reliable interpretation of a chart, a recording or a scanned document.

Mistral (Mistral AI)

Mistral’s catalogue includes different model sizes and access options. If European hosting or downloadable weights matter to your organisation, investigate the specific product, licence and service terms rather than inferring them from the provider’s location.

Compare a hosted option and a deployment you would operate yourself on realistic infrastructure. Include administration, monitoring and maintenance in the assessment. Read the Enterprise AI guide for the wider questions about ownership and operating a use case.

Llama (Meta)

Meta’s official Llama organisation provides model repositories and access to licence information. Downloadable weights give you deployment choices, but each model still has conditions and hardware requirements.

Evaluate the exact weights and serving setup you plan to use. Quantisation, prompts and available compute can change the result; a score reported for another configuration is not a measurement of yours.

DeepSeek

DeepSeek’s API documentation identifies its current endpoints and usage conventions. Model names, aliases and available features can change, so record the exact endpoint used in your evaluation.

Compare task quality, latency and total usage cost. Assess data handling and contractual conditions separately from model capability. A low token price does not answer those questions or establish the cost of a successfully completed task.

Qwen (Alibaba)

Qwen’s official model repository documents a family with different model configurations. Review the model card and licence for the individual release rather than treating every Qwen model as interchangeable.

For multilingual work, test the languages and terminology your users need. For a local deployment, check memory use, serving performance and the quality of the chosen configuration on your own examples.

Open vs closed models

With a managed API, a provider operates the model-serving infrastructure. With downloadable weights, you may operate a model yourself or use a hosting service. These choices change who is responsible for updates, security, capacity and support.

Open-weight does not automatically mean unrestricted use, complete training transparency or legal compliance. Managed does not automatically mean the provider’s terms fit your data requirements. Inspect the licence, hosting region, retention settings and contract for the route you would use.

The EU AI Act introduction helps identify regulatory questions. It is a starting point for governance discussions, rather than a determination that a model or deployment complies with the law.

Which LLM to choose

Write a short evaluation brief before you select a winner. It should identify the task, the people affected, the available information and the quality you need.

Start with a small set of representative examples that you can score consistently. Keep some examples outside the prompt-development process so the comparison does not reward a setup tuned only to familiar questions. Ask reviewers to explain why they accepted or rejected an output.

Choose the simplest candidate that meets your acceptance criteria. If none does, narrow the task or change the workflow before expanding deployment. Re-evaluate when the provider changes the version or your task changes.

For a leadership discussion, bring a concrete question: which task is being considered, what evidence you have, what remains uncertain and who would own the decision. The Enterprise AI guide helps place that model choice in an adoption plan.

If the model comparison informs a leadership decision, a private AI executive briefing can help frame one model-selection question, its evaluation criteria and the evidence to gather. Technical benchmarking and deployment remain with your technical team or a separately appointed specialist.

If this interests you

• Plan adoption: the Enterprise AI guide connects a tool choice to ownership and business use.
• Explore public multilingual models: ALIA and Salamandra provide a Spanish ecosystem reference.
• Investigate governance: the EU AI Act introduction identifies questions for further review.

Sources

• OpenAI model catalogue
• Anthropic model overview
• Google Gemini model catalogue
• Mistral model catalogue
• Meta’s official Llama model organisation
• DeepSeek API documentation
• Alibaba Cloud’s official Qwen repository

Official catalogues consulted on 3 October 2026. Product availability and terms can change after that date.