What does your company brain cost?
Cognee runs on local models — on hardware you already own, the whole pipeline is free. Size your documents, code, and database tables below and compare BYOC on local models against the same pipeline on API LLMs.
Full pipeline: graph extraction + summarization + embeddings
Parsed deterministically (AST) — embeddings only, no LLM calls
Semi-structured + unstructured mix — one extraction pass + embeddings
Every agent run feeds the brain — extraction pass + embeddings
Schema-mapped rows — one small schema LLM pass, then embeddings
Agents and people asking the brain (graph completion, cognee's default)
One cognee Postgres per agent, group of agents, org, or user — hosted on the shared fleet
Workstations or on-prem servers — $0, the whole pipeline is free
Auto-scaled: 1× 48 vCPU / 192 GB node — sized vertically for the data, horizontally for the query volume.
Prices per 1M tokens, as of August 2026
Storage: 255.7 GB total — 243.2 GB of graph/vector/relational data plus 12.5 GB across 25 isolated Postgres instances — included in both tracks at $0.08/GB/mo. Cognee Cloud reference: one-time $4,093 at $1/1M tokens processed.
Methodology & assumptions
- 1 GB of documents ≈ 25M tokens (≈10% of document bytes are extractable text — a fixed assumption, measured at 8.65% on a real contract PDF). A GitHub repository ≈ 1M tokens (~100k lines), API data ≈ 25% extractable, agent traces ≈ 50%, and ~40 tokens per database row — all estimates; measure on your data.
- Each isolated memory instance is its own cognee Postgres hosted on the shared node fleet — per agent, group of agents, org, or user. Modeled at ~0.5 GB base storage and ~0.25 GB RAM per instance (estimates); instance count therefore grows the fleet, not just the disk.
- The local track runs cognee's LLM-free extraction (GLiNER2-class 205M CPU encoder): measured at ~1.6M pipeline tokens in ~22 minutes on one laptop-class CPU — ~1,200 tokens/sec/node — with zero LLM completions. Nodes auto-scale: vertically so the working set fits RAM, horizontally so ingest lands within 24h and queries stay under 50% utilization. Node rates surveyed August 2026 (Hetzner-class VPS vs AWS r6g on-demand).
- Query cost models cognee's default graph completion: ~1,550 input / ~400 output tokens per question (top_k=15). Retrieval-only search types cost $0 in model fees on every track; on the local track completions run on a local model.
- “Pure agentic reasoning” is the no-memory baseline: the agent re-explores the raw data with LLM calls on every query — estimated at ~0.5% of the corpus per question (floor 20k tokens). Past a ~1M-token context budget, every doubling of the corpus adds another ~250k-token exploration round, plus ~3k reasoning tokens out. Nothing is remembered, so the exploration bill repeats every year — and quality degrades as the corpus outgrows what one query can re-read.
- Documents run the API pipeline at graph extraction (~0.78 output ratio) + summarization (~0.39); code is parsed deterministically (AST, no LLM); database rows take one small schema-mapping pass. Storage uses cognee's embedded defaults (LanceDB, Ladybug, SQLite) at gp3 rates; embeddings are not duplicated onto graph nodes (verified against cognee source). API prices as of August 2026. Cognee Cloud meters $1/1M tokens processed, plus $5/workspace on Standard.