AI subscriptions in 2026: ChatGPT, Claude, Gemini or local
AI subscriptions in 2026: what ChatGPT, Claude and Gemini cost, what free tiers include, and which open models run on your own laptop.
A basic paid chatbot subscription in 2026 still costs about $20 a month. That is what OpenAI charges for ChatGPT Plus, Anthropic for Claude Pro and Google for Google AI Pro. Top tiers run from $100 to $300. Free versions cover a couple dozen questions a day. Open models such as gpt-oss, Qwen and Gemma run on a laptop with 16 GB of memory, but they still trail the paid flagships.
Below is the breakdown: what each tier includes, where the limits hide, and when a model on your own computer or server is cheaper than a subscription.
How much do AI subscriptions cost in 2026
Over two years the market settled into three steps: a free tier to try things, a standard plan at $20 and a pro plan at $100-300. Prices in the table come from public pricing pages, taxes and regional discounts excluded.
| Service | Free | Base plan | Top plan |
|---|---|---|---|
| ChatGPT (OpenAI) | yes, limited | Plus, $20 | Pro, $200 |
| Claude (Anthropic) | yes, limited | Pro, $20 ($17 billed yearly) | Max, $100 or $200 |
| Gemini (Google) | yes | Google AI Pro, $19.99 | Google AI Ultra, $249.99 |
| Grok (xAI) | yes | SuperGrok, $30 | SuperGrok Heavy, $300 |
| Perplexity | yes | Pro, $20 | Max, $200 |
| Le Chat (Mistral) | yes | Pro, $14.99 | Team and Enterprise |
| DeepSeek | chat fully free | none | none, API billing only |
Between the free tier and Plus, OpenAI sells ChatGPT Go. It launched in India at roughly $4.50 and then expanded to other countries. The price depends on the region and in most places sits at $5-10 (estimate). Go gives more messages and file uploads than the free tier, but no access to the top models.
What the free tiers of ChatGPT, Claude and Gemini include
Free ChatGPT answers with the flagship model until the limit runs out, then switches to a lighter version until the counter resets. Claude Free offers the current Sonnet model with a message cap that resets every five hours. Gemini gives away the fast Flash model and limited access to Pro. DeepSeek charges nothing for chat at all.
The companies do not publish exact limits. They shift with server load, conversation length and attachment size. A 100-page PDF burns through a daily allowance faster than fifty short questions.
For someone who asks once a day to rewrite an email, translate a paragraph or explain a router settings error, free tiers are enough. Rotating between two or three services makes hitting the ceiling almost impossible.
What free usually lacks: unrestricted reasoning modes, Deep Research more than a few times a month, large file handling, agents that browse websites on their own, and priority at peak hours.
What $20 a month buys
The base paid plan sells three things: limits, bigger models and tools.
Limits. Plus and Claude Pro allow several times more messages than free. Anthropic promises at least five times the Free allowance, OpenAI does not name a multiplier.
Models. At $20 the reasoning modes unlock. They think for a minute or two before answering and are noticeably more accurate in math, code and document analysis.
Tools. Deep research with a ten-page report and sources, voice mode, image and video generation, projects with shared files, cloud drive and email connectors. Claude Pro includes Claude Code, an agent for working with code in the terminal. ChatGPT Plus includes the OpenAI equivalent, Codex.
The math is simple. $20 a month is about 66 cents a day. If the subscription saves at least 15 minutes of work a week, it pays for itself at most salaries in Europe and the US. In Russia the math is worse because of payment, covered below.
Choosing between the three big $20 services is now mostly a matter of taste. On public leaderboards like LMArena, models from OpenAI, Anthropic and Google trade places every couple of months. Claude is often picked for writing and code, ChatGPT for versatility and ecosystem, Gemini for its ties to Gmail, Docs and the 2 TB of cloud storage bundled in.
Who needs ChatGPT Pro at $200 and Claude Max
Top tiers appeared once it became clear that some users run models for hours without a break. These are programmers with agents, analysts with huge spreadsheets, researchers.
ChatGPT Pro at $200 gives near-unlimited access to the main models, Pro mode with the longest reasoning, extended agent limits and Codex. Claude Max comes in two versions: $100 with 5 times the Pro limits and $200 with 20 times the Pro limits. Google AI Ultra at $249.99 adds the most expensive Gemini modes, Veo video generation with a high cap and 30 TB of cloud storage.
The logic becomes clear next to API pricing. A developer who points Claude Code at a project all day can easily burn $1,000-3,000 a month in tokens at API rates (estimate based on user reports). A $200 Max plan at that load looks like a 5-15x discount. That is why in August 2025 Anthropic introduced weekly limits: according to the company, a small share of subscribers kept agents running around the clock.
For a reader who writes emails and sorts out server settings, the $200 tier means overpaying roughly tenfold. The difference from Plus only shows up under hours of daily load.
The catch: limits, data and geography
Limits are not counted in messages
A subscription does not mean a fixed number of requests. Claude counts usage in five-hour windows plus a weekly quota, ChatGPT in rolling windows per model. One request with an attached book can cost as much as a hundred short ones. Companies reserve the right to change limits without separate notice and use that right.
Chats may be used for training
Since September 2025 Anthropic has used chats from Free, Pro and Max users to train models unless the user opts out in settings. For those who agree, data retention grew to five years. OpenAI has had a similar toggle for a long time, switched on by default for personal plans. Business plans and the API at both companies are excluded from training by default.
In 2025 a US court in The New York Times v. OpenAI case ordered the company to keep even chats that users had deleted. The order was later narrowed, but the precedent showed that deleted in the interface does not always mean deleted on the server.
Russia, Iran and China are outside the map
OpenAI, Anthropic and Google do not officially operate in Russia, Iran, China and North Korea. In Russia the services are unavailable without changing the IP address, and Russian cards are not accepted. People pay with foreign cards or through resellers who add a 20-50% markup (market estimate) and do not guarantee the account will not be banned. In September 2025 Anthropic also banned sales to companies controlled from China, Russia, Iran and North Korea, even when they are registered elsewhere.
Hence the interest in open models. They are downloaded once, and after that no card or permission is needed.
Which AI models can you run yourself
Open models (open weights) are files you can download and run on your own computer offline. During 2025 the gap with paid flagships on tasks like translation, summarizing and simple code shrank to a few months.
The main families:
- gpt-oss from OpenAI. The company’s first open language models since 2019, released in August 2025 in two sizes, 20 and 120 billion parameters, Apache 2.0 license.
- Qwen from Alibaba. The widest lineup, from tiny to huge, strong in code and multilingual tasks.
- Gemma from Google. Compact models that work well on weak hardware.
- Llama from Meta. The biggest community and the largest number of fine-tuned versions.
- DeepSeek. Flagship-level models, but the full versions need hundreds of gigabytes of memory.
- Mistral. A European company with compact options.
New versions ship every few months, so model numbers as of November 2026 are worth checking in the Ollama and Hugging Face catalogs.
What hardware a local model needs
The key parameter is memory, not the processor. The whole model loads into video memory or, on Macs with M-series chips, into unified memory. Figures below are for compressed (quantized) 4-bit versions, which almost all home tools use.
| Model size | Memory needed | Example hardware | Speed, tokens/s (estimate) |
|---|---|---|---|
| 3-4B | 3-4 GB | any laptop with 8-16 GB RAM | 15-40 |
| 7-8B | 5-6 GB | 16 GB laptop, MacBook Air M1 or newer | 10-40 |
| 12-14B | 8-10 GB | 12 GB GPU, Mac with 16-24 GB | 15-40 |
| 20B (gpt-oss-20b) | about 14 GB | 16 GB GPU, Mac with 24 GB | 20-60 |
| 30-32B | 18-22 GB | RTX 3090 or 4090 with 24 GB, Mac with 32 GB | 15-40 |
| 70B | 40-45 GB | two 24 GB cards, Mac with 64 GB | 5-15 |
| 120B (gpt-oss-120b) | 60-65 GB | one 80 GB card, Mac with 96-128 GB | 10-30 |
For reference: 10 tokens per second is roughly reading-aloud speed, 30 and above feels like a normal chat. Models with 7-8 billion parameters handle summaries, translation and email drafts. In long reasoning and complex code they make mistakes noticeably more often than paid flagships.
How to run a model in ten minutes
The easiest path is Ollama, a free app for Windows, macOS and Linux. On Linux, install and launch look like this:
curl -fsSL https://ollama.com/install.sh | sh
ollama run gpt-oss:20bFor a weaker laptop, smaller models fit:
ollama run qwen3:8b
ollama run gemma3:4bThe first command downloads a 3-14 GB model file, after which the chat runs in the terminal without internet. For a familiar ChatGPT-style window, people install Open WebUI on top of Ollama:
docker run -d -p 3000:8080 --add-host=host.docker.internal:host-gateway -v open-webui:/app/backend/data --name open-webui ghcr.io/open-webui/open-webui:mainThe interface then opens in a browser at localhost:3000. An option with no command line at all is LM Studio, an app with a model catalog and a download button.
Running your own model on a VPS: when it pays off
Without a powerful computer at home, the model can live on a rented server. There are two options at very different prices.
| Option | What it handles | Monthly price (estimate) |
|---|---|---|
| VPS without GPU, 8 vCPU, 16 GB RAM | 3-8B models, 3-8 tokens/s | $20-50 |
| Hourly RTX 4090 rental | models up to 32B | $0.30-0.70 per hour, $250-500 running 24/7 |
| A100 or H100 80 GB rental | gpt-oss-120b, 70B models | $1.50-3 per hour, $1,100-2,200 running 24/7 |
| ChatGPT Plus or Claude Pro subscription | flagships | $20 |
The comparison is harsh. For one person a $20 subscription is almost always cheaper than a cloud GPU. A server without a GPU is cheap but slow: answering with a paragraph of text takes half a minute or more.
A self-hosted model wins in three cases. First, data that must not leave the premises, such as contracts, medical records or customer databases. Second, large volumes of repetitive work like labeling tens of thousands of reviews, where API bills would grow with every line. Third, a team of 10-20 people for whom one GPU server costs less than a dozen subscriptions.
The home option costs money too. An RTX 4090 draws about 450 W under load. Four hours a day add up to roughly 54 kWh a month, which is $5-15 depending on the electricity rate. The card itself costs $1,500-2,000 (estimate), equal to 6-8 years of a $20 subscription.
Subscription, API or local model: how they compare
| Criterion | Subscription | API | Local model |
|---|---|---|---|
| Entry cost | $0-20 | card and key, pay per token | computer with 16 GB memory |
| Quality | flagships | flagships | a step or two behind |
| Privacy | data stays with the company | with the company, not used for training by default | data never leaves the computer |
| Limits | floating windows | budget only | hardware only |
| Access from Russia | via resellers | via resellers | unrestricted |
| Works offline | no | no | yes |
Where AI money goes in 2026
For most people the split looks like this. Free tiers of two or three services cover everyday questions. One $20 subscription makes sense for anyone who works with text, documents or code every day. The $100-300 plans are built for those who keep agents running for hours, and for them it really is cheaper than the API.
A local model does not yet replace a subscription on quality, but it does not depend on limits, sanctions or data retention policies. A small model on a laptop needs 16 GB of memory and ten minutes of setup. Open models currently trail flagships by roughly six months to a year (estimate), and the gap keeps shrinking. By the end of 2027 the question of paying versus running it yourself may look very different.