Skip to content

Command Palette

Search for a command to run...

AI for Everyone

Choose the right AI model

No model is best at everything. This guide helps you pick by task, data sensitivity and effort, and explains the difference between a model, an app, an API and an agent.

Last reviewed . AI products change quickly; check the linked sources before relying on details.

Model, app, API, agent and platform

  • Model: the trained system that produces text, images or audio. It has no interface by itself.
  • App: a product people use, such as a chat assistant. One app can switch between several models.
  • API: how developers call a model from their own software. Usually billed per use and separate from an app subscription.
  • Agent: a system where the model chooses steps and uses tools such as browsing or email. More capable, and riskier.
  • Platform: a service bundling models, tools, storage and admin controls, such as a cloud AI service.

Where a model runs also matters:

  • Consumer apps: easiest, with consumer data terms.
  • API: for building; business data terms.
  • Open-weight: weights you can download. See open-source AI.
  • Hosted open-weight: someone else runs an open model for you; their terms apply.
  • Local: runs on your computer; data stays with you, capability is limited by hardware.

There is no universally best model

Rankings change monthly, and a model that leads on one kind of task can trail on another. The practical approach is to choose two candidates, run the same real task in both, and keep the one that needs less fixing. Treat any claim of an overall winner as marketing unless it names the task, the date and the method.

Choosing a kind of AI optionStart from your task. Check what it needs: privacy, context length, cost and tools. Those needs point to a kind of option: a consumer app, an API, or an open-weight model you run yourself.Your taskwhat you needdonePrivacycan the data leave?Context lengthhow much to readCostper use or per monthToolssearch, files, codeConsumer appchat in browser or phoneAPIyour software calls itOpen-weight, localyou run the model yourself
Pick the kind of option from your needs first, then compare specific models.

What to compare

This table lists the dimensions that matter and what to look for. It is deliberately qualitative: figures for limits and prices change often, so use the provider links.

DimensionWhat to check
WritingTone, structure and editing quality. Judge on your own text.
ResearchWeb search, citations, and reading uploaded sources. Citations still need checking.
ReasoningMulti-step problems. Reasoning modes are slower but often more reliable.
CodingDepends on the tool around the model (editor, tests, repository access) as much as the model.
Long contextHow much text fits. Fitting is not the same as recalling accurately.
MultimodalWhether it reads images, audio, video, and whether it can also generate them.
DocumentsFile types, size limits and how well tables and scans are handled.
Tool useConnecting to apps, running code, browsing. Look for approval controls.
Speed and usage limitsFree and lower tiers usually have tighter limits. Limits change often.
APIWhether a developer API exists, and whether it is priced separately from the app.
PricingSee each provider's pricing page; this guide does not quote prices. For engineering comparisons use /pricing.
PrivacyTraining use, retention, enterprise terms and where data is processed.
IntegrationsEmail, calendar, office suites, chat apps.
Suitable for non-technical usersApp polish, setup effort, clarity of controls.

Benchmarks: what they show and miss

  • A benchmark measures a fixed test set. It does not measure your task, your documents or your tolerance for errors.
  • Results are often reported by the vendor, on selected tests, with different settings between models.
  • Models can be tuned toward public tests, so scores can overstate everyday quality.
  • Use benchmarks to shortlist, then decide with a small test on your own work.

Recommendations by task

These are starting points described by model family, not rankings. Check the current model names on each provider's page.

Writing and editing

Style and tone matter more than raw capability. Try two and compare on your own text.

  • Claude. Often liked for measured, natural prose and long-document editing. Trade-off: Usage limits on free and lower tiers can interrupt long sessions.
  • ChatGPT. Flexible, many tools for formatting, images and files. Trade-off: Default style can be more generic; give a voice sample.
  • Gemini. Works inside Google Docs and Gmail if you live in Workspace. Trade-off: Quality of tone varies by tier; test with your own examples.

Research with sources

Pick tools that cite and link sources, then open the sources yourself.

  • ChatGPT with search or research modes. Searches the web and returns linked sources. Trade-off: Citations can be wrong or irrelevant; check each one.
  • Gemini. Search integration and long-context reading of many documents. Trade-off: Can blend sources; ask for quotes and links.
  • Claude with web search or uploaded papers. Good at careful summaries of documents you provide. Trade-off: Web features depend on plan and region.

Reasoning and analysis

Use the provider's reasoning or thinking mode for multi-step problems, and verify the result.

  • Reasoning modes from OpenAI, Anthropic or Google. Spend extra computation on multi-step problems. Trade-off: Slower and may count more against usage limits.
  • DeepSeek or other open-weight reasoning models. Strong value; can be self-hosted. Trade-off: Hosted service raises data-residency questions.

Coding

Evaluate on your own codebase. Coding tools wrap models with editors, tests and repository access.

  • Claude (including its coding tools). Widely used for agentic coding. Trade-off: Costs and limits rise with long sessions.
  • OpenAI coding tools. Deep editor and cloud-task integrations. Trade-off: Features differ by plan.
  • Open-weight coding models (GLM, Qwen, Kimi, DeepSeek). Lower cost, self-hostable. Trade-off: Setup effort and hardware needs.

Long documents

Context window size is a ceiling, not a guarantee of accurate recall. Spot-check specific facts.

  • Gemini. Known for very long context windows. Trade-off: Recall of details in the middle can still be weak.
  • Claude. Strong at structured summaries of long reports. Trade-off: Limits on file sizes by plan.

Images, audio and video (multimodal)

Input understanding and content generation are different capabilities; check which you need.

  • Gemini. Broad support for image, audio and video input. Trade-off: Generation features vary by plan and region.
  • ChatGPT. Voice mode and image generation in one app. Trade-off: Some features are plan-limited.

Agentic and tool use

Look for approval controls and logs, not only autonomy. See the agents page for the safety model.

  • Claude and OpenAI agent products. Mature tool-calling and computer-use features. Trade-off: New agent products may be region- or plan-limited (for example reported OpenAI dots).
  • Self-hosted agent frameworks. Control over data and steps. Trade-off: A framework is not a finished product; you own setup and safety.

Privacy-sensitive work

The safest data is data that never leaves your machine. Otherwise use business or API terms, not consumer defaults.

  • Local open-weight models (Ollama, LM Studio, llama.cpp). Data stays on your device. Trade-off: Smaller models, hardware limits, you maintain it.
  • Business or enterprise plans from a major provider. Contractual data terms, admin controls. Trade-off: Cost and procurement; still read the terms.

Low budget

Free tiers are enough for light use; heavy use pushes you to paid plans or local models.

  • Free tiers of ChatGPT, Claude, Gemini. No cost to try several. Trade-off: Limits and older models; terms for data use differ.
  • Local models via Ollama or LM Studio. No per-use cost after hardware. Trade-off: Quality and speed depend on your computer.

The model landscape

As of October 2026, these are the providers most people meet. Names of specific model versions change quickly and several could not be verified on official pages in this review, so each entry says how well it was checked.

OpenAI (ChatGPT)

Secondary source
Consumer app
ChatGPT on web, desktop and mobile, with free and paid tiers.
API
OpenAI API for developers, separate from ChatGPT subscriptions.
Open weights
Has released some open-weight models; most flagship models are closed. Check the model card licence.
Notable strengths
Broad general-purpose assistant; Large ecosystem of integrations and custom assistants; Strong multimodal and voice features in the app.
Privacy and data control
Consumer and business tiers usually have different data terms. Look in the app's settings for data controls (training and chat-history options) and on the provider's business or API terms page for enterprise terms. Check these before pasting anything sensitive. ChatGPT Business and Enterprise plans and the API have their own terms.

Verification: Current flagship model names not verified on an official page. Reports say the latest family is the default in ChatGPT (October 2026); check OpenAI's model page. Always-on ChatGPT agents called dots were reported on 2026-09-29 as limited to Pro and Business Premium and unavailable in the EEA, UK and Switzerland; the official OpenAI page was not fetched.

ChatGPT · OpenAI platform

Anthropic (Claude)

Check provider
Consumer app
Claude on web, desktop and mobile, with free and paid tiers.
API
Claude API and availability through major cloud platforms.
Open weights
No open-weight Claude models. Closed weights only.
Notable strengths
Long-document reading and careful writing; Coding and agentic tool use; Clear, measured tone.
Privacy and data control
Consumer and business tiers usually have different data terms. Look in the app's settings for data controls (training and chat-history options) and on the provider's business or API terms page for enterprise terms. Check these before pasting anything sensitive.

Verification: Model tier names (Opus, Sonnet, Haiku are tiers within a generation) were not verified on an official page in this review. Check the provider's model page for the current generation.

Claude · Anthropic docs

Google (Gemini)

Secondary source
Consumer app
Gemini app; a new Gemini generation was reported available in the app for AI Pro and Ultra plans from 2026-10-13.
API
Gemini API via Google AI Studio and Google Cloud.
Open weights
Gemma is Google's open-weight family, separate from Gemini. Check the Gemma licence terms on the model card.
Notable strengths
Very long context and multimodal input; Tight integration with Google Workspace, Search and Android; Wide choice of speed and cost tiers.
Privacy and data control
Consumer and business tiers usually have different data terms. Look in the app's settings for data controls (training and chat-history options) and on the provider's business or API terms page for enterprise terms. Check these before pasting anything sensitive. Workspace and Google Cloud customers have separate admin controls.

Verification: The Google blog URL is the source for the app availability report; the page itself was not re-fetched for this review.

Gemini availability (Google blog) · Gemini API docs

xAI (Grok)

Check provider
Consumer app
The Grok app and the Grok features inside X.
API
xAI API, a separate product.
Open weights
Some older Grok models have been released with open weights; check xAI's repositories and licence.
Notable strengths
Access to live X content inside the X product; Real-time-oriented answers.
Privacy and data control
Consumer and business tiers usually have different data terms. Look in the app's settings for data controls (training and chat-history options) and on the provider's business or API terms page for enterprise terms. Check these before pasting anything sensitive. Posts on X and chats with Grok may be governed by different terms.

Disambiguation: Distinguish four things: the Grok app, the xAI API, @grok replies on X, and unofficial third-party bots using the name.

Verification: Latest Grok model naming not verified (a coding model was reported as Grok 4.5 in July 2026). The Grok app, the xAI API and @grok on X are different surfaces. Third-party accounts called Grok bots are not xAI products.

xAI

Meta AI

Secondary source
Consumer app
Meta AI app and assistants inside Facebook, Instagram, WhatsApp and Meta glasses; availability varies by country (US first for the newest model).
API
Check Meta's developer pages for current API access.
Open weights
Llama is Meta's open-weight family under the Llama community licence (not an OSI open-source licence). Muse Spark launched with closed weights.
Notable strengths
Reach inside apps people already use; Llama family is a common base for self-hosted models.
Privacy and data control
Consumer and business tiers usually have different data terms. Look in the app's settings for data controls (training and chat-history options) and on the provider's business or API terms page for enterprise terms. Check these before pasting anything sensitive. Meta AI inside social apps is tied to Meta account and app settings.

Disambiguation: Meta Muse Spark and Microsoft Muse are different products from different companies.

Verification: Muse Spark (2026-04-08) is reported as the first Meta Superintelligence Labs model, closed weights at launch, powering the Meta AI app, glasses and Meta apps, US first, with open-weight versions planned per Meta. Source is a news report. Microsoft's Muse (a 2025 gameplay model) is unrelated.

Meta AI · Muse Spark launch coverage (Fortune)

DeepSeek

Check provider
Consumer app
DeepSeek chat app and website.
API
DeepSeek API, typically priced low relative to many rivals (check their pricing page).
Open weights
Several models released with open weights. Check each model card for the licence.
Notable strengths
Strong reasoning and coding for the cost; Open weights allow self-hosting.
Privacy and data control
Consumer and business tiers usually have different data terms. Look in the app's settings for data controls (training and chat-history options) and on the provider's business or API terms page for enterprise terms. Check these before pasting anything sensitive. The hosted service is operated from China; organisations with data residency or regulatory requirements should read the privacy policy and consider self-hosting the open weights instead of using the hosted app.

Verification: A V4 generation was reported but not verified on an official page.

DeepSeek

Moonshot AI (Kimi)

Check provider
Consumer app
Kimi chat app.
API
Moonshot platform API.
Open weights
Kimi models have been released with open weights; check the model card licence for commercial terms.
Notable strengths
Long context and agentic tool use in the open-weight space.
Privacy and data control
Consumer and business tiers usually have different data terms. Look in the app's settings for data controls (training and chat-history options) and on the provider's business or API terms page for enterprise terms. Check these before pasting anything sensitive. Review where the hosted service stores data.

Verification: Current K2.x naming not verified on an official page.

Kimi

Z.AI (GLM)

Check provider
Consumer app
Z.AI chat.
API
Z.AI open platform API.
Open weights
GLM models have been released with open weights; check the model card licence.
Notable strengths
Coding and agent-style tasks in the open-weight space.
Privacy and data control
Consumer and business tiers usually have different data terms. Look in the app's settings for data controls (training and chat-history options) and on the provider's business or API terms page for enterprise terms. Check these before pasting anything sensitive.

Verification: Current GLM 5.x naming not verified on an official page.

Z.AI

Mistral (other option)

Check provider
Consumer app
Le Chat.
API
Mistral API (La Plateforme).
Open weights
Mix of open-weight and commercial models; the licence differs per model, so read the model card.
Notable strengths
European provider, useful if EU hosting matters; Smaller efficient models that suit local use.
Privacy and data control
Consumer and business tiers usually have different data terms. Look in the app's settings for data controls (training and chat-history options) and on the provider's business or API terms page for enterprise terms. Check these before pasting anything sensitive.

Verification: Not checked in detail in this review.

Mistral AI

Key takeaway
Meta Muse and Microsoft Muse. Meta's Muse Spark (reported launch 8 April 2026) is a language model behind Meta AI. Microsoft's Muse, announced in early 2025, is a research model for generating gameplay. They are unrelated.
Production warning
Grok means several things. The Grok app, the xAI API and @grok on X are separate surfaces. Accounts called Grok bots from other people are not xAI products. Check where a Grok answer came from before trusting its terms or data handling.

Model-agnostic prompting

Prompts transfer between models better than most people expect when they are explicit about goal, material, format and rules. Adapt to the provider in small ways:

  • Format: state the output format and length. Some models follow strict structure better than others.
  • Context window: paste only what is needed, and put the key instruction at the start and the end of very long inputs.
  • Provider differences: some assistants prefer short, direct instructions; others do well with detailed role and rules. Try both.
  • Verify facts: models can state wrong things confidently. Ask for what is assumed, and check anything that matters.
  • Switch models when the answers are repeatedly wrong in the same way, the output format keeps breaking, or limits block your work. Switching rarely helps if the prompt is unclear.

Reusable prompt structure

A structure that works on any major assistant.

Model-agnostic
Use it for
Getting consistent results when you switch tools
Who it is for
Anyone using more than one assistant
You will need
  • Your task
  • Source material
  • Desired format
What you get back
A response in your format that flags gaps instead of inventing facts.
Prompt
Role: You are helping me with [TASK TYPE].
Goal: [WHAT I WANT TO ACHIEVE AND WHY].
Context: [BACKGROUND, AUDIENCE, CONSTRAINTS].
Material: [PASTE TEXT OR DESCRIBE ATTACHED FILES].
Format: [LENGTH, HEADINGS, TABLE OR BULLETS].
Rules: Use only the material above for facts. If something is missing, say so instead of guessing. Mark any assumption.

Check the output, safety and limits

Verify

  • Check every factual claim against the material you supplied.

Limitations

  • Different models follow format instructions with different strictness.

Compare two models fairly

Run one identical prompt in two assistants and score the results.

Model-agnostic
Use it for
Deciding with your own work instead of benchmarks
Who it is for
Anyone choosing between tools
You will need
  • A real task
  • Two assistants
What you get back
Comparable answers plus each model's own list of uncertain claims.
Prompt
Task: [A REAL TASK FROM MY WORK].
Material: [PASTE THE SAME MATERIAL EVERY TIME].
Output: Give your answer, then list the three claims you are least sure of.

Check the output, safety and limits

Verify

  • Score on accuracy, usefulness, tone and effort to fix, using the same checklist for both.

Limitations

  • One task is a small sample. Repeat with two or three tasks.

Fact-check a response

Turn an answer into a list of claims you can verify.

Model-agnostic
Use it for
Spotting unsupported statements before you use them
Who it is for
Anyone relying on AI answers
You will need
  • The AI answer you want to check
What you get back
A numbered claim list sorted by how it needs to be verified.
Prompt
Below is a text. List every factual claim as a numbered list. For each, say whether it is (a) stated in the material I provided, (b) common knowledge, or (c) needs an external source. Do not add new facts.

Text: [PASTE ANSWER]

Check the output, safety and limits

Verify

  • The model can mislabel claims. Verify the (c) items yourself from primary sources.

Limitations

  • This reduces but does not remove the need to check.

More templates are in the prompt library.

For engineers

If you are building software, use the detailed engineering tools: the model catalogue, the model finder, comparisons and pricing.

Sources

Frequently asked questions

Which AI model is best for writing?

There is no single best model. For writing and editing, style and tone matter more than raw capability, so try two and compare on your own text. The page lists Claude, ChatGPT and Gemini as starting points, each with a trade-off. Treat any claim of an overall winner as marketing unless it names the task, date and method.

What is the difference between an AI model, an app, an API and an agent?

A model is the trained system that produces text, images or audio, with no interface of its own. An app is a product people use, such as a chat assistant. An API lets developers call a model from their own software, usually billed per use. An agent is a system where the model chooses steps and uses tools, which makes it more capable and riskier.

How do I choose between ChatGPT, Claude and Gemini?

Pick two candidates, run the same real task in both, and keep the one that needs less fixing. Rankings change monthly, and a model that leads on one kind of task can trail on another. Also compare privacy terms, usage limits, file support and integrations, and check each provider's current pages as of October 2026.

Can I trust AI benchmark scores?

Use them only to shortlist. A benchmark measures a fixed test set, not your task, documents or tolerance for errors. Results are often reported by the vendor, on selected tests and with different settings between models, and models can be tuned toward public tests. Decide with a small test on your own work.

Can I use the same prompt in different AI models?

Mostly yes. Prompts transfer better than most people expect when they are explicit about goal, material, format and rules. State the output format and length, paste only what is needed, and try both short and detailed instructions. Switch models when answers are repeatedly wrong in the same way, but switching rarely helps if the prompt is unclear.

Where to go next