Pi installs for free; the open source harness still has three bills

The harness costs nothing to install

Pi installs from one shell line on macOS or Linux, or from npm, and neither asks for an account. curl -fsSL https://pi.dev/install.sh | sh is the whole first step, and Pi needs Node.js 22.19 or newer (pi.dev). Start it inside a project and it runs; provider keys come from the environment, or /login connects a subscription or an API key (earendil-works/pi).

The harness is free in the way a compiler is free. Nothing meters the loop, the four built-in tools (read, write, edit and bash), or the four modes Pi runs in. What the install does not bring is a machine to run on and a model to run against. It also leaves out what Pi skips: sub-agents and plan mode arrive as packages, and so does a permission gate. The Pi harness page takes that trade apart, and the Pi coding agent guide walks the first session.

The model calls are the meter

Every tool call ends as a request to a model, and the provider of that model sets the rate. Pi reads keys from the environment and names Anthropic, OpenAI, Google, Azure, Bedrock, Mistral, Groq, Cerebras, xAI, Hugging Face, Kimi For Coding, MiniMax, NVIDIA, OpenRouter and Ollama among its providers (pi.dev). Two controls sit in the harness. /model switches provider and model mid-session, and a model served through Ollama runs on hardware you already own, so those turns never reach a provider's invoice.

Which account pays depends on the plan around the harness. On Kortix's Free plan your own key answers every call, or the ChatGPT subscription you already pay for answers instead; your agents use it for model calls and no credits are spent. The Credits plan runs managed DeepSeek and GLM models, and each call costs the provider's base price plus a 5.5% platform fee, paid from credits. The same session can spend a provider account or a credit pool, depending on the model it picks.

The machine, and the person who reads the diff

A session needs a machine. On a laptop that is the one you own, at no extra charge and with whatever reach your account has, and running Pi in the cloud weighs that against a hosted one. Celesto keeps Pi, the conversation and the model credentials on your machine while read, write, edit and bash run in a Celesto cloud computer. Its Pi page offers a free plan with no credit card and says fair-use runtime and active-computer limits apply (celesto.ai/pi-coding-agent). Celesto's homepage quotes $13.33 per 20,000 minutes in its CI-runner table for a 2 vCPU / 8 GB runner; that is a runner figure, not a price for a Pi session (celesto.ai, checked 28 October 2026).

Kortix sells the same layer in credits. Free is $0 with 500 credits a month, 1 project and your own key for a premium model, or your ChatGPT subscription. Credits starts at $50 a month at 100 credits per dollar, runs managed DeepSeek and GLM on one shared pool, and rolls unused credits over one month; members add nothing to the price. Enterprise is custom, and self-host is free (Kortix pricing). The stack runs on Kortix Cloud, in your VPC, or on your own on-prem network (Read the docs).

The third bill never arrives as an invoice. A session ends with a diff in the working directory, and a person reads it before it goes anywhere. That reading time is the same whether the machine was free or paid, and no plan itemizes it. Kortix gives the review a shape: the agent commits to a session branch named with the session id, opens a change request, and a person merges it (Read the docs).

A six-person month, worked from the published numbers

Say a team of six puts its work on one Credits plan. The plan is $50 a month, and at 100 credits per dollar that buys 5,000 credits for the account. A month that uses 3,000 of them still costs $50, and the 2,000 left over roll into the next month. A seventh member draws on that pool without adding a seat price.

Model calls come out of the same credits when the session runs on managed DeepSeek or GLM: each call costs the provider's base price plus a 5.5% platform fee, paid from credits. With the team's own key, model usage goes straight to the provider. The first $1,000 a month is free, then Kortix adds 5%, paid from credits; with your own keys, credits pay for the agent computer and 5% of model usage past $1,000 a month. A ChatGPT subscription spends no credits at all.

The same six people on laptops pay no plan fee. Each pays a provider bill and reviews their own diff. Neither version counts the hours the review takes, so the smallest plan line is not the smallest month.