Oh My Pi pricing: what omp costs, and which API to run it on

Oh My Pi costs nothing. The omp client is MIT-licensed open source, and omp.sh publishes no pricing, billing, checkout or sign-up page at all — so "Oh My Pi pricing" is a question about the model tokens you spend through it, not about the tool. The part worth reading before you budget is stranger than a rate card: let omp discover a gateway's models and its own cost readout shows $0 for every one of them, by design.

Checked September 21, 2026 against the GitHub and npm APIs: can1357/oh-my-pi is active, not archived and not a fork record on GitHub, MIT-licensed, with 32,293 stars, and the latest non-draft release is v18.2.8, published 17:31 UTC that day. Date any version you quote — v18.2.6 shipped September 18 and v18.2.7 landed at 02:13 UTC on September 21, so this number moved three times in four days. The package that installs the binary is @oh-my-pi/pi-coding-agent, also 18.2.8, MIT, not deprecated, with bin: omp.

One disambiguation first, because it decides which documentation applies. Oh My Pi is a fork; its README says so in as many words: "omp is a fork of Pi by Mario Zechner, rewritten as a coding-first surface". Pi is a separate, separately maintained client on its own release line — on September 21, 2026 its latest GitHub release was v0.87.0, against omp's v18.2.8 — so omp's major number is not eighteen generations of maturity over Pi. Pi's plans and provider surface do not transfer; see Pi pricing for those and Oh My Pi vs Pi for the differences.

What Oh My Pi charges, and what actually bills you

Two of those lines are the ones people get wrong. The web-search key is a separate axis: pointing omp's models at a gateway buys you no search, and buying a search key buys you no models. And the relay is the project's only hosted component — it carries no inference, and its documentation says every session payload is sealed with AES-256-GCM before it touches the socket, leaving the relay only room ids, connection counts and opaque ciphertext frames. But no price is published for it and nothing establishes who pays for the default host.

The omp-specific trap: gateway models bill at $0 on screen

omp estimates spend locally. Its models documentation says it "estimates token costs from the selected provider/model's catalog pricing, preferring server-reported monetary costs when available", and that completed messages keep their recorded costs — switching models or reopening a session does not reprice what you already spent.

That catalog is the problem for anyone using a gateway. The same page states that discovered proxy and gateway models "are the opposite case: their pricing is provider-specific and rarely matches the bundled catalog, so discovery keeps them at a local-unknown zero cost and no tariff applies to them". If you configure a discovery block and let omp read a gateway's model list, every model it finds arrives with no price, and the cost segment in the status line has nothing to sum. Your provider still bills you normally. Only the readout is empty.

The documented fix is to declare the models you care about by hand, with an explicit cost block. The unit is per million tokens — omp's catalog source documents the type as per-million-token rates for one model pricing tier — but the schema states no currency, so match whatever the card you are copying uses. Rates below come from the live Kunavo catalog:

# Derived from omp's documented schema in docs/models.md and
# docs/providers.md, read 2026-09-21. Not runtime-tested by Kunavo.
providers:
  kunavo:
    baseUrl: https://api.kunavo.com/v1
    api: openai-completions
    apiKey: KUNAVO_API_KEY      # an env var NAME, a literal, or !shell-command
    authHeader: true
    # No discovery block here, deliberately: discovered models arrive with no
    # price, so every model you want costed is listed by hand below.
    models:
      - id: claude-sonnet-5
        name: Claude Sonnet 5
        contextWindow: 1000000
        maxTokens: 128000
        cost:                   # per MILLION tokens; the schema states no currency
          input: 2.00
          output: 10.00
          cacheRead: 0.20
          cacheWrite: 2.50

Three consequences worth planning around. First, models.yml "does not accept a timeBased schedule", and an explicit cost "is a flat-price override and disables inherited time-based pricing for that model" — so a provider with peak and off-peak rates cannot be modelled accurately here, and the ↑/↓ tariff arrow omp shows for scheduled pricing never appears. Second, a key pinned in models.yml deliberately outranks a stored OAuth credential, "so a key supplied for a custom baseUrl or gateway is honored instead of forwarding an upstream OAuth token the proxy would reject" — which removes the most common gateway authentication failure before it happens. Third, mistakes here are quiet: a models.yml that fails validation leaves the registry "operating with built-in models", so a broken config looks like a missing provider rather than a crash, and an apiKey value that does not name an existing environment variable is used as the literal key.

When omp's estimate and the vendor's price card disagree

This is checkable today and it is not a rounding difference. omp's bundled catalog carries DeepSeek's scheduled card, and its documentation says deepseek-v4-pro "initially uses peak rates of $1.32 uncached input, $0.044 cached input, and $3.96 output per million tokens", but that "From 2026-09-14 04:00 UTC, its estimates use the Flash rate card" — that is, $0.30 uncached input and $1.20 output at peak.

Both columns were read on September 21, 2026 — the left from omp's models documentation, the right from DeepSeek's own price card, which still prints $1.32 / $3.96 with no September change note. Read this narrowly: it is a gap between omp's local estimate and a published vendor rate, not evidence that DeepSeek changed its prices. If both are current, omp's readout understates deepseek-v4-pro spend by roughly four times on input. The general lesson holds for every route: omp's dollar figure is the client's own arithmetic, and your provider's ledger is the number that settles.

DeepSeek's schedule is worth knowing separately, because it halves a bill without changing a model. Its pricing page states that "Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday, excluding Chinese public holidays" and that all other hours are off-peak at half rate. No source read here lists the holiday calendar, so treat any given week as approximate.

Cheapest API for Oh My Pi

Answered honestly, no metered per-token rate beats running the model on your own machine, where there is no per-request charge at all. Kunavo's text catalog is the Claude, Gemini and GPT families; it carries no DeepSeek model and runs nothing locally, so neither of those two routes is on its price list. Published vendor rates for the routes omp can reach, all read September 21, 2026:

Two of those vendors have already met omp halfway. DeepSeek publishes its own Oh My Pi page, which tells you to write a full custom provider block at ~/.omp/agent/models.yml with baseUrl: https://api.deepseek.com — note, without /v1 — and warns "Do not rely on the built-in model entries" because recent builds list its ids without three compat fields it needs. Ollama publishes one too, where the whole setup is ollama launch omp, which it says "configures Ollama as a provider, sets up web search tools, and starts OMP". Neither page quotes a cost. If one vendor covers your work, those are the shortest paths on this list.

Cheapest listed rate and lowest cost to finish the task are different claims, and only the first is on the table above. A cheap model that needs three attempts and a review pass can cost more than a dearer one that lands in a single turn. omp ships omp bench, which its CLI reference describes as benchmarking models on "TTFT/prefill vs decode throughput with p50/p95 across chat, prefill, generation, and prompt-cache workloads" — useful, but it measures speed, not dollars, so it does not produce a cost-to-finish figure on its own. AI cost optimization covers the measurement method; the cost calculator does the arithmetic on your own token counts.

A worked estimate at Kunavo's live rates

This is illustrative token arithmetic, not a measured task cost and not a bill ceiling. omp routes work across nine roles that can each name a different model, so the assumption is two of them: a default turn of 200,000 uncached input and 15,000 output tokens, and a smol subagent fan-out of 120,000 input and 8,000 output tokens. Both assume no cache hits. Rates are live catalog prices per million tokens.

The role split is the lever these numbers exist to show. Running Claude Sonnet 5 on the main turn and Claude Haiku 4.5 on the fan-out estimates at $0.614 for that session, against $0.870 if both roles sit on Claude Sonnet 5 — the same work, routed differently. Note also that Claude Sonnet 5 and Claude Opus 5 carry the same Kunavo rate today and therefore the same estimate, which is a fact about one catalog on one date, not a general relationship.

On Claude specifically a list comparison is sourceable, because Anthropic's own page on September 21, 2026 prints Sonnet 5 at $2 / $10, Opus 5 at $5 / $25 and Haiku 4.5 at $1 / $5 per million. Against those, Kunavo sells Claude Sonnet 5 at Anthropic's list rate; Claude Opus 5 about 60% below it; Claude Haiku 4.5 about 60% below it. That comparison is made for the Claude rows only, because those are the three whose catalog reference was checked line by line against the vendor's live page for this guide; the Gemini and GPT rows are printed as Kunavo rates with no "versus official" figure attached. Model ids differ in spelling across surfaces too — OpenAI prints dots, Kunavo's slugs use dashes — so copy the id from the surface you are configuring.

Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. Cache charges and external tools sit outside this example, and whether prompt-cache markers survive a gateway into omp was not tested here — since caching is the largest single lever on a coding-agent bill, treat any example that assumes cache hits as an assumption. The minimum top-up is $10 in prepaid credit, a funding floor rather than a monthly fee or a task price; see billing details and caching.

Best API for Oh My Pi: which route wins when

One boundary is worth stating plainly rather than leaving to a table cell. omp offers /login anthropic, and Anthropic's legal and compliance page says that OAuth authentication "is intended exclusively for purchasers of Claude Free, Pro, Max, Team, and Enterprise subscription plans and is designed to support ordinary use of Claude Code and other native Anthropic applications", and that "Anthropic does not permit third-party developers to offer Claude.ai login into their own applications, or to route requests through Free, Pro, or Max plan credentials on behalf of their users". The same page reserves the right to enforce that without prior notice. Those two facts do not reconcile in a sentence, and no Anthropic statement naming Oh My Pi was found, so this page reports both and recommends neither reading. There is also a credit scheme for apps authenticating through Anthropic's Agent SDK, but its help-centre article opens "Update June 15: We're pausing the changes to Claude Agent SDK usage described below" and was last updated June 16, 2026 — three months before this check — and omp does its own OAuth rather than using that SDK, so whether it applies is doubly unsettled. The unambiguous ways to pay are an API key billed to you or a gateway key.

Set it up and check the first charge

Kunavo publishes a configuration reference for this client, and a published reference is not a compatibility test: Kunavo has not run omp against its endpoint — not a session, not a streamed turn, not a tool round-trip, not a routing check. Everything above is derived from omp's documented schema and each vendor's published pages. One detail that page settles for you: Kunavo's /v1/models route does not emit the field omp's discovery.type: proxy mode reads, which is why the documented setup uses openai-models-list instead; that route also lists media models, so declare the chat models you want by hand if you would rather not see the rest in the picker.

Keep a working route available while you try it, run one bounded task, then read the charge your account actually recorded against omp's on-screen estimate — with a hand-declared cost block they should be close, and without one the estimate is zero. Start at the Oh My Pi integration guide, and create a Kunavo account when you are ready to fund a key. If you are still choosing a client rather than a provider, Oh My Pi vs OpenCode compares the two on workflow and on who sells you the models; how an LLM gateway works covers the concept if that is the new part.

FAQ

Every source linked here was read on September 21, 2026: the GitHub and npm APIs for both oh-my-pi and pi-mono, omp.sh's sitemap, omp's README and its models, providers, collab and CLI-reference documentation, DeepSeek's price card and its Oh My Pi page, Ollama's pricing and Oh My Pi pages, Google's and OpenAI's price pages, and Anthropic's pricing and authentication-policy pages. Oh My Pi was not runtime-tested against Kunavo, and no claim here rests on an observed run. Kunavo token rates come from the live catalog; every dollar figure is illustrative token arithmetic rather than a measured task cost. Version numbers, the star count and the sitemap count all move — re-check rather than trust them.