Claude Opus 5: Base Tokens, Fast Mode and Marketplace Billing

A dated model-specific starting point

Anthropic introduced Claude Opus 5 on 24 July 2026 and stated an API list rate of $5 per million input tokens and $25 per million output tokens, matching its predecessor Opus 4.8 at launch. The model identifier was claude-opus-5. This dated announcement supports a launch rate; it does not establish that later releases, enterprise agreements or third-party clouds necessarily bill identically. The underlying model and the procurement route must both be frozen in a historical snapshot.

Fast mode is a separate rate card

The same announcement described Fast mode at twice the base Opus 5 price on the Claude Platform, implying $10 per million input and $50 per million output tokens for that mode. Faster turnaround is a service option, not a claim that an individual request becomes smaller. On a task where output length expands, the speed choice can significantly change variable cost. Do not assume Fast mode availability on every partner marketplace or automatically apply ordinary batch discounts to it.

Three invoices, not one price

The direct Claude API, consumer Claude Max subscription, and partner-operated AWS or Google Cloud inference are not interchangeable invoices. Cloud partners may apply regional endpoint policies, commitments, platform adjustments or account-specific enterprise discounts; the vendor directs readers to those partner prices. A consumer user seeing Opus in an app has not purchased unrestricted developer API usage. Track biller, SKU, country, region, throughput mode and currency as distinct ledger fields.

A worked inference budget

At the dated launch base rates, 400,000 input tokens plus 120,000 output tokens would cost (0.4 × $5) + (0.12 × $25) = $5 before optional extras. The same hypothetical volume at the twice-price Fast rate would be $10. This is only a metered illustration: actual work might consume different tokens when fast mode changes user behaviour or model reasoning. Add cache writes, cache reads, retries and storage only when independently measured.

Before choosing a paid tier

Obtain the billing quote for the exact Anthropic product, account geography, tier, model or subscription, currency and commitment length. If the provider offers several kinds of meters, forecast each independently rather than summing quantities in incompatible units. The relevant captured meters here include input_tokens_per_1m, output_tokens_per_1m. Build a low, expected and high workload from your own request logs or editorial schedule; a pricing page is not a credible estimate of how many retries or approvals your team will need.

For governance, add an approval threshold for paid overages, reconcile invoices with metered activity, and schedule a quarterly or release-triggered source audit. The absence of an identified enterprise rate, promotional expiry or regional tax line is not evidence that the amount is zero. This guide does not represent a product trial, capability benchmark, procurement offer or warranty of uninterrupted access.

Historical availability and boundaries

Preserve the precise source version, checked date and any stated effective date before reusing these figures. A model launch announcement may confirm an introductory token rate without proving which private offers remain available; a mutable plan page can describe a later product state. If a source changes after this cutoff, publish a new snapshot instead of silently editing the historical record. Specialised product evaluations belong to the relevant AI 02–09 editorial teams, while this article documents the billing model and its limitations.