Key Takeaways
- The cheapest route is rarely the API price. — A free playground, a bundled subscription, or an aggregator’s free model usually beats the headline per-token rate for anything short of steady production volume.
- Free almost always means your prompts train the model. — Google AI Studio, most free API tiers, and several aggregators improve their products on free-tier traffic. The paid tier is partly a privacy purchase.
- Match the route to the usage profile, not the model. — Occasional use → free tier. Daily coding → a subscription harness. Spiky automation → metered API or an aggregator. Heavy steady volume → self-host the math.
- By free-tier generosity, Gemini leads. — Google AI Studio gives away the most usable frontier access of any major provider. DeepSeek is the cheapest paid token. Claude and GPT give the least away for free.
There are only five ways to reach a frontier model
Every route to GPT-5.5, Claude Opus 4.7, Gemini, or DeepSeek collapses into one of five shapes. The differences that matter are not the headline prices on the pricing page — they are the cap, the catch, and the usage profile each route is built for. Pick the route first; the cheapest provider inside it follows.
Free tier / quota
A provider gives away a capped amount of the model — a daily message count, a rate-limited API key, or a browser playground at no charge.
The catch: Limits are real (requests per minute, tokens per day) and on most free tiers your prompts can be used to improve the provider’s products.
Subscription / harness
A flat monthly fee bundles the model inside an app or coding harness — ChatGPT Plus, Claude Pro/Max, Gemini Advanced, Cursor, GitHub Copilot.
The catch: You pay whether or not you use it; heavy automated workloads hit usage caps the marketing pages do not advertise.
API aggregator
One key routes to many models through a single billing surface — OpenRouter and peers — often with a slice of free models attached.
The catch: A routing margin sits on top of the underlying token price; some routed providers log prompts unless you opt out.
Direct API
Metered pay-as-you-go billing straight from the model maker (Anthropic, OpenAI, Google, DeepSeek). You pay per token, nothing when idle.
The catch: Headline per-token rates explain less than half of real spend once context, caching, and retries are counted.
OSS self-host
Run an open-weights model (Llama, Qwen, DeepSeek, Mistral) on your own or rented GPUs. No per-token fee.
The catch: You trade the token bill for a GPU bill and the engineering time to operate inference; rarely cheaper below heavy, steady volume.
Match the route to how you actually use the model
The single most expensive mistake is paying for the wrong shape. A solo developer on a metered API key burns money on idle context; a 200-person team on individual free tiers loses days to rate limits and inherits a data-governance problem nobody signed off on. The route is a function of usage, not of which model has the best benchmark.
- Occasional / evaluating: a free tier or playground. Google AI Studio gives the most usable frontier access at $0.
- Daily, human-in-the-loop: a flat subscription or coding harness. You cannot consume enough to beat the cap.
- Automated, spiky, batch: the metered API or an aggregator. You pay only for what runs.
- Heavy, steady, privacy-bound: self-hosted open weights. The GPU bill replaces the token bill above a real volume threshold.
Pick your model or your route
Each guide below answers one question end to end — every route to a specific model, or every provider inside a specific route — with the limits, the data-usage terms, and the effective cost named. Pricing in this cluster was last verified 2026-05-27; the category moves monthly, so check the date on each page before you act on a number.
Free Gemini API
Google AI Studio, the Gemini API free tier, and where the no-cost route stops — with the data-usage catch spelled out.
Read guide → Per-modelFree DeepSeek API
Where DeepSeek is genuinely free, the off-peak discount window, and the aggregators that front it.
Read guide → Per-modelFree Claude API
Claude Pro vs Max vs Claude Code vs metered API vs Bedrock — which route to Opus 4.7 is actually cheapest for your usage.
Read guide → Per-modelFree GPT-5.5 Access
The free ChatGPT tier, where the API gives nothing away, and the cheapest paid paths to GPT-5.5.
Read guide → Per-routeBest Free LLM API Tiers
Every provider handing out free frontier-model calls — Groq, Cerebras, GitHub Models, OpenRouter free — scored on limits and the catch.
Read guide → Per-routeFree AI Coding Assistant Quotas
The real free-tier limits inside Cursor, GitHub Copilot, Windsurf, and Zed — and which model you actually get for $0.
Read guide → ComparisonSubscription vs API Cost
The breakeven math between a $20/$200 subscription and metered API billing, by usage profile.
Read guide → ComparisonLLM API Aggregators
OpenRouter and its peers on a cost lens — routing margin, free models, and when one key beats many.
Read guide →A note on how fast this rots
Free-tier limits, per-token prices, and even model names change on a monthly cadence — a provider doubles a free quota one week and removes a model the next. Every table in this cluster carries a "last verified" date and reads from a single source so the numbers stay consistent across pages. Treat any figure older than a month as a starting point to confirm, not a quote.
What AI models can I use for free?
Most frontier families have a free door. Google AI Studio gives the widest free access to Gemini, including the current Pro and Flash models, through a browser playground and a rate-limited API key. Claude, ChatGPT, and Gemini all have free chat apps that run a capable (usually not the top) model with daily caps. Groq, Cerebras, GitHub Models, and OpenRouter hand out free API calls to open-weights models such as Llama, Qwen, and DeepSeek. The cost you pay on the free tier is almost always data: your prompts can be used to improve the provider’s products.
Which LLM API is the cheapest?
Among the frontier makers, DeepSeek is consistently the cheapest paid token, and it runs an off-peak discount window that drops the rate further. For open-weights models, Groq and Cerebras price aggressively because their hardware is fast enough to serve more requests per chip. OpenRouter does not have its own cheapest price — it routes to whoever is cheapest and adds a small margin. The honest answer is that the cheapest API depends on the model you need: there is no single winner across all of them.
Is there a completely free AI API?
Yes, with caps. Google AI Studio, GitHub Models, Groq, Cerebras, and OpenRouter all offer API keys that cost nothing up to a rate limit (requests per minute, requests or tokens per day). They are genuinely free for prototyping, learning, and low-volume side projects. None are suitable for production traffic — the limits are designed to push real workloads onto a paid tier, and free-tier data is often used for training.
Can I run an AI model locally for free?
You can run open-weights models (Llama, Qwen, DeepSeek, Mistral, Gemma) locally with tools like Ollama or LM Studio at no per-token cost. "Free" here means no API bill, not no cost: you pay in hardware and the electricity to run it, and a laptop can only run the smaller, weaker variants. Local inference is the right call for privacy-sensitive work and steady high volume; it is rarely cheaper than a free or low-cost API for occasional use.
How do I choose between a subscription and the API?
Count how you use the model. If a human is in the loop most of the day (coding, writing, research) a flat subscription (ChatGPT Plus, Claude Pro/Max, a Cursor or Copilot seat) is almost always cheaper than metered tokens, because you cannot physically consume enough to exceed the cap. If the usage is automated, spiky, or measured in batch jobs, the metered API or an aggregator wins, because you pay only for what runs. Our subscription-vs-API breakdown does the breakeven math by profile.
Ready to Find the Right AI Tools?
Browse our data-driven rankings to find the best AI tools for your team.