Amp Review 2026: Pricing, Orbs, Models & Trade-offs

Disclosure: CodeToolPilot may earn a commission if you buy through links on this page, at no extra cost to you. How we make money.

Last verified: October 12, 2026. This is an editorial research review — how we review. Prices and plans change often; confirm on the vendor’s site.

Status update: Amp has been an independent company (Amp Inc.) since Sourcegraph spun it out in December 2025. Amp restructured its pricing on September 13, 2026.

Amp is a coding agent and development environment from Amp Inc., the company Sourcegraph spun out as an independent business in December 2025. You can run it from a command-line tool on your own machine, in a web app at ampcode.com, or in macOS and iOS apps. Its signature idea in 2026 is the “orb”: a cloud machine Amp creates for each thread, with your code, tools and plugins already set up, so agents keep working after you close your laptop.

Amp is aimed at developers and teams who want to hand larger pieces of work to several agents in parallel and can live with an opinionated product that changes quickly. Since September 2026 the agent itself is free if you bring your own compute and model access, such as a ChatGPT or Claude subscription or API keys. Amp charges for orb time, optional monthly tiers and pass-through model credits.

This is an editorial research review based on Amp’s documentation, pricing page and news posts, plus press coverage of the Sourcegraph spin-out. We did not run our own benchmarks.

Quick verdict

Amp is one of the most ambitious agents in this category. It routes work across several frontier models behind a simple four-position “Dial”, uses specialist subagents, and runs agents on cloud orbs with live previews. It fits teams that want parallel, cloud-hosted agents without tying themselves to one model vendor. The price is predictability: billing spans orb minutes, model credits and linked subscriptions, pricing has changed several times in 2026, and Amp says plainly that it drops features without keeping backward compatibility.

Pro: A free Hobby tier with no Amp token fees when you bring a ChatGPT or Claude subscription or your own API keys.
Pro: Multi-model routing through the Dial (low, medium, high, ultra), plus Oracle, Librarian and Search subagents.
Pro: Orbs give each thread its own cloud machine. They are billed by the minute and pause when idle, or you can run agents on your own runners for free.
Pro: Teams at no extra charge, with free SSO once one member is on a paid tier. Amp’s zero and minimal data retention policy now covers all users.
Con: Several meters (orb minutes, model credits, linked subscriptions) make the total cost harder to forecast.
Con: Opinionated and fast-changing. The old modes were replaced in July 2026, and Amp promises “no backward compatibility, no legacy features.”
Con: A thread’s mode is fixed after the first message, so changing modes means starting a new thread.
Con: The CLI runs on Windows only through WSL, and the docs list no IDE extension.

Visit Amp

What is Amp?

Amp began as a Sourcegraph product. In December 2025 Sourcegraph spun it out as Amp Inc., led by Sourcegraph co-founder Quinn Slack, while Sourcegraph itself refocused on code search. Amp now calls itself “an independent agent research lab.” It also states who it is not for: “If you’re looking for a product that will stay the same and check every procurement box, this isn’t it.”

The same agent and the same threads are available on every surface. The web app needs no install and runs work in an orb. The macOS and iOS apps add notifications, dictation and sharing. The CLI runs Amp on your own machine, in your editor’s terminal or from scripts, and installs with a one-line script on macOS, Linux and WSL or via Homebrew. Puck is an agent that launches and coordinates other agents for you, including by real-time voice, and Amp also runs in Slack.

Key features

The Dial and model routing

Instead of choosing a model, you choose how hard the task is. low is for small, well-defined tasks, and medium is the default for most work. high is for hard changes where a subtle miss is expensive, and ultra is for open-ended work like migrations and architecture. Amp picks the model, reasoning effort, prompt and tools for each mode and changes them as models improve. Per Amp’s news posts, Claude Opus 5.5 has powered medium since September 28, 2026, and Claude Fable 5.1 has powered ultra since September 1. Since September 10 you can pin your own models per mode or add custom agents.

Specialist subagents

The main agent can delegate to subagents with their own context windows: Search finds code, Oracle handles hard reasoning and second opinions, Librarian researches external codebases, Task carries out delegated coding and verification, and Read Thread summarizes other threads.

Orbs and runners

Orbs are per-thread cloud machines. Sizes run from a1.tiny (1 CPU, 2 GB) at $0.08/hour to a1.3xlarge (16 CPUs, 44 GB) at $2.13/hour, each with 60 GB of disk. Billing is per minute, and an orb pauses 5 minutes after the agent stops working or 20 minutes after your last interaction, whichever comes later. Portals expose running apps for preview. If you prefer your own hardware, runners let Amp agents run on your machines, including your Mac, for free.

Bring your own subscriptions and keys

You can sign in with ChatGPT so OpenAI model usage bills to your ChatGPT plan. Since October 10, 2026, you can also link a Claude Pro or Max plan and pick a “Claude Code” mode, which runs on Anthropic’s Claude Agent SDK inside Amp’s orbs and threads. Since September 30, all tiers can use BYOK providers including OpenRouter, Amazon Bedrock, Google Cloud Agent Platform, Azure Foundry, Vercel AI Gateway, Cloudflare AI Gateway, Ollama Cloud and custom endpoints.

Collaboration

Workspaces add shared threads, multiplayer sessions, shared portals, plugins and skills, analytics, pooled credits and SAML/OIDC SSO. Members can sit on different tiers, so a few power users can pay while others stay on Hobby.

Extensibility and automation

Amp supports plugins, skills, MCP servers, AGENTS.md files, an SDK and API, an execute mode and streaming JSON output for scripts. Orbs can be scheduled or triggered by events.

Who should use Amp

Amp fits developers and teams who want several agents working in parallel on cloud machines. It also suits teams that already pay for ChatGPT or Claude and want to reuse those plans in a multi-model agent, and organizations that want collaboration features without per-seat fees.

Who should skip it

Skip Amp if you need a stable product with long-lived features and a single, predictable bill, or if you want a deep IDE extension. It is also a weaker fit if you work mainly on native Windows without WSL, or if your policies do not allow agents to run code on a vendor’s cloud machines.

Strengths and limitations

Strengths

Vendor-neutral routing. Amp blends Anthropic and OpenAI models behind stable modes, lets you pin your own, and reuses subscriptions you already pay for.
Built for parallel, remote work. Orbs, portals, scheduled runs and mobile apps support delegating many tasks and reviewing results later.
Low cost to adopt for teams. Free Hobby seats and a no-extra-charge Teams tier make it easy to trial.
Transparent orb pricing. Published per-hour rates by machine size and clear pause rules make compute spend easy to reason about.
Data handling. Amp has extended its zero and minimal data retention commitment to all users, not just Enterprise.

Limitations

Cost is spread across meters. Model credits, orb minutes, linked subscriptions and tool usage such as web search are billed separately. Some subagent roles bill to Amp credits even when your main agent runs on a ChatGPT subscription.
Constant change. Amp replaced its smart, deep, rush and large modes in July 2026, launched subscriptions on July 18 and restructured pricing again on September 13. Teams that need stable workflows and procurement terms should expect to re-check docs often.
Mode locked per thread. A thread’s mode is fixed after the first message. Amp says this protects quality and caching, but you must judge difficulty up front.
Platform gaps. The CLI supports Windows only through WSL. The current docs describe web, macOS, iOS and CLI clients but no IDE extension.
Cloud-first by design. Orbs run your code on Amp’s machines unless you set up runners, so security teams will want to review secrets handling and network controls. Using Claude Fable models with your own Anthropic key also requires enabling data retention on that key’s workspace.

Pricing (as of October 2026)

Plan	Price	What you get
Hobby	Free	All product features. Pay as you go for orbs, or use your own runners free. Bring ChatGPT or Claude subscriptions or API keys with no Amp token fees or limits.
Megawatt	$20/month ($10/month for students and teachers)	45,000 minutes of orb time, included agent usage, unlimited public and private repositories, and custom domains for portals.
Gigawatt	$200/month	A larger included allowance, plus access to the largest `a1.3xlarge` orb size.
Teams	No extra charge	Members on any tier, shared billing and pooled credits, team permissions, analytics and SAML/OIDC SSO.
Enterprise	Custom	Pooled credits with no paid seats, SCIM, spend limits, audit logs, network and MCP controls, IP allowlisting and regional endpoints.
Paid credits	Provider list price, no markup (individuals and non-enterprise workspaces)	Model and tool usage beyond your included amount. Web search also uses credits, and purchased credits expire after 12 months.

Amp says Megawatt and Gigawatt members get average discounts of 60% and 65% on their Amp usage. Included usage resets monthly and does not roll over. The Gigawatt price appears when you select that tier on the pricing page. Confirm current figures on Amp’s pricing page before subscribing, since the model has changed several times this year.

See Amp pricing

Prices verified on October 12, 2026 against the official Amp pricing page. Plans change often — confirm on the vendor’s page before you buy.

Use cases

Amp is built for handing work to several agents at once, often on cloud machines, while reusing model access you may already pay for. The scenarios below follow from its documented features, and each one notes the caveat that matters most for that kind of work.

Running several scoped tasks in parallel overnight

A tech lead can split a backlog of well-defined chores, such as dependency bumps, missing tests or small refactors, into separate threads that each run on their own orb. Orbs keep working after you close your laptop, and portals expose running apps for preview, so results can be reviewed later from the web or iOS app. Orb time is billed by the minute until the machine pauses, so this pattern suits teams that watch that meter or point Amp at their own runners instead.

Reusing an existing ChatGPT or Claude plan in a multi-model agent

An individual developer who already pays for ChatGPT, or for Claude Pro or Max, can link that plan to Amp’s free Hobby tier and avoid Amp token fees. Bring-your-own-key providers such as OpenRouter, Amazon Bedrock and Ollama Cloud are also available on all tiers. The caveat is that some subagent roles still bill to Amp credits even when the main agent runs on a ChatGPT subscription, so keep a small credit balance.

Planning a migration or architecture change

For open-ended work such as a framework migration, Amp’s docs point to the ultra mode, and the main agent can hand hard reasoning to the Oracle subagent and research on external codebases to the Librarian. In our assessment, this is where Amp’s routing is most useful, because the developer picks the difficulty instead of tracking which model is strongest this month. A thread’s mode is fixed after the first message, so decide up front whether a task needs ultra or a cheaper mode.

Scheduled or event-driven maintenance jobs

Platform teams can script Amp through its SDK and API, execute mode and streaming JSON output, and orbs can be scheduled or triggered by events. That makes recurring repository chores a reasonable fit, such as a weekly dependency check that opens a thread for review. Orbs run your code on Amp’s machines unless you set up runners, so have your security team review secrets handling and network controls before these jobs get credentials.

A team where only some members need paid usage

Amp workspaces carry no extra charge and let members sit on different tiers, so a few heavy users can pay while others stay on Hobby. Shared threads, multiplayer sessions, pooled credits and SAML/OIDC SSO are included, and SSO is free once one member is on a paid tier. Organizations that need SCIM, audit logs or spend limits will have to move to Enterprise.

Alternatives

Claude Code: Anthropic’s agent with broad terminal, IDE and cloud coverage and strong admin controls, limited to Claude models.
OpenAI Codex: an open-source CLI, IDE extension and Codex Cloud, included in ChatGPT plans.
Cline: an open-source, model-agnostic agent for VS Code, JetBrains and the terminal that uses your own keys.
Antigravity CLI: Google’s terminal agent, which replaced Gemini CLI for individual users in June 2026. It is not yet reviewed on CodeToolPilot.

Final verdict

Amp is a forward-leaning agent for teams that want parallel cloud agents and multi-model routing, and its free Hobby tier and bring-your-own-subscription model make it cheap to try. The trade-off is a product and pricing scheme that change quickly, with costs spread across several meters. If that pace and cloud-hosted execution suit you, Amp belongs on your shortlist.

Visit Amp

FAQ

Is Amp free?

Yes, with conditions. Since September 2026, the Hobby tier includes all product features and charges no Amp token fees when you bring a ChatGPT or Claude subscription or your own API keys. You still pay as you go for orb time unless you run agents on your own runners, and tools such as web search draw on paid credits. Paid monthly tiers add included usage and orb time, so check Amp’s pricing page for current figures.

Is Amp still part of Sourcegraph?

No. In December 2025, Sourcegraph spun Amp out as an independent company, Amp Inc., led by Sourcegraph co-founder Quinn Slack, while Sourcegraph refocused on code search. Amp now describes itself as “an independent agent research lab.”

Does Amp train models on my code?

Not unless you opt in, according to Amp’s security page: neither Amp nor its subprocessors train models on your data without explicit opt-in. On a workspace, turning training on requires admin approval, and it can never be enabled on Enterprise workspaces. Minimal data retention applies to all users, and a thread’s data is removed within 30 days of deleting it. If you link a personal ChatGPT subscription, the provider’s consumer terms and data controls apply instead, so turn off training on that account before linking it.

Does Amp work in VS Code or JetBrains IDEs?

Amp’s current documentation lists web, macOS, iOS and CLI clients but no IDE extension. You can run the CLI in your editor’s integrated terminal on macOS, Linux, or Windows through WSL. If an in-editor panel matters to you, Cline works inside VS Code and JetBrains, and Claude Code covers both the terminal and IDEs.

Can Amp run on my own machine instead of Amp’s cloud?

Yes. The CLI runs Amp locally and installs on macOS, Linux and WSL with a one-line script or via Homebrew. Runners let Amp agents execute on your own hardware, including a Mac, at no charge, instead of on paid orbs. Teams whose policies forbid running code on a vendor’s cloud machines should plan around runners from the start.

Amp vs Claude Code: which should I choose?

Claude Code is Anthropic’s own agent, with broad terminal, IDE and cloud coverage and strong admin controls, but it is limited to Claude models. Amp routes work across Anthropic and OpenAI models and adds per-thread cloud orbs. Since October 2026, it can also run a “Claude Code” mode on a linked Claude Pro or Max plan. In our assessment, Claude Code is the steadier choice for teams that want one vendor and deep editor integration, while Amp suits teams that want multi-model routing and parallel cloud agents and can live with frequent product changes.

Amp Review (2026): Multi-Model Coding Agent With Cloud Orbs