Factory Droid pricing: plans, API costs and provider choices

Factory's Droid CLI is not free and has no published free tier: individual plans list at $20/mo (Pro), $100/mo (Plus) and $200/mo (Max), Teams at $60/mo per team plus $40/mo per seat, and Business and Enterprise are custom-priced. Bringing your own API key does not remove that subscription. Factory's own documentation says BYOK is free only "up to an allowance" and is charged after that, so the honest budget is plan fee plus your provider's token bill plus whatever Factory meters past the allowance.

This is Factory's agent — the Droid CLI, the Factory App desktop client and the headless droid exec mode — not a Motorola DROID handset. One sourcing note: factory.ai now 307-redirects to factory.com on both the root and /pricing (checked September 18, 2026), while docs.factory.ai and app.factory.ai still serve directly and Factory's own navigation still renders "Factory.ai". The migration is mid-flight, so everything below is cited at the resolved URL.

Three separate bills, not one

Most confusion here comes from collapsing three charges into one number: the plan fee (the CLI quickstart says "if prompted, sign in via your browser", and the CLI reference carries /login and /logout commands); Factory-metered model usage, meaning included Standard Usage plus prepaid Extra Usage past it; and your own provider bill if you configure a custom model. The third is the one comparisons get wrong. On some agent clients a personal API key replaces the vendor's meter outright. On Factory it is additive.

Factory plan prices

Checked September 18, 2026 against Factory's pricing page, the individual-plans documentation and the organization-plans documentation. Marketing page and docs agree on the three individual prices, so these are not a stale landing page. No annual price, free tier or trial appears on the served pricing page.

How Factory meters usage, and why it cannot be converted to dollars

Individual usage is not a monthly token bucket. Factory's individual-plans documentation describes three rolling windows — 5-hour, 7-day and 30-day — that "each runs independently", and you need headroom in all three to send a request. A plan can be exhausted on a Tuesday afternoon and usable again on Wednesday without anything resetting monthly.

Consumption is counted in Factory Standard Credits, which the Analytics API reference describes as computed from raw input and output tokens with cache discounts. Each model carries a published multiplier against that unit rather than a per-token price:

Read off Factory's model list on September 18, 2026. Several entries are promotional with hard expiry dates — GPT-5.6 Sol bills at 1.6x "through November 22, 2026, then returns to 2x", Gemini 3.8 Flash at 0.3x "through January 1, 2027, then returns to 0.6x" — so any multiplier table needs an as-of date, this one included.

Factory does not publish the dollar value of one Standard Credit. It is not on the pricing page, on either plans page or in the Analytics reference, and we found no conversion anywhere else in the docs. The consequence for buyers is direct: the published figures do not let you compute what a given task costs on a Factory subscription. The multipliers rank models against each other on Factory's meter; they do not convert to a per-million-token rate. A dollar figure for Factory model usage quoted anywhere else is therefore not taken from Factory's published pricing or docs — ask Factory for the conversion before trusting one.

What a custom API key does and does not change

Factory calls this BYOK, and it is real, but narrower than the phrase suggests. Five limits from Factory's BYOK documentation, its individual-plans pricing page and its org-control reference, checked September 18, 2026:

• It is local-surface only. "Custom models are available in the Droid CLI and the desktop app, which read your local settings.json; they don't appear in Factory's hosted web or mobile platforms." Delegation through the hosted product keeps running on Factory-billed inference.
• It is still metered. The individual-plans page says "BYOK is free up to an allowance on all Individual plans. After that, usage is charged according to your specific plan." The allowance size is stated nowhere, and the BYOK page itself does not mention the allowance at all. Whether BYOK traffic also consumes the three rolling windows is likewise not stated; treat both as unverified.
• An admin can remove it. modelPolicy.allowCustomModels: false disables user BYOK entirely, and allowedBaseUrls can force all custom-model traffic through one approved host. A net-new custom model added through org-managed settings cannot use a static apiKey — Factory says to use apiKeyHelper, a keyless endpoint or a Bedrock model instead, while "existing static-key models continue to work and can still rotate their key". apiKeyHelper itself runs a shell command, so it is honoured only from org-managed settings and is stripped from user, project and folder settings.
• Some features do not follow. For generic providers, "prompt caching support varies by provider and cannot be guaranteed" — and an agentic client re-sends its context on every step, so whether that context is cached can move the bill a long way. The CLI reference lists /cost as "show usage statistics", so read that rather than assuming a route caches — and check your provider's own usage record too.
• Factory limits its own claims. "Only Anthropic and OpenAI models accessed via their official APIs are fully tested and benchmarked," and models below 30 billion parameters "have shown significantly lower performance on agentic coding tasks." That is Factory's wording about its own product, not ours.

The CLI is proprietary — the npm package declares UNLICENSED — so "self-host it and pay only for tokens" is not available here the way it is for an open-source client. The only mode that drops the Factory account entirely is the airgapped build, which "skips sign-in", offers your models only, and is "not distributed through the public download channels".

Configuring a custom endpoint

Per the settings reference and the BYOK page, custom models go in ~/.factory/settings.json (%USERPROFILE%\.factory\settings.json on Windows) under a customModels array, with overlays at ~/.factory/settings.local.json and <project>/.factory/settings.local.json. Required fields are model, baseUrl and provider. In user settings, provider must be exactly one of three values, and each selects a different wire protocol: anthropic for the Anthropic Messages API, openai for the OpenAI Responses API, and generic-chat-completion-api for OpenAI Chat Completions.

Kunavo serves all three of those protocols. The configuration below is derived from both products' documented protocols; it has not been runtime-tested with the Droid CLI here, so treat it as a starting point and keep your working route available while you try it.

{
  "customModels": [
    {
      "model": "claude-sonnet-5",
      "displayName": "Kunavo Claude Sonnet 5",
      "baseUrl": "https://api.kunavo.com",
      "provider": "anthropic",
      "apiKey": "${KUNAVO_API_KEY}",
      "maxOutputTokens": 8192
    }
  ]
}

Two details matter. The Anthropic-style base URL on Kunavo is the origin with no /v1 — adding it produces requests to /v1/v1/messages, which 404 (see the base URL reference); Factory's own table gives https://api.anthropic.com, origin-only, for the official Anthropic route. And authMode can be left out, because Kunavo's Messages endpoint accepts Anthropic's x-api-key header as well as Authorization: Bearer, and x-api-key is what Factory sends by default. For the Chat Completions path instead, set provider to generic-chat-completion-api and baseUrl to https://api.kunavo.com/v1. Whether Droid's validator accepts a given model string verbatim is unconfirmed.

Two traps: the legacy ~/.factory/config.json format with snake_case custom_models and base_url is still loaded and merged, with settings.json taking priority — but Factory documents that ${VAR_NAME} expansion does not apply there, so a copied config can quietly send the literal placeholder as your key — use settings.json. And two official pages disagree on naming a custom model at the command line: the Droid Exec overview says custom: plus the display name and an index, the subagents page says custom: plus the model field and explicitly not the display name. Both checked September 18, 2026. Use the interactive /model picker and confirm the flag with droid exec --help on your version.

Create a Kunavo account to try that route; keys are issued in the dashboard and the balance is prepaid.

What the tokens cost: a worked estimate

Because Factory's credit price is unpublished, the only side of this comparison that can be computed honestly is the token side. Assume one headless run consuming 400,000 uncached input tokens and 30,000 output tokens — a stated assumption, not a measurement of any Droid session. At current Kunavo catalog rates:

Prices are per million tokens. Anthropic list rates checked September 18, 2026; Kunavo rates read from the live catalog. On Claude Sonnet 5 the catalog rate is the same as Anthropic's list price, so there is no saving on that model — the saving is on other rows, and the comparison is worth making per model rather than per vendor. At forty such runs in a month, Claude Haiku 4.5 estimates at $8.80 and Claude Sonnet 5 at $44.00.

This is token arithmetic, not a measured task cost and not a bill ceiling. It excludes cache charges, tool calls and tax, and assumes the smaller model finishes without extra attempts — the assumption that actually decides whether a cheaper rate produces a cheaper month. Agentic clients re-send context every step, so the cache-read rate usually matters more than the headline input rate: Claude Sonnet 5 cache reads bill at $0.20 per million here. Do not budget on cache hits over Factory's generic-chat-completion-api path, where Factory says caching cannot be guaranteed. Kunavo's minimum top-up is $10 in prepaid credit with no subscription; the billing guide explains why a catalog estimate is a floor rather than a ceiling.

Which route wins, and when

Keep "cheapest listed price" and "lowest cost to finish the task" apart. The lowest multiplier on Factory's published table is GLM-5.3-Flash at 0.06x, and Claude Haiku 4.5 is the cheapest of the three rows in the estimate above — not the cheapest rate in the Kunavo catalog, which carries smaller models below it. Neither figure tells you which model finishes your task in one pass. Run one bounded task per candidate and read the CLI's /cost usage statistics before moving scheduled work.

Comparing clients rather than providers: Claude Code pricing, OpenCode pricing and Cline pricing cover clients whose model routing differs from Factory's, and AI cost optimization covers the levers common to all of them. Kunavo publishes setup guides for Claude Code, OpenCode and Cline; there is no Droid setup page, which is why the configuration above is inline and marked untested.

FAQ

Factory plan prices, multipliers, rate-limit rules and BYOK wording checked September 18, 2026 against factory.com/pricing and docs.factory.ai; the factory.ai to factory.com redirect was checked the same day. Kunavo rates are read from the live catalog, and the cost figures are illustrative token arithmetic under the stated assumptions. Kunavo has not runtime-tested the Droid CLI.