Amp — the coding agent at ampcode.com from Amp Frontier Corporation, not Google's Accelerated Mobile Pages and not the ASX-listed AMP Limited — is free on its Hobby tier, $20 a month on Megawatt and $200 a month on Gigawatt. Those are software prices. What you actually spend runs on three meters that move independently: the tier, the cloud machines Amp calls orbs, and model tokens. Getting the Amp Code API cost right means pricing all three, because cutting one of them does nothing to the other two.
Two stale stories dominate the search results for this product, and both are worth clearing before any number. Amp is no longer a Sourcegraph product: it spun out on December 2, 2025 with the line "Amp is becoming a separate company. We're spinning out of Sourcegraph to become an independent research lab", and its Terms of Service name Amp Frontier Corporation. And "Amp Free" — the ad-funded mode from October 15, 2025 — is not the current free tier. Its own post now opens with "Update 4: Amp Free is paused. We don't intend to bring it back, but you never know." The free tier that exists today is Hobby, announced September 13, 2026.
The three meters behind Amp Code pricing
| What bills you | Price | Source, checked September 19, 2026 |
|---|---|---|
| The Amp client and all product features on Hobby | $0 | Pricing page: "All product features" on the Hobby card |
| Monthly tier | $0 Hobby, $20 Megawatt, $200 Gigawatt, no extra charge for a Teams workspace, custom for Enterprise | Pricing page |
| Orb compute (Amp's remote machines) | $0.08 to $2.13 per hour by size, billed by the minute | Orb sizes and costs |
| Local runners (agents on your own computer) | $0 | Free Agent post: run your agents "for free" |
| Model tokens charged to Amp credits | The provider's API price, with no markup added for individuals and non-enterprise workspaces | Docs pricing |
| Model tokens through your own key or subscription | Billed by that provider or plan; no Amp token fee on Hobby, Individual or Teams | Pricing FAQ, read September 18, 2026 |
| Non-model tools such as web search | Consume Amp credits regardless of how models are routed | Docs pricing |
Two figures a budget wants are simply not published, and guessing them is how most third-party write-ups go wrong. First, the pricing FAQ says a paid tier includes "orbin' time" and "Amp credits (which you can use for inference on all models)" — but no credit amount appears on the pricing page for either paid tier. The only first-party included-credit figure on any of the pages cited here is on the education page, which offers Megawatt at $10 a month for students and teachers with "$10 of included agent usage", described there as headroom for incidental usage rather than a working allowance; that discount does not apply to Gigawatt. Second, the docs say only that "larger orbs use the allowance faster, and smaller orbs use it more slowly", with no multiplier — so the included 45,000 minutes cannot be converted into hours of any particular machine size. Treat any write-up that prints those conversions as unverified.
Amp tiers, and what each one includes
| Tier | Published price | Included orb time | Notable terms |
|---|---|---|---|
| Hobby | Free | None — pay-as-you-go orbs, or free local runners | All product features, BYOK, linked subscriptions, no Amp token fees or limits |
| Individual / Megawatt | $20 per month | 45,000 minutes | Adds unlimited public and private repositories and your own portal domains; early access to the extra BYOK providers |
| Individual / Gigawatt | $200 per month | 480,000 minutes | The only individual tier that can use a1.3xlarge orbs — Enterprise workspace members can too |
| Teams workspace | No extra charge | Per member's own tier | Members can sit on different tiers; shared billing and pooled credits; SAML/OIDC SSO requires at least one paid member |
| Enterprise | Custom | Not published | Pooled credits only with no paid seats; SCIM, audit logs, IP allowlisting, spend limits; the one tier excluded from the no-BYOK-fee promise |
| Megawatt, education discount | $10 per month | 45,000 minutes | Students and teachers; $10 of included agent usage; Megawatt only |
A sourcing note on the Gigawatt row, because it is the one figure here that cannot be read from the page source: the $200 price and the 480,000 minutes sit behind a client-rendered tab on the pricing page and are absent from the static HTML. They were read in a browser on September 18, 2026. Everything else in this table was re-read from the live page on September 19, 2026. The pricing FAQ answers sit behind an accordion and were read on September 18, 2026; they are the source for the Teams and cancellation terms — cancellation is self-service in user settings and the paid tier runs to the end of its current monthly period.
One wording conflict to be aware of when reading Amp's own material: every tier card carries the flat line "No Amp token fees or limits", while the September 13 announcement writes it as "No BYOK token fees or limits for anyone (except our Enterprise tier)". The pricing FAQ resolves the two — no fees on Hobby, Individual and Teams, custom pricing for Enterprise — but if you are on an enterprise agreement, confirm the term in your own contract rather than from a card.
Orb compute: the meter most budgets forget
| Orb size | CPU / memory | Standard price per hour | Availability |
|---|---|---|---|
| a1.tiny | 1 CPU / 2 GB | $0.08 | All tiers |
| a1.small | 2 CPU / 4 GB | $0.17 | All tiers |
| a1.medium | 4 CPU / 8 GB | $0.33 | All tiers |
| a1.large | 8 CPU / 16 GB | $0.66 | All tiers |
| a1.xxlarge | 16 CPU / 32 GB | $1.32 | All tiers |
| a1.3xlarge | 16 CPU / 44 GB | $2.13 | Gigawatt subscribers and Enterprise members only |
Every size includes 60 GB of disk. Billing is by the minute and, in Amp's own words, "a paused orb costs nothing" — but read the pause rule before assuming that protects you. An orb pauses five minutes after the agent finishes or twenty minutes after your last interaction, whichever comes later. A session you walked away from mid-conversation therefore keeps billing for up to twenty idle minutes, and opening a portal or sending a message wakes it and restarts the meter (orb sizes and costs, September 19, 2026). This is the meter that no amount of BYOK touches: bringing your own key changes who bills your tokens and changes nothing about compute. Local runners are the lever there, and Amp states they cost nothing.
Amp Code BYOK: how it is wired, and the gate on it
Start with a correction, because the page that still ranks for this query is a reversal of itself. Amp's "No More BYOK" post from May 8, 2025 removed Isolated Mode and reads as a flat refusal; it now opens with an update saying Amp supports customer-managed model provider connections, including API keys for certain providers and ChatGPT subscriptions, set up on the Model Routing page. That reversal, not the 2025 headline, is current policy.
Configuration is server-side. There is no base-URL setting, no API-key setting and no model-provider environment variable anywhere in Amp's published CLI settings reference; keys and subscriptions are added in Amp's account settings under Model Routing, and workspace-wide under Workspace Model Routing. That follows from the product's shape: Amp's security page describes a multi-tenant cloud service at ampcode.com that brokers inference, and states plainly that Amp "doesn't offer self-hosted deployments, but it can use your own API keys for certain model providers." Your key lives on Amp's server and your endpoint has to be reachable from Amp's cloud — which is consistent with the early-access list naming Ollama Cloud rather than a local Ollama.
Two settings then interact. Model Routing decides which connection serves a model; Tune Modes, shipped September 10, 2026, decides which model a built-in mode's main agent, Oracle or subagents use. The Dial documentation puts it as "Tune Modes selects the model, while Model Routing determines which provider serves it."
The gate. An arbitrary custom endpoint URL is not generally available. Per the September 13, 2026 announcement, re-checked September 19, 2026, "Custom endpoint URLs" — along with OpenRouter, Amazon Bedrock, Google Cloud Agent Platform, Azure Foundry, Vercel AI Gateway, Cloudflare AI Gateway, Ollama Cloud and OpenCode Go — is early access for Megawatt and Gigawatt members only, behind an experimental feature named "More AI Routers & Subscriptions". Amp says it will roll this out to everyone soon and publishes no date. So on the dates checked, routing Amp through a third-party gateway costs at least $20 a month of Amp subscription before a single token. Outside that toggle, the only BYOK routes named in public documentation are direct keys for "certain model providers" — an Anthropic key is the one named in public docs, in Amp's own Claude Fable retention procedure — plus linked ChatGPT and Grok subscriptions. The full generally-available provider list sits behind the login-gated Model Routing page and is not published.
Four things BYOK does not buy, all documented by Amp itself. The separate code-review check and supporting system models such as image analysis, image generation and titling stay on Amp's own routing and credits — the Dial docs say outright, "Do not assume that pinning all three roles puts every model call on your subscription." A pinned model that becomes unavailable or is restricted by your workspace falls back to Amp's automatic model selection. With a linked ChatGPT subscription, ultra mode's main agent continues to use Amp credits. And non-model tools such as web search spend credits regardless. Data retention flips the other way and in your favour: requests sent with your key are covered by the retention terms of your provider account.
What Amp routes to by default, and what those tokens cost
This is the comparison a BYOK buyer actually needs, because Amp publishes its default routing and Amp credits bill at the provider's own API price. The modes page (September 19, 2026) sets out which model serves each role, so the table below puts the provider rate next to the Kunavo catalog rate for the same model. Rates are per million tokens.
| Amp role | Default model | Provider rate in / out | Kunavo catalog in / out |
|---|---|---|---|
| Low mode, main agent; Read Thread subagent | GLM-5.3-Flash / GLM-5.2 | Not applicable | No equivalent in the Kunavo catalog |
| medium mode, main agent | GPT-5.6 Sol | $5.00 / $30.00 | $2.00 / $12.00 |
| high mode, main agent; Oracle in low, medium and ultra | GPT-6 Astra | $10.00 / $50.00 | $4.00 / $20.00 |
| ultra mode, main agent; Oracle in high | Claude Fable 5.1 | $10.00 / $50.00 | $7.00 / $35.00 |
| Search subagent | GPT-5.6 Terra | $2.00 / $12.00 | $0.70 / $4.20 |
Read the provider column carefully. It is what Amp credits bill at on a tier with no discount, since Amp states it "does not add a markup to providers' API prices" for individuals and non-enterprise workspaces. Amp's September 13 post adds that on average, Megawatt members get a 60% discount and Gigawatt members a 65% discount on their Amp usage. What that average applies to cannot be settled from published material: the same post describes both orb usage and model usage as discounted for paid members, while the pricing docs say credits are charged at providers' API prices with no markup for individuals and non-enterprise workspaces. Amp publishes no base figure, no worked example and no split between the two meters, so the average cannot be turned into a per-token rate for your month — which is why the two columns above are the honest comparison and no discounted number is shown. Note also that Amp's low mode and its Read Thread subagent run GLM models that Kunavo does not carry, so a gateway route would not reproduce Amp's cheapest lane as configured.
A worked Amp Code API cost estimate
These figures are illustrative token arithmetic, not measured task costs and not a bill ceiling. Assume one session that sends 200,000 uncached input tokens and receives 25,000 output tokens on the main agent. Prompt caching, retries, subagent traffic and tool calls are all excluded, and orb compute is excluded on both sides because it is billed the same either way.
| Model | At the provider rate Amp credits use | At the Kunavo catalog rate |
|---|---|---|
| GPT-5.6 Sol | $1.750 | $0.700 |
| GPT-6 Astra | $3.250 | $1.300 |
| Claude Fable 5.1 | $3.250 | $2.275 |
| GPT-5.6 Terra | $0.700 | $0.245 |
Scale by your own sessions per day before treating this as a budget, and remember what the left column omits: Amp advertises a paid-tier discount but publishes no figure you can apply to it, and a paid tier costs $20 or $200 a month whether or not you use it. The right column omits Amp's subscription entirely but requires one anyway today, because the custom-endpoint route is gated to paid tiers. 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. The minimum Kunavo top-up is $10 in prepaid credit, which is a funding minimum rather than a task fee or a subscription — see billing details.
One warning about reading a table like this as a shopping list. The cheapest listed rate and the lowest cost to finish a task are different claims, and Amp makes the gap wider than usual because a model that needs more turns also burns more orb minutes. A cheaper model that reruns the work twice can cost more in total than an expensive one that lands it first time, and no price table can tell you which is which for your repository. Use the rates to build a shortlist, then run one bounded task on each candidate and read the charge your own account recorded — that number, not this one, is the comparison. AI cost optimization covers the method.
Which route wins when
| Route | Wins when | What you give up |
|---|---|---|
| Hobby plus Amp credits | Occasional use; you want one balance and no monthly commitment, at the provider's own rate with no markup | No included orb minutes, and none of the discount Amp advertises for paid tiers |
| Megawatt or Gigawatt | Steady daily use with real orb time; the included minutes apply, and Amp advertises a paid-tier discount on top | A fixed monthly charge, included usage that does not roll over, and an included credit amount you cannot see before subscribing |
| Linked ChatGPT or Grok subscription | One consumer plan already covers your volume | The provider's consumer account terms and data controls apply, and ultra mode's main agent still draws Amp credits |
| Direct provider API key | You want one vendor's own caching and batch terms, and that vendor's retention terms on your traffic | A separate account per vendor, and only "certain model providers" are named as generally available |
| A gateway behind a custom endpoint URL | You switch models per task and want one key and one balance | Early access on Megawatt or Gigawatt only as of the dates checked, behind an experimental toggle, with the wire protocol undocumented |
| Local runners | You have a machine that can run the work and want the compute meter at zero | Parallelism and the walk-away workflow that orbs exist to provide; model tokens still bill normally |
For the general version of the gateway row, what an AI gateway is explains what sits between a client and a model vendor, and choosing an LLM gateway compares the options on terms rather than on any one client's configuration.
Before you try to point Amp at a gateway
Kunavo publishes no Amp setup guide, and that omission is deliberate. Amp does not document what its custom endpoint field expects — whether OpenAI chat completions, OpenAI Responses or the Anthropic Messages protocol, whether the field takes a base URL or a full path, or how a model id is written — and the Model Routing page that would answer it is behind a login. Kunavo serves chat-completions, Responses and Messages endpoints, but which of those Amp would send to an endpoint you configure is not something anyone can state from outside, and no runtime test of Amp against Kunavo has been performed. Nothing on this page should be read as a compatibility claim.
What that means practically: if you want to try it, you need a Megawatt or Gigawatt tier, the "More AI Routers & Subscriptions" experimental feature enabled, and a willingness to discover the protocol yourself. Keep a working route available while you do, run one bounded task, then read the charge your account recorded for it. For the background on what an endpoint has to speak, see OpenAI-compatible API, and for how Amp's configurability compares with other agents' see the AI agent API directory. If a client with a documented and settled configuration path suits you better today, OpenCode pricing and Cline pricing cover two agents with a published Kunavo setup. Create a Kunavo account when you are ready to fund a key.
FAQ
How much does Amp Code cost?
Amp, the coding agent at ampcode.com, has a free Hobby tier and two paid individual tiers: Megawatt at $20 a month and Gigawatt at $200 a month. A Teams workspace carries no extra charge and Enterprise is custom-priced. Those are the software prices only. Three other meters run alongside them: orb compute, which is billed by the minute from $0.08 to $2.13 an hour depending on machine size; model tokens, charged to Amp credits at the provider's own API price with no markup for individuals and non-enterprise workspaces; and non-model tools such as web search, which also consume credits. Running agents on your own computer with runners costs nothing. Prices read from ampcode.com/pricing and ampcode.com/docs/orbs/sizes-and-costs on September 19, 2026.
Is Amp free?
Two different things share the name and only one of them is current. The free Hobby tier, announced September 13, 2026, is live: Amp's own post says it is "now free to use when you bring your own compute and model subscriptions/keys", and the Hobby card lists all product features, pay-as-you-go orbs or free local runners, BYOK, and no Amp token fees or limits. The older ad-funded "Amp Free" mode, launched October 15, 2025, is finished — its own post now carries "Update 4: Amp Free is paused. We don't intend to bring it back, but you never know." Search results still conflate the two, so check which one a write-up is describing before budgeting from it.
Does Amp Code support BYOK?
Yes, with a gate worth reading carefully. Direct API keys for certain model providers, plus linked ChatGPT and Grok subscriptions, are listed as features of Hobby, Individual and Teams, and Amp's pricing FAQ states there are no Amp token fees or limits on those tiers when you use your own subscription or key, with Enterprise carved out under custom pricing. But an arbitrary custom endpoint URL is not generally available: as of the September 13, 2026 announcement, re-checked September 19, 2026, custom endpoint URLs and gateway providers such as OpenRouter, Amazon Bedrock, Azure Foundry and Vercel AI Gateway are early access for Megawatt and Gigawatt members only, behind an experimental feature named "More AI Routers & Subscriptions". Amp says it will roll this out to everyone soon but publishes no date. Keys are configured server-side on the Model Routing page in Amp's account settings; Amp's published CLI settings reference documents no base-URL, API-key or model-provider setting and no environment variable for one.
What is the best API for Amp Code?
It depends on which meter dominates your month, because Amp splits the bill three ways. A linked ChatGPT or Grok subscription wins when one consumer plan already covers your daily volume and you accept that plan's consumer account terms. A direct provider key wins when you want one vendor's own caching and batch terms and the vendor's retention terms applied to your traffic. Amp credits win on simplicity, since Amp bills model usage at provider API prices with no markup for individuals and non-enterprise workspaces; Amp also describes paid tiers as carrying a discount on Amp usage, but states it only as an average, with no base figure and no split between orbs and model tokens. A gateway behind a custom endpoint URL wins when you switch models per task and want one balance — but that route requires a paid Amp tier and the experimental toggle today, and Amp publishes nothing about which wire protocol the endpoint field expects. Local runners are orthogonal to all of this: they remove orb compute cost, not model cost.
What is the cheapest API for Amp Code?
Cheapest listed rate and cheapest way to finish the task are different questions, and with Amp a third question sits on top of both: compute. A cheaper model that needs more turns burns more orb minutes as well as more tokens, and orbs are billed by the minute from $0.08 an hour for a1.tiny to $2.13 an hour for a1.3xlarge, so token arithmetic alone understates the difference. The lowest-cost configuration for most people is local runners, which cost nothing to run, plus whichever model finishes your work with review effort you accept. No provider should be picked as an absolute lowest price: run one bounded task, then read the charge your own account recorded for it.
How much do Amp orbs cost, and when am I being billed for one?
Amp's published standard prices are $0.08 an hour for a1.tiny (1 CPU, 2 GB), $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 (16 CPU, 44 GB), which is restricted to Gigawatt subscribers and Enterprise workspace members. Every size includes 60 GB of disk. Billing is by the minute and a paused orb costs nothing. An orb pauses five minutes after the agent finishes or twenty minutes after your last interaction, whichever comes later — so an idle session you walked away from keeps billing for up to twenty minutes. Opening a portal or sending a message wakes it and resumes billing. Read from ampcode.com/docs/orbs/sizes-and-costs on September 19, 2026.
Is Amp still a Sourcegraph product?
No. Amp spun out on December 2, 2025 — "Amp is becoming a separate company. We're spinning out of Sourcegraph to become an independent research lab." The legal entity named on Amp's own Terms of Service and Privacy Policy is Amp Frontier Corporation, and sourcegraph.com/amp now redirects to ampcode.com. Sourcegraph's own blog post about the split calls the new company "Amp Inc."; the first-party legal pages say Amp Frontier Corporation. Sourcegraph remains a separate live company selling code search, and its pricing has nothing to do with Amp's.
Amp's pricing page, orb cost table, docs pricing, security page, modes page, education page and its Free Agent, Amp Free, No More BYOK and spin-out announcements were read on September 19, 2026; the Gigawatt tier's price and minutes are client-rendered behind a tab and were read in a browser on September 18, 2026. The pricing page's FAQ answers sit behind an accordion and were read on September 18, 2026. Kunavo token rates come from the live catalog, and every dollar example here is illustrative token arithmetic rather than a measured task cost. No runtime test of Amp against Kunavo was performed.