There are three ways to use today’s top AI models, and the cheapest one depends on how much you use them. The free tiers in consumer apps cost nothing and are enough for most individuals. A monthly subscription makes sense when you hit the free limits or need a flagship model inside the app. Paying per token through an API is usually cheapest for apps, automations and batch work, and it lets you mix very cheap models with expensive ones. Self-hosting open-weight models is a fourth route, but it only pays off in specific cases. All prices below were recorded on 14 and 15 October 2026.
Quick answer
- Occasional or everyday personal use: stay on free plans such as ChatGPT, Claude, Gemini or DeepSeek.
- Heavy daily chat or flagship models in the app: one subscription, starting around $7.99 to $8 per month.
- Apps, scripts and automation: the API, starting with a cheap model such as GPT-6 Luna, Claude Haiku 5.5, Gemini 3.5 Flash-Lite or DeepSeek-V4.1-Flash.
- Private data, offline use or fine-tuning: an open-weight model such as gpt-oss-120b, Gemma 4 or Mistral Small 4 on your own hardware.
Option 1: the free tiers in consumer apps
Free plans are more generous than many people expect. ChatGPT Free includes unlimited text chats with GPT-6, subject to abuse guardrails. Claude Free runs the Sonnet and Haiku models with web search, file creation and code execution. Gemini’s free tier uses the lighter Flash-Lite model but adds Deep Research and Canvas. DeepSeek offers DeepSeek-V4-Pro and DeepSeek-V4.1-Flash in its app and on the web at no charge, and Alibaba describes its Qwen chat app as free to use. If your needs are covered by these, the cheapest option is simply to stay on them and switch between apps when one hits a limit.
Option 2: a monthly subscription
A subscription buys a flat monthly cost, higher limits and access to stronger models inside a familiar app. The entry tiers are close in price: Google AI Plus starts at $7.99/mo, ChatGPT Go and Grok Plus are $8/month, and Claude Pro is $20/month billed monthly. Some flagship models are subscription-only in the consumer apps: Claude Opus 5.5 is not on Claude Free, and Gemini 4 Argon is rolling out to Google AI Pro and Ultra subscribers. Subscriptions still have usage limits, so a flat fee does not mean unlimited use.
Option 3: paying per token through the API
An API bills you for tokens, the small chunks of text a model reads and writes. Input (your prompt and documents) is cheaper than output (the model’s answer), and repeated input can be billed at a lower cached rate. For light or bursty use, a few cents of tokens can replace a subscription; for heavy use, the bill grows with every request. The table lists prices per 1M tokens exactly as recorded.
| Model | Provider | Input per 1M tokens | Output per 1M tokens | Notes |
|---|---|---|---|---|
| GPT-6 Astra | OpenAI | $10 | $50 | Flagship; $1 cached input |
| GPT-6.1 Sol | OpenAI | $2 | $10 | $0.10 cached input |
| GPT-6 Luna | OpenAI | $0.10 | $0.50 | Cheap, fast tier |
| Claude Fable 5.1 | Anthropic | $10 | $50 | Slowest and priciest Claude |
| Claude Opus 5.5 | Anthropic | $4 | $20 | Anthropic’s recommended default |
| Claude Sonnet 5.5 | Anthropic | $2 | $10 | Everyday coding and documents |
| Claude Haiku 5.5 | Anthropic | From $0.10 | From $0.50 | $0.50 / $2.50 for prompts above 100k tokens |
| Gemini 4 Argon | $2.00 | $10.00 | Through Nov 30, 2026; $4.00 / $20.00 from Dec 1, 2026; no free API tier | |
| Gemini 3.8 Flash | $0.75 | $3.75 | Through Dec 31, 2026; $1.50 / $7.50 from Jan 1, 2027; free tier | |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | Free tier available | |
| Grok 4.7 | xAI | $2.00 | $6.00 | Prompts under 200k tokens |
| Grok 4.3 | xAI | $1.25 | $2.50 | Prompts under 200k tokens |
| Muse Spark 1.3 | Meta | $1.25 | $4.25 | Contributor tier $0.10 / $0.20, data used by Meta |
| Mistral Large 4 | Mistral AI | $1.36 | $4.18 | List price; sale price $0.68 / $2.09 |
| Mistral Small 4 | Mistral AI | $0.15 | $0.60 | Apache 2.0 open weights |
| DeepSeek-V4-Pro | DeepSeek | $0.66 off-peak / $1.32 peak | $1.98 off-peak / $3.96 peak | Cache-miss input |
| DeepSeek-V4.1-Flash | DeepSeek | $0.15 off-peak / $0.3 peak | $0.6 off-peak / $1.2 peak | Cache-miss input |
| Qwen3.8-Max | Alibaba | $2 | $6 | International (Singapore) |
| Qwen3.8-Flash | Alibaba | $0.15 | $0.47 | International (Singapore) |
Cheap and fast models vs flagships
The gap between tiers is large. GPT-6 Luna costs $0.10 input and $0.50 output, while GPT-6 Astra costs $10 and $50: one hundred times more per token. Claude Haiku 5.5 matches Luna’s starting price but rises 5x for prompts over 100k tokens. Gemini 3.8 Flash and Gemini 3.5 Flash-Lite include a free API tier, though Google notes that free-tier content may be used to improve its products. DeepSeek-V4.1-Flash is among the cheapest options and is cheaper still in off-peak hours.
Cheap models are built for classification, extraction, routing, summarisation and simple agent steps. Flagships earn their price on long, hard reasoning and coding. A practical pattern is to send most requests to a cheap model and pass only the hard steps to a mid-tier model such as GPT-6.1 Sol or Claude Sonnet 5.5, both at $2 input and $10 output. Our DeepSeek-V4-Pro vs GPT-6.1 Sol comparison shows how a low-cost model stacks up against a mid-priced one, and the benchmarks page helps check quality before you switch.
Watch the fine print. Gemini 4 Argon’s introductory price doubles on 1 December 2026, and Gemini 3.8 Flash’s on 1 January 2027. Grok 4.7 charges more for prompts of 200k tokens or more, applied to every token in the request. Mistral Large 4’s lower price is a sale price, and Meta’s cheap Muse Spark 1.3 contributor tier lets Meta use your data.
Open-weight models: when self-hosting makes sense
Open-weight models can be downloaded and run on your own hardware. The model itself is free, but GPUs, electricity and maintenance are not. The main options:
- gpt-oss-120b (OpenAI, Apache 2.0): a reasoning model that runs on a single 80GB GPU. Text only, with a knowledge cutoff of June 2024.
- Gemma 4 (Google, Apache 2.0): sizes from phone-class E2B up to 31B for a single-GPU server.
- Mistral Small 4 (Apache 2.0): also cheap on Mistral’s API at $0.15 input and $0.60 output.
- DeepSeek V4: open weights, but DeepSeek-V4.1-Flash has a 552B backbone, too large for most teams to host, and the published V4-Pro weights are the April preview.
- Qwen: Qwen3.8-27B runs on a single GPU; Qwen3.8-Max weights are about 4.9 TB and use a custom license.
Self-hosting makes sense when data must not leave your network, when you need to work offline, when you want to fine-tune, or when steady high volume makes owning hardware cheaper than paying per token. It rarely makes sense for occasional use, because hosted cheap models already cost cents per million tokens. Read each license: Apache 2.0, modified MIT and custom community licenses carry different conditions.
Checklist before you pay
- Have you tested two or three free plans on your real tasks?
- Do you know roughly how many tokens a typical request uses, input and output?
- Have you tried the cheapest model first and moved up only where quality falls short?
- Do you know when introductory prices end?
- Does any cheap tier use your data for training, and is that acceptable?
- If self-hosting, have you counted hardware, power and staff time, not just the free download?
Bottom line
For most people the cheapest setup is free apps, plus one subscription when limits start to bite. For developers it is the API with a cheap model by default and a flagship only for hard steps. Self-hosting is for privacy, offline work or high, steady volume.
