A guide to current AI models

SingulX · Updated 22 September 2026

Our monthly reviews look at new releases, how they compare with existing options and whether they give you a reason to switch.

The full comparison

This table contains 208 model and family entries. Scroll inside the table to see the full list, or use the filters to narrow it. “Show all models” clears the filters while keeping the table scrollable.

Updated 22 September 2026. Models and named variants are listed individually where available; some entries cover a family.

Reasoning and general work

For analysing a proposal, checking an argument or working through a difficult problem, start here. The published results favour different models on different tasks.

A founder checking a business plan, a researcher comparing papers and a student learning a concept all need answers they can verify. The best score alone does not tell them how much checking remains.

Writing and judgement

A useful writing model should help you say what you mean, in language your reader understands. Fluent sentences are only part of the job: a model can produce a polished draft while adding facts you never supplied, making a promise you did not intend or removing an important qualification.

What does a good result look like?

For an article or report, give the model your source material, the audience and the point you want to make. Then check whether a reader can follow the argument without already knowing the subject. Look for unexplained terms, unsupported claims and repeated paragraphs. If the model adds a figure, quotation or example, check where it came from before keeping it.

For a customer email, give it the customer’s actual question and the facts you are allowed to communicate. A good reply answers the question directly, sounds appropriate for the relationship and explains what happens next. Check that it has not invented a refund, delivery date or other commitment. Friendly wording does not make an inaccurate answer useful.

For translation, decide who will read the text and what it is for. A good translation preserves the intended meaning, level of formality and specialist terms. Someone fluent in the target language and familiar with the subject should review important work. A sentence can read smoothly while changing a condition or losing the intended tone.

To compare models, give them the same brief and a short sample of writing you like. Count the corrections needed to make each result usable. The model that saves you the most editing on your actual work may differ from the one that leads a general benchmark.

Small and self-hosted models

These models give you more control over where your work runs. Some fit on personal devices; others need substantial servers. Downloadable does not mean small.

For private notes, classroom tools or an internal company assistant, hardware and data flow may decide the choice before a benchmark does.

Coding and automation

Completing a line of code, writing a function and repairing a project are different jobs. The table includes small local models as well as coding agents.

A learner needs clear explanations. A developer needs correct changes and tests. Someone building an app without coding experience also needs a tool that can show, run and explain the result.

Images and design

Compare image creation and image editing separately. A striking picture is not enough when a logo, product or sentence must remain exact.

Illustrators may prioritise style. Retailers need product accuracy. Designers may need editable vectors. These are different reasons to choose a model.

Video

The main differences are control over references and motion, whether you can edit existing footage, and whether the model also produces sound.

A social clip, a teaching demonstration and footage for a film have different requirements. Count the clips you can actually use, not just the price of one generation.

Transcription

These models turn speech into text. Recorded-audio scores and live conversation performance are separate comparisons.

For interviews, meetings, captions and accessibility, check names, speaker changes and the languages actually spoken.

Voice and live conversation

These models speak or take part in spoken conversations. A pleasant voice and an accurate transcript are different things.

Narration needs pronunciation and expression. A live assistant also needs to respond at the right time and handle interruptions.

Music and sound

Compare control over the song, vocals, editing and export. A short demo cannot establish that one system will suit every musician or production.

A musician developing an idea and a business commissioning background audio also have different licensing needs.

Searching your documents

A document assistant first has to find the right passage. Some models help find it, some reorder the results, and others read text from scanned pages.

For a researcher, this means finding the right paper. For a team, it may mean searching internal files. For an individual, it could mean finding an answer in years of personal documents.

Other specialist uses

Some models are components for safety, forecasting, vision or robotics. They belong in the landscape, but a chatbot league table does not describe them.

Use a test that measures the actual job: forecasts, extracted values, detected risks or physical actions.

What happens to your data?

Several chat apps let you turn model training off. That is a useful control, and you do not need to use an API to have it. Check the setting in the account you actually use, including a paid personal account.

Turning training off does not necessarily delete your conversations or prevent safety checks. Training, chat history, memory and connected apps can have separate controls. For example, an assistant may remember your preferences without using that conversation to train its underlying model.

The controls in everyday chat apps

These are the providers' published rules, checked on 22 September 2026. They describe commitments and exceptions, rather than independent verification of what happens inside each service.

Chinese cloud services need the same scrutiny

DeepSeek is one example, not a proxy for every Chinese developer. Some services store information in mainland China; others operate international products with different arrangements. An international website or a Singapore company address does not, by itself, establish where every part of a service processes your files.

For some services, the documents reviewed do not clearly explain how to stop your conversations being used for training. Treat that as an unanswered question. Check the app's current settings or ask its support team before uploading private material. The same question remains to be checked for MiMo, Yuanbao, Yuewen and individual services hosting Wan models. Their inclusion in a model comparison does not establish how those services handle your data.

Can you be certain with a cloud service?

You cannot be completely certain simply by switching a setting off or reading a privacy policy. When you use a cloud service, your material is processed on systems you do not control. You rely on the provider to apply its stated rules and protect those systems. Its policy may also allow limited staff access, service providers or legally required disclosure. Z.ai's policy, for example, describes these forms of access and explicitly recognises that internet transmission cannot be completely secure. Z.ai privacy policy.

Turning training off still has value: the provider is committing to exclude covered content from model training. But that commitment does not say that every copy is immediately deleted or that nobody can access it for another permitted purpose. Contracts and independent security assessments can give you stronger grounds for confidence, without making a service risk-free. Uncertainty is not evidence that a provider is secretly ignoring its promises.

Think about the consequence of disclosure. A public product description is different from a customer database, a private family conversation or an unreleased invention. For sensitive material, use a service whose terms meet your needs, remove identifying details where possible, or keep the work in a properly secured local setup. Keeping files local gives you more control, but the security of your device still matters.

Geography matters across providers. OpenAI's consumer policy describes processing in the US and other jurisdictions. Google's policy describes servers around the world. Mistral describes EU hosting with possible transfers for particular features. A familiar Western brand does not automatically mean your information stays in your country. OpenAI privacy policy, Google privacy policy, Mistral locations.

For an ordinary draft, start by checking your training and history settings. For client files, unpublished plans or internal records, establish which account and service your organisation permits and what its contract covers. Paying for a personal subscription does not automatically give you the terms of a business workspace.

If you need the material to stay on your own equipment, a local model can be useful. Check the entire setup: a local model connected to a cloud search tool, remote file store or external service can still send information out. Equally, running a downloaded Chinese model locally does not automatically send your prompts to its original developer.

Teams, APIs and connected tools

Business workspaces and APIs can offer different defaults, retention options and contractual protections. Check the named product and plan. For example, ChatGPT business offerings exclude business data from training by default, while Mistral documents separate chat and API opt-out switches. There is no sound rule that every API is private or every paid chat account has the same protections. OpenAI account controls, Mistral product controls.

The same care applies when an assistant searches your files or acts on your behalf. Check which folders it can read, which services receive the material and whether it can edit or send anything. The useful privacy decision is about the whole service you are using, not just the model name in a menu.