Agent.Space combines optional model plans with a credits balance. Use a plan when its covered models and recurring allowance fit your normal workload; use pay-as-you-go credits for occasional work, models outside the plan, or continued use after the allowance runs out.
If you arrived through an older Share/Flex reference, use the current plan and credit rules below rather than the old assumption that those funding paths never cover continued usage.
Checked September 29, 2026. This guide uses the public Agent.Space product description, pricing page, and Developer API page. It explains documented rules, not an authenticated account test or a guarantee that one option costs less for your workload.
Plans and credits at a glance
The unit is straightforward: one Agent.Space credit represents US$1. That does not mean one credit buys the same number of requests from every model. Token rates, context length, repeated attempts, and the selected channel affect consumption.
Understand the order before starting a long task
Current product rules use eligible plan allowance first. When that allowance is exhausted, requests automatically continue on pay-as-you-go credits; requests outside the plan are also metered.
Three consequences matter:
- A plan is not necessarily a hard spending cap. A long task can use allowance and then credits.
- The harness name does not determine coverage. Choosing Codex or Claude Code is separate from selecting the model and channel.
- A successful request does not prove which funding source paid. Inspect allowance and request records rather than assuming every call was included.
The current public plan example is Codex Dedicated. It provides supported OpenAI model access through Agent.Space, not a shared OpenAI login, an official ChatGPT subscription, or a transfer of funds into an OpenAI account. Check the live catalog before assuming another model family has an equivalent available plan.
For unattended runs, separate the funding choice from Agent spending limits and run budgets. An allowance ending, a notification, and a hard stop are not the same event.
Separate billing periods, allowance resets, and credit expiration
These are three different clocks.
Plan billing determines when you pay for the subscription. Annual billing is paid upfront; it does not turn the whole year into one unrestricted allowance pool.
Allowance resets currently use fixed seven-day cycles. Unused allowance does not roll forward. Compare a normal cycle with a busy one, not just a monthly average that hides concentrated work.
Credit expiration depends on where the credits came from:
Do not generalize the top-up rule to every number in the wallet. Before buying, inspect the source and expiry of the credits you expect to spend.
Credits can also be used to purchase a model plan. When comparing costs, do not count the same money once as a top-up and again as a plan purchase. Separate cash paid, credits consumed, and credits still available.
Which route fits your workload?
Occasional or uncertain use
If you are still evaluating a supported model or only have a few bounded tasks, pay-as-you-go can avoid committing to recurring allowance you may not use. It is not automatically cheaper: begin with a small task and review actual consumption before extending it.
Regular work on covered OpenAI models
A Codex Dedicated plan is worth evaluating when its current coverage and seven-day allowance fit your regular work. Compare the fixed commitment plus any extra metered usage against what the same workload would cost without the plan.
A busy day is not enough evidence to choose a larger recurring tier. Check whether the pattern repeats and whether you can use the allowance before each reset.
Work across model families
Combine eligible allowance with metered credits when different tasks genuinely need different available models. Check the selected model and channel for each task; a plan for one family does not make every other family included.
For example, implementation and review might use different models. That is a planning example, not a claim that switching models always improves quality or reduces cost.
Local API access is a separate setup choice
You can use supported local clients or a server integration without moving your project into the cloud Workspace. External API access currently requires an active plan or purchased credits; signup bonus credits alone do not satisfy that eligibility requirement.
Follow the Developer API setup guide for current model discovery, client configuration, and request-history checks. A key authorizes a model request; it does not migrate project files, grant an official provider subscription, or establish that every tool feature is compatible.
If you need project files, Sessions, and handoffs to stay together in the cloud, how Agent.Space works explains that additional layer. Funding a request and preserving a project are separate decisions.
Before paying, check five things
- Coverage: the exact model and channel, not just the harness or plan name.
- Timing: the billing commitment, next allowance reset, and relevant credit expirations.
- Continuation: the rate and funding source after allowance runs out.
- Evidence: a bounded task, its request record, and an accepted result.
- Control: when you will stop and review usage before approving more work.
For longer tasks, also review the persistent coding Agent Sessions guide. A plan funds model use; it does not guarantee that a process runs forever.
Check current Agent.Space plans and rates, choose the smallest suitable starting point, and revisit the decision using actual usage. The useful comparison is fixed commitment plus consumed extras for work you accept—not the cheapest headline price.