Match the ChatGPT alternative to the constraint you actually have

The answer first

Four ChatGPT alternatives, on the question that actually changes the choice

Prices and plan matrices are from each vendor's own pages, read on 6 October 2026. This is not a test of which model writes a better email.

Why the table refuses to name a winner

People who ask for a ChatGPT alternative are usually stuck on one of four things: the bill, a limit that hits mid-task, a missing strength, or a rule about where the prompt may go. Those are different decisions, and a numbered list of twenty products treats them as one. ChatGPT Plus is $20 a month on OpenAI's help page, billed monthly, with no annual option. Claude Pro is the same $20, or $17 a month when $200 is paid up front. Mistral Pro is $14.99. If the only problem is the subscription, the cheaper plan is a real answer. If the problem is a contract, the cheaper plan is irrelevant.

The free tiers are where listicles copy each other and get it wrong. Anthropic's pricing matrix includes Sonnet and Haiku on the free plan and marks Opus as off until Pro. Pro also unlocks Claude Code, and Anthropic describes Pro as at least five times the free usage per five-hour window. There is no fixed message count on either plan, because the allowance depends on length, model and features. OpenAI's Plus article says the free product exists and that Plus buys broader model access, and it also says Plus may still cap messages when demand is high. A page that promises unlimited free ChatGPT, or the same Claude model on every plan, is ahead of the documents.

Privacy splits into three controls, and only one of them is 'switch the chatbot'. ChatGPT Business and Enterprise, and Claude Team and Enterprise, state that they do not train on your business content by default. ChatGPT Plus and Claude Free, Pro and Max are opt-out. Training is about whether your chats improve their models. Residency is about which region processes them: OpenAI lists Enterprise residency in the US, the EU, the UK, Japan, Canada, Korea, Singapore, India, Australia and the UAE, and the same row is off for Business. Neither control means the prompt stayed in your building. DeepSeek's privacy policy says it collects, processes and stores personal data in the People's Republic of China. That is a fine API for a public benchmark and a bad one for a customer file. Mistral's help centre says data is hosted in the EU by default, with a US endpoint only if you ask for it, and with a written caveat that some features can still transfer data outside the EU.

The specialist case is narrower than the ads. Perplexity's own API docs show a citations list and search results with URLs on the response, which is the property you want when a report has to be checked. Claude's current Sonnet, Opus and the models around them advertise a 1 million token context, and from Claude 4.6 onward Anthropic bills that full window at the standard per-token rate. Sonnet 5.5 is $2 per million input tokens and $10 per million output tokens. DeepSeek V4 Pro, on a cache miss in off-peak hours, is $0.66 and $1.98 on the same units. The gap is real for anyone calling an API all day. It does not tell a person which chat window to open for a letter. Gemini is the remaining case, and it belongs to teams who already work in Google's apps: the free Gemini app includes 3.6 Flash, varying access to 3.1 Pro, and Deep Research, with a daily research limit Google does not publish as a fixed count. Google AI Pro raises ordinary usage limits to four times the free allowance.

Give the same real task to two assistants and compare the answers line by line. The free tiers make that test free, and it uses your work instead of a vendor benchmark. It will not answer the hosting question. A better letter from Claude still left your network. ChatGPT Business is the purchase when the gap is a workspace: standard seats are $20 per user per month billed annually, from two seats, with no training on business data by default, for teams of 2 to 200. Enterprise is the purchase when you also need residency in one of the ten regions OpenAI names, including the EU. The Business column on that same page leaves residency off.

Once the work is an application rather than a chat window, the unit changes from a seat to a million tokens. Claude Sonnet 5.5 is $2 in and $10 out. DeepSeek V4 Pro, on a cache miss in off-peak hours, is $0.66 and $1.98. Mistral's pricing page gives Mistral Large at $0.50 in and $1.50 out. Those are list prices for an API call. They are not a reason to put a customer contract on the cheapest host. Claude Enterprise adds $20 per seat per month on top of usage billed at API rates, and that plan is annual. A comparison that stops at the $20 chat subscription never sees this bill.

Same task, two hosts

The payload does not change. The host is the decision.

Illustrative only. Neither call was run for this article. The model name is a placeholder because the point is the URL, and inventing a current model id would age in a week.

# 1) Hosted. The clause leaves your network.
POST https://api.openai.com/v1/chat/completions
Authorization: Bearer $PROVIDER_KEY
{"model":"<model your key can call>","messages":[
  {"role":"user","content":"<contract clause>"}]}

# 2) Self-hosted. Same JSON, a host you operate.
POST https://llm.internal.example/v1/chat/completions
Authorization: Bearer $GATEWAY_KEY
{"model":"<model you serve>","messages":[
  {"role":"user","content":"<contract clause>"}]}

Swap api.openai.com for Anthropic, Google or Mistral and the shape is unchanged: a vendor still holds the clause. Only the second host can log it, budget it and keep it inside the estate.

One real strength each

Three options a self-hosted cluster does not beat, and the question to ask before you take them

• Gemini

  It is already inside the Google account the company works in. The free app includes Deep Research and access to 3.6 Flash, and there is no GPU cluster to staff on day one.

  Is this prompt allowed to be processed by Google, including a Deep Research run that can read Gmail or Drive if someone connects them?

• Mistral

  A European vendor that hosts data in the EU by default, with Pro at $14.99 a month and private deployments on Enterprise, before anyone rents a GPU.

  Which features still transfer data outside the EU, and will the Enterprise plan turn those features off for our organisation?

• DeepSeek

  The API price. V4 Pro, cache miss, off-peak, is $0.66 per million input tokens and $1.98 per million output tokens, against $2 and $10 for Claude Sonnet 5.5.

  Is processing and storage in China acceptable for this data, and is that answer in writing from the person who owns the data?

Which row is yours

The condition that picks each option

Sources: OpenAI: What is ChatGPT Plus? · OpenAI: Business pricing · Anthropic: Claude plans and pricing · Anthropic: API pricing · Anthropic: Claude Sonnet 5.5 · Perplexity: Sonar and Agent API · Mistral: pricing · Mistral: where data is stored · DeepSeek: API pricing · DeepSeek: privacy policy · Google: Gemini app · Google: Gemini usage limits. Read 6 October 2026. Vendor documentation changes; verify against the current release.

Where to go once the constraint is clear

• Sovereign private AI — the deployment when the prompt cannot leave, and the page this post is a path into
• Open WebUI — the self-hosted chat interface in front of a model you run
• Open-weight model hosting — which model to put behind that interface, and what it costs to serve
• Claude in the EU — when the choice is Claude and the remaining question is the region