Amp Pricing Explained (2026): Credits, Megawatt & Gigawatt Cost

Amp bills on pay-as-you-go credits, not per seat. You top up from a $5 minimum and spend credits as the agent runs, with zero markup for individuals and non-enterprise workspaces — you pay the underlying LLM API cost and nothing on top. If you'd rather have included usage, two optional plans layer on: Megawatt at $20/mo and Gigawatt at $200/mo. Here's how each part works and who it's cheapest for.

How Amp pricing works

Amp charges for what the agent actually spends on model calls, not for the number of people using it. There's no subscription required to start, and credits are pooled across everyone in a workspace.

Amp is Sourcegraph's agentic coding agent, and it runs across a CLI, the web, and editor extensions for VS Code, JetBrains, Neovim and Zed. Unlike seat-priced rivals, its whole cost model is credit-based: you buy credits, the agent burns them as it works, and you top up when you run low. New users get a small free credit grant — around $10 — to try it before spending anything. We put the tool through its paces in our Amp review.

Pay-as-you-go credits

The core model is dead simple: each Amp thread deducts credits equal to the real underlying LLM API cost, with no markup for individuals and non-enterprise workspaces. The minimum top-up is $5.

Say a thread incurs $2 in Anthropic usage plus $0.50 in OpenAI usage — Amp deducts $2.50 from your balance, matching the provider pricing exactly. Because Amp passes token pricing straight through, you're never paying a platform premium on top of the model. Credits are pooled across all members of a workspace, so a team shares one balance rather than paying separately, and unused credits expire only after 1 year of inactivity.

That structure is the opposite of a fixed monthly seat fee. Light weeks cost little; a big refactor week costs more. The trade-off is that your bill tracks your usage rather than sitting at a predictable flat number.

Subscriptions: Megawatt vs Gigawatt

The two optional monthly plans aren't a markup — each one gives you at least its cost back in included agent usage, plus extra perks like orb hours and access to more expensive modes. Megawatt is $20/mo; Gigawatt is $200/mo.

"Orbs" are async cloud agents that run tasks in the background, and the plans bundle a large block of orb hours. Modes control how much the agent spends per run — Low is cheapest, then Medium, High and Ultra — and the higher plan unlocks the pricier modes. Here's every plan side by side.

Plan	Price	Included usage / orbs	Modes
Pay-as-you-go	$5 min top-up	No included usage; pay the raw API cost with zero markup. ~$10 free credit to start.	Per your top-up / plan
Megawatt	$20/mo	$20 of agent usage included · 750 hours of small orbs	Low & Medium only
Gigawatt	$200/mo	$200 of agent usage included · 1,000 hours of large orbs	All modes (High & Ultra included)
Enterprise	Usage-based (~50% higher)	Custom · BYO inference keys, SSO, zero LLM data retention	All modes

Megawatt suits a steady individual who wants ChatGPT or X Premium+ subscriptions linked in and is happy on the cheaper modes. Gigawatt is for heavier users who need High and Ultra and lean on large orbs. Because the included usage matches the price, subscribing is essentially free relative to pay-as-you-go if you'd spend that much anyway — you just get the extra orb hours and mode access on top.

Enterprise

Enterprise is still usage-based, but priced roughly 50% higher than individual and team pricing in exchange for controls that larger orgs need.

You get bring-your-own inference keys, SSO and directory sync, zero data retention by the LLMs, plus custom controls, analytics and support. If your compliance team requires those guarantees, the premium is the cost of routing usage through your own governed setup rather than the default shared pipeline.

How to keep costs down

Because you pay the real model cost, the biggest lever is the mode you run. Drop to Low or Medium for routine edits and reserve High or Ultra for genuinely hard problems — the cheaper modes cut credit burn substantially.

Beyond that: keep threads focused so the agent isn't re-reading huge contexts, pool credits at the workspace level so a team shares one balance instead of over-provisioning, and match the plan to your real usage. If you'd reliably spend $20 a month anyway, Megawatt hands that back as included usage plus orb hours. If your usage is spiky, staying pure pay-as-you-go avoids paying for a subscription you don't fill.

Is Amp worth it?

Amp is worth it if your usage is variable or spiky, or you run a team that doesn't want per-seat lock-in. Zero markup and pooled credits mean you pay for actual work, not headcount, which is hard to beat for bursty projects.

The catch is predictability. A flat $20 tool like Cursor or Claude Code Pro gives you a fixed bill; Amp's can swing with your workload, and heavy daily users may spend well over $20–$100/mo. If you code with an agent all day and want a known number, a subscription tool can be calmer. For everyone else — occasional, team-shared, or unpredictable use — Amp's credit model is genuinely cheaper. See how the two stack up in Amp vs Claude Code and the wider best AI coding assistants roundup, or check the current rates on the official Amp pricing page.