# Credits and billing

> Account credits, content usage, and automatic top-ups.

Your account balance pays for processed source input and readable memory returned.
Paid usage credits are offered for work, business, or organizational purposes,
including individual developers using Augenta commercially.
During rollout, billing applies only to explicitly enrolled accounts. **Credits and billing** in your user menu shows
whether billing and new purchases are enabled for your account. Earlier usage is never
back-billed. An account can spend granted credits while purchases are paused.

When Augenta's welcome offer is enabled, new users receive **$5 credit** after email verification.
The one-time grant belongs to your personal account, including if you also join a team.
No card is required to spend it. At the current source price this is equivalent to 5 million
source tokens; returned readable memory uses the same dollar balance at its lower rate.
It is not a fixed combined-token quota. Existing users are not automatically backfilled.

Open **Getting started**, choose **Create API key** and copy the first request. You can start
without creating an organization or assigning a Workspace. Save the key before leaving;
its secret is shown once. Credits and billing shows the exact available balance and your grant history.

| Customer content                    | USD per million tokens |
| ----------------------------------- | ---------------------- |
| Source input successfully processed | $1.00                  |
| Readable memory returned            | $0.05                  |

Each account has its own USD credit balance. An organization shares that balance across its Workspaces, plugin capture,
console content operations, and API keys, including one-shot Consolidate.
Workspace permissions still govern content access. Only organization owners and admins
can manage billing or read organization billing history.

## Complimentary usage

Augenta staff can designate an account for complimentary usage. No balance or card is
required; normal API limits continue to apply. **Credits and billing** shows the normal
value of settled usage and the amount actually charged. Complimentary work is charged $0.
Usage receipts expose `ratedAmount` for the normal value, `amount` for the charge, and
`complimentary: true`. These are customer prices, not our model-provider costs. Mode changes
apply to new work; accepted requests and sources retain their original treatment on retries.
Earlier unmetered usage is not reconstructed. Purchases and automatic top-ups are unavailable
while the account is complimentary.

## Add credits

Open the user menu at the top right of the console → **Credits and billing**. Purchase credits in $5 increments using the embedded
Stripe checkout. Applicable tax is shown separately; a $10 credit purchase grants $10
of usage credit. Credits appear only after confirmed payment. A redirect or a completed
browser form alone does not credit the account.

Optionally save cards and enable **Auto top-up**. Choose a positive balance threshold,
a purchase amount in $5 increments, and up to three cards in priority order. There are
no silently selected threshold or purchase defaults. Each crossing below the threshold
triggers one purchase, including enabling the rule while already below it. The rule
re-arms when the settled balance recovers to the threshold. Choose a refill amount
that restores sufficient headroom. Temporary reservations do not trigger top-ups.

The primary card is tried first, followed by backups once each. Required card
verification pauses the rule. If every card declines, the rule pauses. Update the
payment methods or add credits and save the rule to resume. Credits are never granted
for a pending or declined payment. Confirmed refunds/disputes adjust the ledger and
can leave insufficient available credit.

## Exactly what is measured

Billing counts the tokens in the text you send that Augenta reads, and in the memory text Augenta
returns: message text, each message step of a trajectory once, a document's full text, and returned
claim, summary and context text. Tool calls and results are never processed into memory, so they
are not charged either. Queries, metadata, receipts and Augenta's internal processing are not
charged.

Measurement details

Price version `content-2026-11-07-v1` uses `o200k_base` through `js-tiktoken@1.0.21`,
with a frozen vocabulary digest and `customer-content-v2` text-selection policy.
Each selected text leaf is encoded independently, without trimming, Unicode
normalization, special-token interpretation, or added framing.

The selected text is exactly the text memory extraction reads. Never selected:

* Tool activity: tool calls (names and arguments), tool results, denied tools, and
  `tool` or `error` trajectory steps.
* Opaque signed or encrypted content: `encrypted_content`, a `signature` whose value
  is an opaque token (a readable value such as a function signature is kept),
  redacted thinking, and armored blocks labelled `PGP SIGNATURE`, `PGP MESSAGE`,
  `SSH SIGNATURE`, `AGE ENCRYPTED FILE`, `ENCRYPTED PRIVATE KEY`, `PKCS7` or `CMS`.

Per input shape:

* Native messages: supplied message text before Augenta adds adapter framing. A
  `tool` role message, tool blocks and `tool_calls` are excluded.
* Trajectories: each message step's text once, as extraction reads it. Duplicate raw
  transcript telemetry is excluded.
* Documents: the complete decoded text before segmentation, overlap, or extraction
  retries. Source revisions that actually require reprocessing charge their full text.
* Output: returned claim/note, engram-summary, and memory-context text. Identical
  memory text of the same kind within one operation is counted once across its
  representations. A separate successful read is a new delivery.

Queries, metadata, acceptance receipts, raw-source echoes, opaque snapshots, internal
prompts, provider usage, embedding, re-embedding, and internal retries are excluded.
Generated answer text has no memory-output charge. Empty/noMatch output costs zero.
The existing `usage.tokensSent` and token-budget estimates are informational estimates,
not billing counts. There are no additional platform, storage, transport, compute,
answer/prose, discount, promotional-credit, or expiry charges in this version.

Usage measured before this version keeps its original price version,
`content-2026-10-23-v1` with `customer-content-v1`. That version counted tool names,
arguments and results, and every submitted trajectory step.

## Successful work and interrupted requests

Funds are reserved before billable work or output is admitted. Available credit is
the remaining balance minus reservations; concurrent requests cannot spend the same
available dollars.

Durable input settles when processing succeeds, even after a caller disconnects or
the surrounding Consolidate operation fails later. Terminal failure, lifecycle
cancellation, unsupported input, and superseded/conflicting document writes release
the input reservation. Explicitly re-driving failed work reserves its input again.

Stateless input and returned memory settle after the serving runtime completes the
authoritative response. Provisional output is not an additional charge. Failed or
incomplete responses release their output holds. Server completion is not proof of
client receipt; unknown completion after a crash is released rather than charged.

## Top up and retry

`402 insufficient_balance` means add credits and explicitly retry. The error is not
automatically retryable: repeatedly sending it cannot fund the account. An automatic
top-up may be pending; wait for payment confirmation before retrying. Streams report
the equivalent error in an SSE error event after their HTTP headers have been sent.

Supply a stable `Idempotency-Key` header for the operation, including stateless Experiences
and `record: false` Recall when retry protection is needed. Retry explicitly with the same
key and unchanged content. Changed content under the same key is a conflict.

Durable ingestion binds the key, authenticated destination, and request fingerprint
in the intake transaction. An unchanged retry returns the original acceptance and
source receipt references without admitting another job. Accepted work keeps its
reservation through processing even if the HTTP client disconnects; these retries
do not consume the stateless execution allowance.

Large documents may still be undergoing exact preflight measurement. A retryable 503
with `Retry-After` means retry the same request and key; no source has been admitted
or reserved yet. Completed measurements reuse counts and hashes instead of decoding
the same source again.

Stateless billing stores counts, hashes, and receipts, without source or result bodies.
A same-key retry may recompute a different result, with no duplicate debit for the
operation. Up to three executions are allowed within 24 hours. The settlement record
remains afterward; an expired key is refused instead of being silently charged again.
`billing_retry_exhausted` names that limit. This is not a global content-deduplication
or exact-response-replay promise.

Enabled responses carry an operation reference in `X-Augenta-Billing-Operation` and
final `billing` metadata. Billing counts and amounts refer to the original receipt on
a replay; `measuredMemoryTokens` describes the current reconstruction. Durable
acceptance responses include source `billingOperationId` references, whose settlement
can be inspected in the owner's/admin's billing history. Monetary API values retain
nine decimal places; per-request charges are not rounded to cents.

Deleting a Workspace removes its source paths and acceptance details from billing
receipts. Financial entries, token counts and request-key hashes remain for audit
and duplicate-charge protection; deleted request keys cannot be reused.

The Credits page provides purchase status, receipt/invoice links, reservations,
and basic usage receipts. Aggregate usage administration is tracked separately.
