Factory Droid vs Claude Code: models, billing and migration

Factory Droid and Claude Code are both proprietary, closed-source terminal coding agents, and the choice between them is decided less by the terminal than by one question: what happens when you point the client at an endpoint you chose. Droid will take any model over three wire protocols, but its plan fee survives a custom key and Factory keeps metering you past an allowance it does not publish. Claude Code's gateway credential replaces the subscription outright — and Anthropic "doesn't support routing Claude Code to non-Claude models through any gateway." Cheaper tokens and a wider model shelf are, on these two clients, two different products.

Both are current and both shipped the same day. On the npm registry on September 21, 2026, Factory's droid is at 0.223.0, published September 19, 2026, declaring UNLICENSED; Anthropic's @anthropic-ai/claude-code is at 2.1.278, published September 19, 2026, declaring SEE LICENSE IN README.md. Both ship near-daily, so check the registry rather than trusting these two numbers. Neither is open source: the GitHub API reports no license on either Factory-AI/factory or anthropics/claude-code, and neither repository is archived. So "fork it and pay only for tokens" is off the table for both, and every configuration route below is the one the vendor documents. One sourcing note: factory.ai now answers HTTP 307 to factory.com (checked the same day), while docs.factory.ai still serves directly, so prices below cite factory.com and behaviour cites docs.factory.ai.

Claude Code vs Factory Droid: which one to pick

Take the row that describes you and stop there. The rest of this page is the evidence behind these five lines.

Two of those rows deserve a warning. The FAQ on Factory's pricing page answers "Which products are included in each plan?" with "All products; IDE, Web, CLI, Slack etc are included in all plans", and that is true of plan access; Factory's BYOK page separately says custom models "don't appear in Factory's hosted web or mobile platforms." Both sentences are published, they are about different things, and a buyer who reads only the first one and then buys a key will budget wrong. And "pay only for tokens" on Claude Code holds only when you set a credential: Anthropic's documentation is explicit that setting ANTHROPIC_BASE_URL alone, without a gateway credential, does not replace the subscription — the traffic routes through your endpoint while your claude.ai login stays the active credential and its limits and billing still apply.

The custom-endpoint door, side by side

This is the mechanical difference, and it explains every consequence further down. Droid takes a config file and a protocol name; Claude Code takes two environment variables and assumes one protocol.

Read against Factory's custom-models documentation, its individual-plans page, its org-control reference, Anthropic's gateway overview and its connect guide, all on September 19, 2026. Kunavo serves all three of Droid's protocols — /v1/messages, /v1/responses and /v1/chat/completions — so both configurations below are expressible. Neither has been runtime-tested here with either client; they are derived from the vendors' documented protocols and should be treated as a starting point while you keep your working route available.

{
  "customModels": [
    {
      "model": "claude-sonnet-5",
      "baseUrl": "https://api.kunavo.com",
      "provider": "anthropic"
    }
  ]
}

For the Anthropic protocol the base URL is the origin with no /v1, since Droid appends the path itself; see the base URL reference. For the Chat Completions path instead, set provider to generic-chat-completion-api and the base URL to https://api.kunavo.com/v1. Claude Code needs no file at all:

# Claude Code: environment variables, no config file required.
export ANTHROPIC_BASE_URL=https://api.kunavo.com
export ANTHROPIC_AUTH_TOKEN=sk-your-key   # travels as: Authorization: Bearer

# ANTHROPIC_API_KEY is the same credential in the x-api-key header instead.
# Anthropic's docs: "If you weren't told which kind, use ANTHROPIC_AUTH_TOKEN."
# Optional, off by default — adds the endpoint's models to the /model picker:
export CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1

One asymmetry worth knowing before you enable that last line. Claude Code's discovery request is GET /v1/models?limit=1000 with a three-second default timeout, and any redirect — even http to https — fails it silently; it then keeps only entries whose id contains "claude" or "anthropic". Kunavo's Claude ids are the bare Anthropic ones, such as claude-sonnet-5, so they pass that filter, but whether the endpoint answers inside the timeout has not been measured here. Discovery is off by default and the built-in model list works without it.

What each client switches off behind a third-party endpoint

Both clients lose capability when they stop talking to their own vendor, and both vendors document it. The lists are different in kind: Factory's losses are about where your key reaches, Anthropic's are about which features stop.

Anthropic's availability matrix has columns for a subscription, the Console and four cloud providers, and none for a generic third-party endpoint. Its own rule fills the gap: availability "matches the underlying provider the gateway forwards to, except for the features Claude Code itself turns off." That is a rule, not a list, so a feature not named above is genuinely unresolved rather than confirmed either way — auto mode is one such case, and this page does not claim it works or that it does not.

A second class of breakage is protocol-level rather than product-level, and it falls on the endpoint. Anthropic's gateway compatibility guide is blunt about it: cache_control markers must be forwarded unchanged or "the conversation bills as uncached input on every turn"; where a capability pairs a beta header with a body field — context management, beta tool fields, output_config — the two must travel together or the request returns a hard 400, while a header-only capability such as extended context goes silently missing instead; streams must not be buffered and keep-alive pings must be relayed, or Claude Code aborts after 300 seconds of silence; error bodies must be forwarded unmodified or its automatic capability-downgrade retry stops matching. Token counting is the one optional endpoint — Kunavo does not implement /v1/messages/count_tokens, so /context falls back to a character-based estimate.

Migration cost, in both directions

The instruction file is the cheap part. Factory's AGENTS.md documentation says Droid looks for AGENTS.md and also accepts agents.md, Agents.md, CLAUDE.md and Claude.md, searching the directory, then .factory/, .agents/ and .agent/ up to the git root — its stated reason is that "existing instruction files from other coding tools keep working." Anthropic's memory documentation goes the other way: Claude Code reads AGENTS.md as project instructions when there is no CLAUDE.md at or above your working directory, and reading it directly requires v2.1.277 or later — the version published on September 18, 2026. One condition attaches on this page's own subject, and Anthropic states it twice: reading AGENTS.md also needs a session that fetches feature flags from Anthropic, and the environment-variable reference lists the sessions that skip that fetch — a third-party provider such as Amazon Bedrock, a Claude apps gateway session, or any session with telemetry or GrowthBook turned off. A generic ANTHROPIC_BASE_URL gateway is not on that list and not excluded from it either, so whether a Kunavo-routed session reads AGENTS.md is undocumented and untested here. Where it does, Anthropic's fallback is to import the file from a CLAUDE.md, which works in every session.

Everything else is per-client. Droid has /commands for custom commands, /droids for subagents, /skills, and /mcp for MCP servers; Claude Code has its own commands, subagents, skills, hooks, plugins and MCP configuration, which Anthropic lists among the features that work on every provider. The formats and directories differ, so budget a rewrite of those, not a copy. Autonomy is the other translation cost: Factory exposes Off, Low, Medium and High autonomy levels alongside Normal, Spec and Mission interaction modes, while Claude Code uses its own permission modes — there is no mapping table between the two, so re-derive your guardrails deliberately rather than assuming a like-for-like setting exists.

What the tokens cost, and what the plans cost

Only one side of this comparison can be priced honestly. Factory publishes per-model multipliers against Factory Standard Credits but never the dollar value of a credit, so no one outside Factory can convert a Droid session into money — the Factory Droid pricing guide works through the plan table, the credit model and the rolling windows in full. What can be priced is the token path both clients share.

Assume one working day of agent sessions at 1.5M uncached input tokens and 100K output tokens, and a 20-day month. That is a stated assumption, not a measurement of any session. At current Kunavo catalog rates:

Prices are per million tokens; Anthropic list rates read from its pricing page and Kunavo rates from the live catalog, September 19, 2026. This is illustrative token arithmetic under the stated assumptions — not a measured task cost, not a bill ceiling, and it excludes cache charges, tool calls and tax. Compare the row for the model you will actually run rather than a vendor-wide gap. On Claude Sonnet 5 the catalog rate is Anthropic's own list price, so there is no saving on that row at all, while Claude Opus 5 carries the same $10.00 output rate as Claude Sonnet 5 against a much higher official reference. The gap is per model, not per vendor.

Now the reality check, because a flattering assumption is how these comparisons go wrong. Anthropic's own cost documentation reports that across enterprise deployments "the average cost is around $13 per developer per active day and $150-250 per developer per month, with costs remaining below $30 per active day for 90% of users." That $13 is roughly 3.3× the modeled day above at the Claude Sonnet 5 rate, so the day modeled here is a light one. Anthropic's published monthly range straddles the $200 that Factory Max and Claude Max 20x each cost: the bottom of it, $150, is below both, and the top, $250, is above both. A rate card alone therefore cannot settle the plan-versus-tokens question — where you land inside that range decides it, and Anthropic reports the range across enterprise deployments, not for your team.

Factory prices from docs.factory.ai, Anthropic's from claude.com/pricing, both September 19, 2026. Name the vendor every time you say "Max": Factory Max and Claude Max 20x are different products at the same $200. Note also that the two Max tier prices are not on claude.com/pricing, which says only "From $100 Per month" — the $100 and $200 figures come from Anthropic's Help Center, which adds that they are "for web subscriptions only." Where two official pages differ like this, confirm the amount in checkout before you commit.

If the token path is the one you want, Kunavo's minimum top-up is $10 in prepaid credit with no plan attached, and the catalog amount is a billing floor rather than a ceiling — the billing guide explains why. Cheapest listed rate and lowest cost to finish the task are still different claims: a smaller model that needs three attempts and a human review can cost more than one pass on a stronger one. Create a Kunavo account if you want to measure that on your own repository; Claude Code's side of the setup is written up at the Claude Code integration page, and there is no Droid setup page here, which is why the Droid configuration above is inline and marked untested.

Surfaces and execution model

Read from Factory's documentation index and CLI reference, and from Anthropic's feature-availability page, September 19, 2026. Note what the cost readouts are: both are the client's own arithmetic over reported tokens, not a statement from a billing system, so reconcile either against your provider's usage record before you trust a monthly figure. For Kunavo that record is the usage endpoint and the dashboard ledger.

If neither client fits, the adjacent comparisons on this site take the same approach: OpenCode vs Claude Code, Claude Code vs Codex and Goose vs Claude Code cover clients whose model routing and account models differ again, and the Claude Code provider comparison covers the endpoint choice on its own.

FAQ

Checked September 19, 2026: both npm packages' latest version, publish date and license field, and both GitHub repositories' archived and license fields through the GitHub API; the factory.ai to factory.com redirect; factory.com/pricing including its FAQ answers; Factory's individual-plans, organization-plans, custom-models, org-control, AGENTS.md, CLI-reference and documentation-index pages; Anthropic's gateway, gateway-connect, gateway-protocol, feature-availability, memory, environment-variable, permission-mode, model-deprecation and costs pages; claude.com/pricing and the Max plan Help Center article; and Factory's Terminal-Bench announcement. The two npm version numbers and the two GitHub license fields were re-read on September 21, 2026. Factory's per-model credit multipliers were not re-read for this page and are covered on the pricing guide instead. Kunavo token rates are read from the live catalog and the cost figures are illustrative arithmetic under the stated assumptions. Kunavo has not runtime-tested either client against its API, so every configuration claim here is read from the vendors' documented protocols rather than from a session.