Stack guides / AI APIs

AI model APIs and vector databases

For apps with AI features. Call models from your server only, cap usage per user, and keep embeddings next to your data when you can.

13 options. Free tiers, catches and prices checked by hand on 2026-10-27; each one links to its sources.

AI API

Cloudflare Workers AIService

Serverless inference for open models on Cloudflare’s GPU network.

Best for: Embeddings and small LLM tasks inside Workers apps. Skip it if: You need frontier-model quality.

Official Cloudflare MCP.

10,000 Neurons/day

  • 10,000 Neurons/day (resets 00:00 UTC)
  • Then $0.011 per 1,000 Neurons
  • gpt-oss-20b: $0.20 / $0.30 per 1M tokens

Workers Free accounts must upgrade to go past the daily allocation.

Workers Paid $5/mo plus Neuron usage

Checked 2026-10-27: developers.cloudflare.com

Google Gemini APIService

Google’s API for Gemini and Gemma models, with a free tier.

Best for: Free prototyping with fast Flash models. Skip it if: Users send personal data while you’re on the free tier.

Gemini API docs publish llms.txt.

Free tier

  • Free: Gemini 2.5 Flash, Flash-Lite and Pro, 3.x Flash and Flash-Lite, Gemma 4
  • Paid-only: 3.1 Pro Preview, Gemini 4 Argon
  • Per-model RPM/RPD shown only in AI Studio; daily limits reset at midnight Pacific

Free-tier data is used to improve Google products and may be human-reviewed; apps for EEA, Swiss or UK users must use Paid Services.

Gemini 2.5 Flash-Lite $0.10 input / $0.40 output per 1M tokens

Checked 2026-10-27: ai.google.dev, ai.google.dev, ai.google.dev

GroqService

A fast inference API for open-weight models on Groq’s own chips.

Best for: Low-latency chat and transcription with open models. Skip it if: You need high free throughput.

OpenAI-compatible API; llms.txt.

Free

  • gpt-oss-20b/120b: 30 requests/min, 1K/day
  • 8K tokens/min, 200K tokens/day
  • whisper-large-v3: 20 requests/min, 2K/day

Token caps are tight, and Llama 3.x models are now 'contact sales’.

Developer pay-as-you-go: gpt-oss-20b $0.075 / $0.30 per 1M tokens

Checked 2026-10-27: console.groq.com, console.groq.com

OllamaMIT

An open-source tool for running open-weight LLMs on your own machine.

Best for: Local development and private prototypes. Skip it if: You need hosted production inference without managing GPUs.

OpenAI-compatible /v1 chat, responses and embeddings endpoints; llms.txt.

Local (free, unlimited)

  • Local models: unlimited
  • Cloud models: pay per token, no free allowance
  • 3 concurrent cloud requests

Free only on your own hardware.

Cloud per token, or Pro $20/mo with $60 credits

Checked 2026-10-27: ollama.com, docs.ollama.com

OpenRouterService

A single OpenAI-compatible API that routes to hundreds of models.

Best for: Trying many models behind one key. Skip it if: You send sensitive data through free models.

Official remote MCP (mcp.openrouter.ai/mcp); llms.txt; OpenAI-compatible.

Free (':free' models)

  • 25+ free models
  • 20 requests/min
  • 50 requests/day; 1,000/day after buying $10+ credits

Some free endpoints log and may train on prompts (controlled by a separate free-model privacy setting).

Provider prices plus 5.5% fee on credit purchases

Checked 2026-10-27: openrouter.ai, openrouter.ai, openrouter.ai

Claude APIService

Anthropic’s pay-per-token API for Claude models.

Best for: App features that need careful reasoning, writing or tool use; Haiku 5.5 is cheap enough for high-volume tasks. Skip it if: You need a free tier to prototype at $0.

Official TypeScript and Python SDKs; Claude Code and the Agent SDK use the same API.

No free plan

Pay per use: Haiku 5.5 $0.10 in / $0.50 out per 1M tokens; Sonnet 5.5 $2 / $10; Opus 5.5 $4 / $20

Checked 2026-10-27: platform.claude.com

OpenAI APIService

OpenAI’s pay-per-token API for GPT models, embeddings and moderation.

Best for: Cheap classification, extraction and short chat. Skip it if: You need free prototyping.

Its API format is supported almost everywhere; docs publish llms.txt.

No free plan

Usage-priced; cheapest is gpt-5-nano

Checked 2026-10-27: developers.openai.com, developers.openai.com

Vector DB

ChromaApache-2.0

An open-source embedding database, also offered as serverless Chroma Cloud.

Best for: Local RAG prototyping in Python or JS. Skip it if: You want a fixed monthly bill.

chroma-mcp server; llms.txt.

Open source, or Cloud Starter ($5 credits/month)

  • $5 Cloud credits/month
  • 10 databases
  • $2.50 per GiB written, $0.33 per GiB-month stored

$5 covers only about 2 GiB of writes a month.

Usage beyond credits; Team $250/mo

Checked 2026-10-27: trychroma.com, github.com

pgvector (Supabase or Neon)PostgreSQL

A Postgres extension for storing embeddings and running vector similarity search.

Best for: RAG and semantic search next to relational data. Skip it if: You have tens of millions of vectors.

Plain SQL; Supabase MCP can enable and query it.

Included on Supabase Free and Neon Free

  • Supabase Free: 500 MB database
  • Neon Free: 1 GB storage, 100 CU-hours per project/month
  • HNSW index up to 2,000 dims (4,000 with halfvec)

A 1,536-dim vector takes about 6 KB, so free storage fills fast.

Free; pay your Postgres host

Checked 2026-10-27: github.com, neon.com, supabase.com

PineconeService

A fully managed serverless vector database.

Best for: Managed vector search for RAG. Skip it if: You’d rather keep vectors next to your Postgres data.

Official Pinecone Developer MCP server.

Starter

  • 2 GB storage
  • 2M write units and 1M read units/month
  • 5 indexes, 1 project

Starter runs only in AWS us-east-1.

Builder $20/mo flat (new); Standard $50/mo minimum

Checked 2026-10-27: pinecone.io, docs.pinecone.io

Qdrant CloudApache-2.0

Managed hosting for Qdrant, an open-source vector search engine.

Best for: Filtered vector search you can later self-host. Skip it if: Your app sits idle for weeks.

Official mcp-server-qdrant; llms.txt.

Free Tier cluster

  • 0.5 vCPU, 1 GB RAM, 4 GB disk
  • About 1M vectors of 768 dims
  • No credit card

Suspended after 1 week unused; deleted after 4 weeks of inactivity.

Usage-based by vCPU, RAM and disk

Checked 2026-10-27: qdrant.tech, qdrant.tech

Turso vector searchMIT

Native vector columns and a DiskANN index in Turso/libSQL, a SQLite-compatible database.

Best for: Per-tenant SQLite databases with embeddings built in. Skip it if: You’re already on Postgres (pgvector is simpler).

Official Turso Cloud MCP; llms.txt.

Free

  • 100 databases, 5 GB storage
  • 500M rows read, 10M rows written/month
  • Vectors up to 65,536 dims

The DiskANN index is approximate and must be queried explicitly (vector_top_k).

Developer $4.99/mo

Checked 2026-10-27: turso.tech, docs.turso.tech

Upstash VectorService

A serverless vector database priced per request.

Best for: Small serverless RAG apps. Skip it if: You need more than 10K queries a day for free.

Official remote Upstash MCP.

Free

  • 10K queries and updates/day
  • Max 1,536 dimensions
  • 1 GB data

The 1,536-dimension cap rules out 3,072-dim embedding models.

Pay-as-you-go $0.40 per 100K requests

Checked 2026-10-27: upstash.com

Not sure what you need? Describe your app and the Build Kit picks a whole free stack, with starters and a plan.

More guides