Gratis

Every free LLM, behind one OpenAI-compatible endpoint.

Point your existing OpenAI code at it, pay nothing, and know which free model is actually worth using.

---

Free LLM tiers are everywhere and nobody can keep track of them. Six providers, hundreds of models, wildly uneven quality, and a catalogue that rots weekly as models get retired or moved behind a paid plan.

Gratis does two things about that. It proxies every free tier behind one OpenAI-compatible API, so your existing code keeps working and stops costing money. And it ranks them, so you know which free model is genuinely good instead of just genuinely free.

⚡ Quick start

git clone https://github.com/orkait/gratis && cd gratis
cp .env.local.example .env.local   # add whichever provider keys you have
python dev.py                      # backend :3460, UI :3470

dev.py is the whole local stack. It installs deps (bun install, uv sync), loads .env.local, starts both servers with interleaved logs, and shuts them down together on Ctrl+C. Every key is optional: providers you have no key for are skipped, and the rest still work.

🛠️ All dev.py commands

Command	What it does
python dev.py	backend :3460 + UI :3470
python dev.py backend	backend only
python dev.py ui	UI only
python dev.py test	vitest + pytest (test ui / test backend to narrow)
python dev.py lint	eslint
python dev.py build	next build
python dev.py docker up	docker compose up --build (--minimal for the 256MB backend-only image)
python dev.py docker down	stop the compose stack (logs, ps also work)
python dev.py deploy	backend → Railway, UI → Cloudflare Workers (deploy ui / deploy backend to narrow)

Flag	What it does
--env NAME	load .env.NAME instead of .env.local
--no-install	skip the dep install, for fast restarts
--backend-port N, --ui-port N	move a service off its default port
--force-port	kill whatever holds the port, even if it is not Gratis
--allow-dirty	deploy a tree that is dirty or unpushed

If a port is held by something outside this repo, dev.py names the process and stops rather than killing it.

🚀 Deploying

python dev.py deploy          # both
python dev.py deploy ui       # frontend only

The backend URL is compiled into the client bundle, so a build that forgets NEXT_PUBLIC_API_BASE_URL inlines the localhost dev default and the deployed site talks to a laptop. dev.py deploy removes the chance to get that wrong: it asks Railway for the service's public domain, builds against it, and refuses to upload a bundle that does not contain it or that still mentions localhost.

Step	Behaviour
Preflight	A dirty or unpushed tree does not reach prod
Railway	Waits for the build. A failed build leaves prod on the old version, and says so
Bundle check	Deploy is aborted if the compiled bundle points anywhere but the real backend
Cloudflare	wrangler prefers CLOUDFLARE_API_TOKEN over an OAuth login, so a stale one in your shell breaks every deploy. It is verified, and ignored when Cloudflare rejects it
Retries	Cloudflare's API answering 5xx is retried; a rejected credential is not
Verify	Backend /health and a 200 from the published route, after the deploy

Variable	Default
RAILWAY_SERVICE	gratis-backend
PROD_API_BASE_URL	asked of Railway
PROD_SITE_URL	the route in wrangler.jsonc

curl localhost:3460/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "gratis-auto", "messages": [{"role": "user", "content": "hi"}]}'

gratis-auto is the point of the whole thing: you do not name a model, and it picks the most capable free one available right now.

Formerly zero-cost-intelligent. That id still works and always will.

🎯 What you get

🔌 Drop-in OpenAI API	/v1/chat/completions, /v1/embeddings, /v1/models. Streaming included. Change the base URL, change nothing else.
🎭 Paid names route to free models	Send gpt-4o, get Llama 3.3 70B on Groq. Your code does not know the difference.
🧠 Pick-for-me routing	gratis-auto selects the strongest free model at request time.
📊 A real market, not a list	Models scored on benchmarks, Artificial Analysis intelligence and LMArena Elo, with task lenses for coding, agents, reasoning, price and speed.
🧹 Dead models remove themselves	Every listed model has actually answered. Nothing is hand-maintained.

🧹 The self-cleaning catalogue

Most "free LLM" lists are a lie within a month. Models get retired, moved behind a paid plan, or gated behind an agreement, and the list keeps advertising them.

Gratis keeps no list. A model earns its place by answering:

a call fails
   ├── 403 gate · 404 gone · "requires a paid plan" · "data policy" ──→ PERMANENT ──→ quarantined
   └── 429 rate limit · 5xx · timeout · anything unrecognised ────────→ TRANSIENT ──→ stays listed

plus a background sweep that verifies free models nobody has tried yet

A rate-limited free model is a working model, so only failures we positively recognise as permanent remove anything. Quarantines expire, which means the catalogue heals itself: accept a model agreement or upgrade a plan, and the model returns on its own. Paid models are never probed, because that would spend your money to learn something you did not ask for.

On a real run this caught Gemini 2.0 Flash variants that Google had quietly retired. A hardcoded list would still be recommending them.

🔑 Providers

Bring whatever keys you have. Six are routed directly, and OpenRouter fans out to dozens more upstream.

Provider	Env var	Free tier
Groq	GROQ_API_KEY	Yes, generous
Google AI Studio	GOOGLE_AISTUDIO_API_KEY	Yes, Flash models
Cerebras	CEREBRAS_API_KEY	Yes
Cloudflare Workers AI	CLOUDFLARE_API_KEY + CLOUDFLARE_ACCOUNT_ID	Yes, 10k neurons/day
Ollama Cloud	OLLAMA_API_KEY	Yes
OpenRouter	OPENROUTER_API_KEY	Free models need a data-policy opt-in

⚙️ All configuration

Backend

Variable	Default	What it does
LOCAL_API_KEY	(unset)	Require Authorization: Bearer on the proxy. Unset means open.
MODEL_ALIASES	built-in map	JSON object, merged over the defaults.
RANKINGS_CACHE_TTL	1800	Seconds to cache the assembled market.
AA_API_KEY	(unset)	Artificial Analysis intelligence index.
AVAILABILITY_ENABLED	1	Kill switch for the self-cleaning catalogue.
AVAILABILITY_UNAVAILABLE_TTL	86400	How long a quarantine lasts before the model is retried.
AVAILABILITY_PROBE_CONCURRENCY	3	Parallel probes in the background sweep.
AVAILABILITY_PROBE_BATCH	25	Models verified per sweep.
AVAILABILITY_PROBE_INTERVAL	900	Seconds between sweeps.
AVAILABILITY_DB_PATH	backend/availability.db	Where verdicts persist.
CLOUDFLARE_API_BASE	Cloudflare v4	Override, mainly for testing.

Frontend

Variable	Default	What it does
NEXT_PUBLIC_API_BASE_URL	http://localhost:3460	Where the UI finds the backend.
NEXT_PUBLIC_RANKINGS_STALE_MS	1800000	Client cache window, mirrors the backend TTL.
BACKEND_URL	falls back to the public URL	Server-side base for the chat route.

🔀 Built-in model aliases

Send an OpenAI or Anthropic model name and it is transparently routed to a free equivalent. Override any of these with MODEL_ALIASES.

You send	You get
gpt-4, gpt-4-turbo, gpt-4o	groq/llama-3.3-70b-versatile
gpt-4o-mini, gpt-3.5-turbo	groq/llama-3.1-8b-instant
claude-3-5-sonnet, claude-sonnet-4	groq/llama-3.3-70b-versatile
claude-3-5-haiku, claude-haiku-4	groq/llama-3.1-8b-instant
claude-3-opus	openrouter/anthropic/claude-3-opus

📡 API reference

Method	Route	Notes
POST	/v1/chat/completions	OpenAI-compatible. stream: true supported.
POST	/v1/embeddings	OpenAI-compatible.
GET	/v1/models	OpenAI-compatible listing. Unavailable models filtered out.
GET	/v1/models/{id}	Single model.
GET	/v1/rankings	The scored market. Powers the UI. Not an OpenAI route.
GET	/health	{"status": "ok"}

🏗️ How it fits together

┌──────────────────────┐        ┌────────────────────────┐        ┌────────────────┐
│  Next.js 16 UI       │        │  FastAPI backend       │        │  Providers     │
│                      │        │                        │        │                │
│  /models   market    │ ─────▶ │  /v1/rankings          │ ─────▶ │  Groq          │
│  /chats    chat      │ ─────▶ │  /v1/chat/completions  │        │  AI Studio     │
│                      │        │                        │        │  Cerebras      │
│  Zustand             │        │  scoring · arena       │        │  Cloudflare    │
│  TanStack Query      │        │  availability (sqlite) │        │  Ollama Cloud  │
│  TanStack Table      │        │                        │        │  OpenRouter    │
└──────────────────────┘        └────────────────────────┘        └────────────────┘
                                            │
                                            └── litellm routes every provider

🧪 Development

Command	What it does
python dev.py	Backend on :3460, UI on :3470, logs interleaved
bun run dev	UI only
bun run test	Frontend tests (Vitest)
cd backend && uv run pytest providers	Backend tests
bun run build	Production build
docker compose up	Full stack, UI on :3030, backend on :3040

Backend dependencies are managed with uv, frontend with bun.

📁 Layout

backend/
  main.py                  proxy, model routing, rankings assembly
  providers/
    availability.py        the self-cleaning catalogue
    scoring.py             composite model scoring
    intelligence.py        Artificial Analysis index
    arena.py               LMArena Elo overlay
    {groq,cerebras,cloudflare,openrouter,models}.py
src/
  app/models/              the market
  app/chats/               the chat surface
  components/app/          market table, filters, drawers
  lib/stores/              Zustand slices
  lib/query/               TanStack Query hooks

_{Built because free models are abundant, uneven, and impossible to keep track of.}