Six measured failure modes when a free AI API meets an agent harness: declared max_tokens billing, a 413 reported as context overflow, and why the same config is fatal on one provider.
Groq's free tier advertises 131k context. We measured 7,500 usable tokens: the 8,000 TPM budget is the real limit, and it counts the max_tokens you request.
We plugged four free AI models into a real agent harness and gave them real jobs. Nemotron was fastest at debugging but silently did nothing on file tasks; Kimi K3 was most reliable.
We benchmarked 12 free AI APIs in August 2026: half now require a card. Groq hit ~521 tok/s; OpenRouter ran NVIDIA's own model faster than NVIDIA.
Gemini Live's preview models are the only speech-to-speech API listed as free. OpenAI Realtime has none; Hume gives 5 min/month. Limits and code inside.
OpenAI Moderation is still free: 5,000 checks/day, text and images. Plus Groq free safeguard models, and why Perspective API shuts down Dec 31, 2026.
Crawl4AI is the free, open-source (Apache 2.0) web crawler that turns any site into clean, LLM-ready Markdown — self-hosted, no API key, no quota.
Qwen's free OAuth tier ended April 2026, but Qwen3-Coder is Apache 2.0. Run the API free via OpenRouter, self-hosting, or Puter.js — with runnable code.
Six free ways to OCR PDFs and scans: OCR.space, Cloudflare, Gemini, Cloud Vision, Mistral OCR 4 and self-hosted Docling. Real quotas, license traps, code.
Five ways to generate AI images from text at bash in 2026 — Pollinations, Cloudflare Workers AI, Together AI, Hugging Face & Gemini Nano Banana compared.
