→
Should I self-host Perplexity?
Perplexica reproduces the interface, but every answer still burns metered LLM API tokens on your key. Heavy use costs more à la carte than the $20 bundle; light use doesn't justify the stack. You'd be self-hosting the wrapper while renting the brain.
The math (nothing hidden, including your time)
| Perplexity (Pro) | $20.00/mo |
| VPS share + storage | −$0.50/mo |
| Your maintenance: 20 min/mo at $20/h | −$6.67/mo |
| Net saving | $12.83/mo |
| Setup: 15 min measured (one-time) | $5.00 |
| Break-even | 0.4 months |
| Markup Index (price ÷ real self-host cost) | 2.8× |
Price: source, checked 2026-08-06. Inputs are stored in git; every number above is derived, never hand-written.
What you lose
- Subsidized frontier-model access — the subscription is the cheap part
What you're paying Perplexity for
- AI answers with citations
- web search synthesis
- follow-up threads
Ranked alternatives
An answering engine with citations, assembled from a SearXNG you run and a model key you already pay for. Renamed from Perplexica in 2026.
One chat window in front of Anthropic, OpenAI and Google, with the conversation history in a database you own.
Hardware
Runs comfortably on a real VPS (2 GB+).
Verification
Protocol v1 · verified by zernonia on 2026-08-10 · setup measured at 15 min · AI assistant: claude-code (allowed & stated)
Full timed log
# Perplexity — verdict log - **Date:** 2026-08-10 - **Verdict:** NOT REALLY — keep paying - **Protocol:** v1 (2GB reference env, Docker preinstalled, AI assistant: claude-code) ## Why no compose file - This verdict is decided by economics/capability, not setup effort — the rubric doesn't require a timed install to conclude the subscription should stay (or that the replacement's cost lives elsewhere, e.g. GPUs, egress, or an installer script). ## The call Perplexica reproduces the interface, but every answer still burns metered LLM API tokens on your key. Heavy use costs more à la carte than the $20 bundle; light use doesn't justify the stack. You'd be self-hosting the wrapper while renting the brain.
Verdict history
- 2026-08-10: unscored → NOT_REALLY — Initial verdict: Perplexica reproduces the interface, but every answer still burns metered LLM API tokens on your key.