🧠 AI & Tokens 101 — the primer

New here? This is the plain-English backstory behind every other guide on this site: what a "token" actually is, why so much AI compute is free, and how we got to a 2026 where you can run a studio-grade creative + dev stack for $0.

Contents

• What is a token?
• Why is any of this free?
• A short history (2022 → 2026)
• The four ways to pay $0
• The one thing that bites creators: licensing
• Glossary

What is a token?

A token is the unit language models read and write. It's roughly ¾ of a word — a chunk of characters the model treats as one step. "Unbelievable" might be 3 tokens (un + believ + able); a space, an emoji, or a CJK character can each be their own.

Two rules of thumb that explain almost all AI pricing:

• ~1 token ≈ 4 characters ≈ 0.75 words in English. 1,000 tokens ≈ 750 words.
• You pay for input tokens (your prompt + context) and output tokens (the reply), usually at different rates, per million tokens (often written $/Mtok).

So a "128k context window" means the model can hold ~96,000 words of prompt at once, and a bill of "$0.50 / 1M input" means a 10,000-token prompt costs half a cent. Everything else — caching, batching, the "100M tokens for $1" tricks — is just moving where those tokens are counted and how often they're recomputed. (The Token-Maxxing guide is the deep dive.)

Why is any of this free?

It feels too good to be true, so it's worth knowing the actual incentives. "Free" AI compute exists for five honest reasons:

1. Land-grab. Cloud and model vendors give away credits to win developers before the market settles. Free tiers are a customer-acquisition cost, not charity.
2. Open weights. Labs (Meta, Mistral, Qwen, DeepSeek, Black Forest Labs, Stability, Tencent) release model weights publicly. Once weights are open, anyone can run them on free or cheap hardware — the marginal cost of inference collapses.
3. Spare capacity. Research programs (Google's TPU Research Cloud, Kaggle, Colab) and serverless platforms hand out idle GPU/TPU time to seed an ecosystem.
4. Loss-leader pricing. Aggressive players — especially Chinese labs in 2025–26 — price tokens near cost (or below) and lean on prompt caching to make repeat work nearly free.
5. Your own hardware. A modern laptop — especially Apple Silicon with unified memory — can run real models locally. Local inference is unmetered and private: the truest $0.

The catch is never the price — it's quotas, idle-sleep, data-sharing, and licensing. This site exists to map exactly those.

A short history (2022 → 2026)

The throughline: every year, more of the stack falls to $0 — first tokens, then audio, then finishing, then the backend. What stays scarce is trustworthy information about which free option is actually free, actually good, and actually safe to monetize.

The four ways to pay $0

Every resource on this site is one of these. Mix them with the stacking playbook:

1. Free tiers — recurring monthly allowances (Gemini API, Groq, Cloudflare, Neon). Best for steady low-volume use; watch the quota and the idle-sleep.
2. Trial credits — one-time fuel ($300 GCP, $200 Azure, $25 Predibase). Best for a burst of heavy work; they expire.
3. Open weights, self-hosted — run the model yourself on free GPU (Kaggle/Colab) or local (Ollama, ComfyUI). Unmetered, private, and the only truly unlimited option.
4. Your own legitimate keys, orchestrated — route many of your free keys through one hub (LiteLLM/OpenRouter) with caching and fallback so you degrade gracefully instead of hitting a wall.

The one thing that bites creators: licensing

If you monetize anything — a YouTube video, an app, a product — the licence on a "free" model matters more than its quality. A huge share of popular open models are research / non-commercial only, and several are silently bundled into the GUIs everyone uses.

• ✅ Safe to monetize: MIT, BSD, Apache-2.0 — and even GPL (copyleft only governs redistributing the software, never your render/output).
• ⚠️ Not without a paid licence: CC-BY-NC, "research only," and bespoke non-commercial licences (CodeFormer, SUPIR, BRIA RMBG, SeamlessM4T/NLLB, Suno free, Coqui XTTS-v2, FLUX.1-dev, and more).

Each guide carries a licensing-landmines section for its domain. When in doubt, prefer permissive weights and read the actual LICENSE file — quality you can't legally use is worth zero.

Glossary

• Token — ~¾ of a word; the unit models read/write and bill by.
• Context window — how many tokens a model can consider at once (e.g., 128k ≈ 96k words).
• Open weights — the trained model parameters are public; you can run/fine-tune them.
• Free tier — a recurring no-cost usage allowance from a vendor.
• Prompt caching — reusing a fixed prompt prefix at a deep discount (often ~90% off).
• Quantization — shrinking a model (e.g., to 4-bit) so it runs on smaller/free GPUs.
• LoRA / QLoRA — cheap fine-tuning that trains a small adapter, not the whole model.
• Inference vs training — running a model vs teaching it; inference is what you mostly pay for.
• Scale-to-zero / idle-sleep — a free service that pauses when unused (cold starts on wake).
• MCP — Model Context Protocol; the open standard "USB-C" that plugs tools into AI agents.
• LUFS — loudness units; streaming platforms normalize to a target (see Post-Production).

Ready to go deep? Start with Token-Maxxing for LLMs, or the Free Compute list for CPUs/GPUs. New tools and prices land in The Wire on the home page, refreshed automatically.

📝 Spotted a stale quota or a license that changed? This guide is open source — edit it on GitHub.