First identify who bills the request
| Route | Credential and billing source |
|---|---|
| Personal Claude subscription | A personal Claude login, governed by the subscription’s current rules and limits. |
| Direct provider API | A provider-issued API credential, billed by that provider under its API terms. |
| Superfastcode gateway | A Superfastcode gateway key, charged under your gateway channel rates and plan. |
Check the base URL and credential source in Claude Code’s status view. With a gateway credential active, a saved personal login is not evidence that your subscription is paying the model request. Verify the corresponding gateway usage record.
For current official subscription and API prices, consult Anthropic’s pricing. For this gateway’s rates, use Superfastcode Claude Code pricing; a provider’s headline price is not automatically your gateway price.
Calculate token charges without double-counting
cost = (uncached_input × input_rate
+ cache_write × write_rate
+ cache_read × read_rate
+ output × output_rate) / 1,000,000The four counts must be disjoint. If a client reports total input including cached tokens, first separate cache reads and writes from ordinary input according to that usage schema. Do not bill the same cached token twice in an estimate. Billed output can include reasoning tokens even when they are not displayed as final text.
Select the exact model, channel, context tier and cache duration. A long-context threshold can change rates for a request; a one-hour cache write can use a different rate from a five-minute write. The current pricing page documents model-specific rules. Tool charges and other separately priced features need their own line items.
Try your own token-cost estimate
Copy the rates for your exact model, channel and context tier from the Claude Code pricing page. Rates start blank so this calculator never silently uses an outdated price.
| Token bucket | Token count | USD / 1M tokens |
|---|---|---|
| Uncached input | ||
| Cache write | ||
| Cache read | ||
| Output (including billed reasoning) |
An estimate of token charges, not a checkout price or guaranteed deduction. Use disjoint counts: exclude cache reads and writes from uncached input. Enter 0 for a bucket that does not apply. Select the correct cache duration; special tools, context tiers, channel rules and actual billing records can change the total.
Separate API Credits from the cash you paid
The model catalog expresses metered rates in USD per million tokens. A plan or top-up can grant a different amount of API Credits from its purchase price. Compare three numbers separately: cash paid, credits actually granted, and usage deducted.
For a hypothetical $20 payment granting 100 usage credits, spending 10 of those credits allocates $2 of that purchase when all 100 are used. If only 20 credits are used before expiry, the $20 was still paid. That is why dividing a monthly price by the maximum advertised grant can overstate savings when weekly credits go unused.
Check the actual grant, release schedule, reset time, carryover, wallet rules and expiration in your chosen plan and Console. Credits are prepaid service credits; refund eligibility follows the Terms and checkout terms, not this estimate.
Reduce cost while checking useful output
- Start new tasks with relevant context rather than repeatedly pasting an entire repository.
- Use a smaller available model for bounded tasks when it passes your quality checks.
- Measure cache hits from actual usage records; a repeated prompt does not guarantee a hit.
- Inspect output and retry counts. A low per-token price can still cost more when a task needs many failed attempts.
- Compare cost per completed, reviewed task alongside latency and correctness.
Sources checked October 16, 2026: Anthropic gateway billing behavior; Superfastcode’s pricing rules, Console usage records and credit terms.