Amp Review (2026): Sourcegraph's Agentic Coding Tool, Tested

Amp is a frontier agentic coding tool built by Amp Frontier Corporation, the company that spun out of Sourcegraph on December 2, 2025. It grew up inside Sourcegraph, the code-search company, as its first AI-native coding agent, and it became the practical path forward for individual developers after Cody shifted toward enterprise-only use. This review covers what Amp does, the models behind it, how much it costs in 2026, and how it stacks up against Cursor and Claude Code.

What Amp is

Amp is an agentic coding tool, not an autocomplete assistant. You give it a high-level instruction, and it plans the work, edits files across your entire codebase, runs commands and tests, reads the output, and keeps iterating until the task is finished. Unlike some earlier tools, it does not ask you to approve every edit step by step; the design philosophy is to let agents run and then verify the result.

A defining feature is where and how those agents run. Amp works as a command-line tool, an editor extension (VS Code and compatible editors), on the web with nothing to install, and on your phone. Agent sessions run in what Amp calls 'orbs' - remote machines that keep working even after you close your laptop - and you can start work on one device and pick it up on another. You can also dispatch multiple agents in parallel, which is where Amp puts a lot of its emphasis: several specialized agents working simultaneously on a large codebase.

At a glance

Pricing

Amp's pricing changed meaningfully in 2026. For most of its life it was strictly pay-as-you-go: you paid for model tokens at provider API prices, historically with no markup, and there was no subscription. That made it powerful but potentially expensive, and it earned a reputation as the premium option among coding agents.

In July 2026, Amp launched monthly subscriptions in beta, which are still available today:

A couple of details matter. The subscriptions include 'at least the subscription cost per month' in agent usage, and if you exhaust the included usage you either link a ChatGPT subscription or add paid credits to keep going. High and ultra modes, which lean on the priciest Anthropic models, generally require the Gigawatt plan or extra credits. There is no advertised ongoing free plan: you can sign up and run Amp on the web without installing anything, but usage is billed.

Models and modes

Amp does not lock you to a single model. It routes work across frontier models - Anthropic's Claude, OpenAI's GPT-5, and Google's Gemini - and uses cheaper models such as GLM for low-effort tasks. This is exposed through selectable modes (low, medium, high, ultra) that trade cost against reasoning power, so routine edits can run on cheap tokens while hard reasoning is reserved for the expensive models. If you link a ChatGPT or X Premium+/SuperGrok subscription, Amp can use those tokens with no additional per-token fee.

Strengths

The autonomy is real. Amp is comfortable planning a multi-step task, touching many files, running the test suite, and fixing what breaks, and users repeatedly describe it as feeling more genuinely agentic than tools that pause for constant approval. The parallel-agent and orbs model is the other standout: being able to fire off several agents that keep running on remote machines, then review their work with screenshots, videos, and live dev portals, is a different way of working than a single in-editor assistant.

Cross-model routing is a practical strength too. Because Amp is not tied to one provider, it can push cheap models for the easy parts and frontier models for the hard parts, which helps control cost without giving up quality. The tooling itself is polished and fast, and it is available in more places - terminal, editor, web, phone - than most competitors.

Limitations

The biggest caveat is cost and predictability. Pure pay-as-you-go at frontier-model prices can add up quickly on heavy workloads, and while the new $20 and $200 subscriptions help a lot, the cheaper plan restricts you to low and medium modes. There is no free tier to lean on indefinitely.

The editor experience is less mature than an IDE-first tool like Cursor, since Amp's center of gravity is the terminal and the web rather than a bespoke editor. The low-approval, let-it-run philosophy is efficient but can feel risky for teams that want tight control over every change. And as a newly independent company - the spin-out only happened in December 2025 - the product is evolving fast, which is exciting but means workflows and pricing can shift.

Pros and cons

Pros

• Genuinely agentic: plans, edits across many files, runs tests, and iterates
• Runs in the CLI, editor, web, and on mobile, with resumable sessions
• Parallel agents in remote orbs keep working after you close your laptop
• Cross-model routing across Claude, GPT-5, and Gemini with cost-tiered modes
• New $20 and $200 subscriptions make pricing far more predictable

Cons

• No advertised free plan; you pay by subscription or usage
• Frontier-model usage can get expensive on heavy workloads
• High and ultra modes need the pricier plan or extra credits
• Editor UI is less mature than IDE-first rivals
• Young standalone company with a rapidly changing product

How Amp compares

Against Cursor, Amp is less of an IDE and more of a cross-surface agent: Cursor wins on the in-editor experience, while Amp leans into terminal, web, and parallel autonomous agents. Against Claude Code, which is Anthropic's single-model command-line agent, Amp offers multi-model routing and its orbs-based parallel workflow, though Claude Code has a deep skills ecosystem and Anthropic's own model tuning. Against Cody, Amp is effectively the successor for individual developers now that Cody has moved toward enterprise code search and assistance. And against GitHub Copilot, which remains primarily an in-editor assistant, Amp is aimed at fully autonomous, multi-file agent work rather than inline suggestions.

Who it is for

Amp is a strong fit for professional developers and teams working in large, complex codebases who want an autonomous agent rather than autocomplete, and who are comfortable letting agents run with lighter supervision. Power users who like the terminal, want cross-model flexibility, and value parallel agents will get the most out of it. Developers who prefer a tightly controlled, IDE-native experience, or who need a permanent free tier, will find better fits elsewhere.

Verdict

Amp is one of the more ambitious agentic coding tools available in 2026: polished, fast, aggressively focused on autonomous parallel agents, and flexible across models and surfaces. Its early reputation as an expensive premium tool has softened considerably now that $20 and $200 subscriptions sit alongside pay-as-you-go pricing. If you want an agent that genuinely takes the wheel across your codebase and you are comfortable with usage-based or subscription costs tied to frontier-model quality, Amp is well worth testing.

FAQ

Is Amp the same as Sourcegraph Cody? No. Amp and Cody are now separate products from separate companies; Amp spun out of Sourcegraph in December 2025, while Cody remains Sourcegraph's code-search and assistance product.

Does Amp have a free plan? There is no advertised ongoing free plan. You can sign up and use it on the web, but usage is billed via subscription or credits.

Which models does Amp use? It routes across Claude, GPT-5, and Gemini, plus cheaper models for low-effort tasks, selectable through cost-tiered modes.