Developers and small software teams can use an eligible ChatGPT subscription in Amp, the AI coding agent, without paying for an Amp monthly plan. Amp removed that requirement in its September 13, 2026, Free Agent announcement.
That does not make every Amp task free. Your plan can pay for supported OpenAI model work; some tools and Amp-hosted machines still cost extra. The practical choice is where the work runs: on your own computer or Amp’s cloud machines.
What your subscription covers
OpenAI lists Amp as a partner for eligible ChatGPT Plus and Pro subscribers. Usage draws on the Codex/ChatGPT Work allowance already in the plan, not a new pool. Signing in and authorizing Amp to consume that allowance are separate permissions.
Work or service | How it is paid for |
|---|---|
Supported OpenAI model requests | ChatGPT allowance, when served through the connected subscription |
Model requests served by Amp | Amp usage or credits, not your ChatGPT allowance |
Painter image generation, web search, webpage reading and voice | |
Amp-managed cloud machines, called orbs | Amp orb allowance or credits; not covered by ChatGPT |
A runner on your own machine | No Amp runner fee; you supply and operate the machine |
Linking ChatGPT is not enough
Linking your account does not change Amp’s selected models. Apply ChatGPT Only in Tune Modes and save to route the main agent, its Oracle adviser and subagents—helper agents—to OpenAI models in low, medium and high. The ultra mode keeps Amp’s own routing. Check the Model Routing preview to see which connection serves each model.
Do not judge the whole bill by the main model’s name. Amp’s Raw Models picker changes only the main agent; other agents can still use separately billed models. Custom plugin modes have separate settings, and supporting system work also needs checking. The preset is not a promise that every call is included.
Choose where the work runs
Use your own machine when you can supply the computing capacity. Amp’s free Hobby tier supports own-machine runners. A runner is an Amp instance on a computer you provide, which can receive tasks started from Amp’s website. You remain responsible for that computer’s availability and operating costs. This does not mean the ChatGPT model runs locally or offline.
Use an orb when the work must continue independently of your computer. Amp documents orbs as managed remote machines that can keep working while a laptop is closed and retain their environment while paused. You are paying for a place to run the task; this does not expand your ChatGPT model allowance.
A cloud-cost example, not a savings claim
Amp’s published standard rates include $0.17 an hour for a1.small, with two CPUs and 4 GB of memory, and $0.33 an hour for a1.medium, with four CPUs and 8 GB.
Illustrative calculation: 100 aggregate billed hours cost 100 × $0.17 = $17 on small machines, or 100 × $0.33 = $33 on medium machines. This assumes one size throughout, pay-as-you-go standard rates and no included allowance or discount. It excludes subscriptions, model and tool usage, taxes and other fees. Equal hours do not imply equal completed work.
Count every orb’s runtime, including idle time. Billing is by the minute and stops while paused. Automatic pausing requires both five minutes without agent work and twenty minutes without interaction. Opening an unarchived thread can wake it and restart billing; Orb Saver Mode prevents that particular trigger.
The Individual Megawatt listing advertises $20 a month and 45,000 orb minutes. But larger machines consume the allowance faster. The checked pages do not establish the size conversion needed to calculate a reliable break-even point against pay-as-you-go use. Do not treat the advertised minutes as interchangeable across sizes.
The ChatGPT allowance is still shared
In ChatGPT Settings → Usage, you can cap Amp at a percentage of your overall weekly usage. A cap does not reserve that capacity: other connected apps can still consume the shared allowance. Check remaining usage before assigning a long task.
For Plus, OpenAI says the five-hour usage limit is shared across apps. That five-hour limit does not apply to Pro, but this is not unlimited usage. Weekly limits can still apply, and consumption varies with the model, task, context and tools. No guaranteed number of Amp tasks per plan is established here.
If a request hits a limit, check both Amp’s app cap and the overall plan balance. OpenAI’s integration documentation says an app-specific cap can cause a usage-limit error. Amp’s exact behavior after subscription exhaustion was not verified; do not assume it automatically switches to paid Amp credits.
Check a real thread before choosing a larger plan
Amp documents a per-thread cost breakdown behind the dollar amount in the sidebar, with balance checks in settings or through the amp usage command. After a representative task of your own, compare that breakdown with ChatGPT’s usage display. A low Amp charge alone will not tell you how much shared subscription capacity the task consumed.
For an existing subscriber with suitable hardware, Hobby plus an own-machine runner is a starting point that avoids an Amp monthly fee and hosted-compute charges. Choose cloud execution when independence from that hardware is worth paying for. If you need a new ChatGPT subscription or an upgrade, include that expense in the comparison—not just the Amp bill.
Comparing coding-tool bundles? See our related guide to what X’s subscription includes in Cursor. Its billing terms are different; do not transfer those inclusions to Amp.
Methodology: Prepared with AI assistance from Amp and official OpenAI documentation, rechecked October 14, 2026. This is a documentary billing guide, not a hands-on review. No customer spending or savings was measured; account eligibility and tool-specific charges were not independently tested.