Amp Code vs Claude Code: workflow, models and total cost

Amp Code (ampcode.com) and Claude Code are both agentic coding tools, but they sit on opposite sides of one line: where the agent runs and who holds the model credential. Claude Code is a local CLI you point at any Anthropic-format endpoint with two environment variables and no plan at all. Amp Code is a hosted service whose threads and provider connections live on Amp's servers, and where an arbitrary custom endpoint URL is, as of September 21, 2026, an early-access feature for Megawatt, Gigawatt and Enterprise customers, with no announced general-availability date.

Two disambiguations before the comparison, because both are actively wrong in the search results. This page is about Amp Code, the coding agent at ampcode.com — not Google's Accelerated Mobile Pages. And Amp is not a Sourcegraph product: its License Terms name "AMP FRONTIER CORPORATION", its About page opens "We are an independent agent research lab", and Sourcegraph announced the separation on December 2, 2025. Amp's tier structure also changed on September 13, 2026, so any comparison written before that date describes a product that no longer exists at those prices.

Which one should you pick

Pick Claude Code if your work is one developer in one repository on one machine, you want the model bill to be per token with nothing in front of it, and Claude models are what you want to run. It installs from Anthropic's installer or npm — the published package @anthropic-ai/claude-code was at 2.1.278 on September 21, 2026 — and the endpoint it talks to is two variables you own. There is no tier gate on redirecting it, and the migration cost of trying a different endpoint is unsetting two variables.

Pick Amp Code if you want the things a local CLI structurally cannot give you: agents running in parallel on remote computers, threads other people can open, a single dial that routes across several model vendors, and connectors for existing subscriptions. Those are real capabilities with a real price — orb minutes, a hosted thread store, and an experimental gate on custom endpoints — and they are the reason to pay, not an accident.

On permissions and execution, they default in opposite directions. Amp's tools documentation states "By default, Amp does not ask for approval before running tools", with a custom plugin as the documented way to change that. Claude Code's default permission mode, labelled Manual, "Prompts for permission on first use of each tool", with separate modes that auto-accept edits or skip prompts entirely. If you are introducing an agent to a repository you cannot afford to have mangled, that difference matters more than any model choice.

On plan limits, the asymmetry is sharp. Claude Code requires a paid Anthropic identity to sign in — "Claude Code requires a Pro, Max, Team, Enterprise, or Console account. The free claude.ai plan does not include Claude Code access" (setup documentation) — but a third-party endpoint replaces that identity outright rather than supplementing it. Amp inverts this: the client is free on Hobby, and it is the ability to bring an arbitrary endpoint that is gated behind a paid tier.

On migration cost, moving between them is cheap in both directions and the friction is instruction files, not code. Amp uses AGENTS.md; Claude Code reads CLAUDE.md and can also read AGENTS.md directly, with an @AGENTS.md import documented for the sessions that cannot, plus an /import command that requires v2.1.213 or later (memory documentation). Neither locks your repository.

How the two clients actually differ

Sources re-read September 21, 2026: Amp's CLI page, security page, models page, the Dial, Model Routing, and Anthropic's gateway, gateway connection and feature availability pages.

What each one costs, with a date on every number

Amp's tiers were restated on September 13, 2026, so treat any figure from an earlier write-up as void.

Prices from Amp's pricing documentation, read September 21, 2026. A gap worth knowing about: the Gigawatt price is not in the server-rendered HTML of Amp's pricing page — a fetch on September 21, 2026 finds only the Megawatt tier at $20 a month with 45,000 minutes, and no Gigawatt price or minute count, because that card only renders after you click its tab. A browser check on September 18, 2026 read $200 per month with 480,000 orb minutes. Confirm it in Amp's own checkout rather than trusting either reading here. Amp's pricing FAQ also carries an "Is there an Education Discount?" entry whose answer is collapsed and not in the served HTML, so no student price is quoted on this page.

Orb compute is metered separately from inference: Amp's pricing documentation says "Orbs use a separate usage allowance" and bills them by the minute, so a connection that moves model spend to your own provider account leaves this line untouched. Amp's published standard prices are $0.08 per hour for a1.tiny, $0.17 for a1.small, $0.33 for a1.medium, $0.66 for a1.large, $1.32 for a1.xxlarge and $2.13 for a1.3xlarge, all with 60 GB of disk, with a paused orb costing nothing. No subscriber rate card for orb time appears on that page, so no discounted per-hour figure is quoted here.

Read from claude.com/pricing on September 21, 2026. For a like-for-like anchor on what the work itself costs, Anthropic publishes one field figure, scoped to enterprise deployments rather than to all users: "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" (cost management). Amp publishes no equivalent per-developer figure, so this anchors one side only.

Pointing either client at your own endpoint

This is where the two products stop being comparable, and it is the decision most buyers are actually making.

Claude Code takes two variables and no plan gate. The base URL is an origin, not a path — Anthropic's own verification request appends /v1/messages to it — and the credential variable you choose decides the header: ANTHROPIC_AUTH_TOKEN goes in Authorization: Bearer, ANTHROPIC_API_KEY in x-api-key, and apiKeyHelper in both. Kunavo's Messages route reads either header, so either variable authenticates.

# Claude Code: the whole third-party-endpoint configuration.
export ANTHROPIC_BASE_URL=https://api.kunavo.com   # an origin, no /v1
export ANTHROPIC_AUTH_TOKEN=sk-kn-...              # sent as Authorization: Bearer

# Anthropic's own verification request appends the route itself:
curl -X POST "$ANTHROPIC_BASE_URL/v1/messages" \
  -H "Authorization: Bearer $ANTHROPIC_AUTH_TOKEN" \
  -H "anthropic-version: 2023-06-01"

The consequence is billing, not just plumbing: "the credential replaces the subscription login for that session, and the subscription's usage limits don't apply", and that traffic is "billed per token to whoever owns the credential the gateway forwards" (gateway documentation). Setting only the base URL with no credential does not replace the subscription. See our ANTHROPIC_BASE_URL reference for the origin-versus-path trap that causes most 404s here.

Amp's equivalent is tier-gated, and it is documented. The September 13, 2026 announcement lists "Custom endpoint URLs" among providers that "Amp Megawatt and Gigawatt members are also getting early access to", behind an experimental feature it calls "More AI Routers & Subscriptions". Amp's Model Routing documentation states the gate a shade wider — "AI model routers and gateways are experimental and in early access for Megawatt, Gigawatt, and Enterprise customers only. We will roll them out to everyone soon" — and it is also where the shape of the connection is written down. A Custom URL connection takes a base URL plus one of three wire formats: chat-completions and responses append those paths, anthropic-messages appends /v1/messages, and a base URL that already ends in /v1 is "accepted and not doubled". The credential goes in an Authorization: Bearer header "in every format", with per-connection headers and query parameters for anything else a provider needs, and a model-mapping list deciding which of Amp's canonical model IDs the connection serves. You add one in Model Routing in personal or workspace settings, or from the CLI with amp config model-providers.

One retention detail specific to Amp: its models documentation states that "Anthropic requires data retention for Claude Fable models, including Claude Fable 5.1", and that "credentials from an Anthropic zero data retention workspace do not have access to Claude Fable". Requests you send with your own key "are covered by the retention terms of your provider account", per Amp's security page.

Kunavo does not serve dictation, speech-to-text or any other transcription model, so Amp's Dictation slot is not one a Kunavo endpoint could fill.

A worked token estimate

This is token arithmetic on assumed volumes, not a measured task cost and not a bill ceiling. Assume one session that sends 250,000 uncached input tokens and receives 15,000 output tokens, at live Kunavo catalog rates per million tokens. The right-hand column shows Anthropic's published per-token rate for the same model, so you can see where the catalog sits.

The two matching rows are not a rendering fault: Claude Sonnet 5 and Claude Opus 5 carry the same catalog rate here — $2.00 in and $10.00 out per million tokens — so the same session estimate falls out of both, while Anthropic's own published rates for the two differ, as the right-hand column shows.

Anthropic's rates read from its pricing page on September 21, 2026, which also notes that Sonnet 5's $2/$10 rate is now standard rather than introductory. Two things this table cannot tell you. First, the same page warns that Claude 4.7-and-later models use a newer tokenizer that "produces approximately 30% more tokens for the same text", with the exact increase depending on the content, so comparing a newer model against an older one on rate alone understates the newer one. Second, cheapest listed rate and cheapest way to finish the task are different claims: a model that needs three attempts can cost more than one that needs one. Nothing here is a promise of the lowest price anywhere.

Set that against the other side of an Amp bill. Amp's pricing page says that for individuals and non-enterprise workspaces "Amp does not add a markup to providers' API prices" and gives a worked example where a thread using $2 of one provider and $0.50 of another deducts $2.50 from credits; purchased credits expire twelve months after purchase. On top of that sit orb minutes at the hourly rates above whenever work runs remotely. Amp never states what one credit is denominated in, so no conversion is asserted on this page.

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. The minimum top-up is $10 in prepaid credit — a funding minimum, not a task fee or a subscription. See billing details.

Setting it up, and what this page has not tested

Kunavo publishes a configuration reference for Claude Code. That is a published setup guide, not a compatibility test: no Claude Code session was run against Kunavo while writing this page, and Kunavo has never been connected to Amp through a custom endpoint at all, so nothing here claims either integration is verified. Keep a working route available while you try it, run one bounded task, then read the charge your account actually recorded for it. Start at the Claude Code 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, Claude Code without a subscription covers the gateway route in depth, the best API for Claude Code compares the endpoints you can put behind it, and the agent API directory records how several dozen clients accept a custom base URL. For a second hosted-versus-local comparison, see OpenCode vs Claude Code.

FAQ

Re-checked September 21, 2026 against Amp's pricing, orb-cost, tools, security, models, models-and-subagents, Model Routing, the Dial, Terms and Free Agent pages, and against Anthropic's setup, permissions, gateway, gateway-connection, feature-availability, cost and pricing documentation plus claude.com/pricing and the npm registry. The Sourcegraph separation post was read through a summarizer, not raw HTML. That pass corrected an earlier draft that had said Amp published no model-routing documentation, that described Amp's subscriber orb savings as published averages, and that put the word "BYOK" inside a quotation whose original reads "No Amp token fees or limits". No runtime test of either client against Kunavo was performed, and no performance or compatibility claim is made. Kunavo token rates come from the live catalog; every dollar example is illustrative token arithmetic.