A single request that fills the 1M-token window costs $5.00 in input before the model writes a word [3][4][1]. Take the reply to the 128k ceiling and that adds $3.20 [2], so the most expensive single call claude-opus-5 will accept is $8.20 [3]. That ceiling is not the interesting number. The loop is.
Agents re-send their state every turn. A tool-using run that carries 500k tokens of context through ten turns bills 5M input tokens, which is $25 at list [4], and ten turns is unremarkable for a coding agent. The material we have quotes input, output, Fast mode and batch rates but no cache-read rate for this model [14], so the arithmetic above assumes you pay full price to re-read the same context on every turn. If caching applies, it moves the total more than any discount below.
Rate limits and budget are the same conversation. Anthropic's published tiers run from Tier 1 at 50 requests and 500k tokens per minute to Tier 4 at 400 requests and 4M tokens per minute, per documentation the post dates to May 2026 [9]. One full-context request therefore needs two minutes of a Tier 1 account's entire token allowance [5], which makes the 1M window a spec at that tier rather than something you can spend. At the top tier, saturating 4M input tokens a minute costs $20 a minute, or $1,200 an hour [6]. Committed-use discounts and dedicated throughput are negotiable, with enterprise contract approval running 3 to 10 business days against instant self-serve activation [10].
Fast mode is priced at 2x for roughly 2.5x throughput [5], which works out to $10 and $50 per million [9]. It does not reduce the token count, so the trade is double the spend for about 40% of the wall clock [7]. Defensible for interactive coding, hard to justify overnight, particularly against the 50% discount the documentation describes for asynchronous batch work, whose exact Opus 5 rate the post says to confirm on the official pricing page [6].
The gateway route is the one number that moves unit economics. Api.Airforce lists $3.58 and $17.88 per million through an OpenAI-compatible endpoint, roughly 28% under list [7]. On the ten-turn run above, that is $17.90 of input instead of $25, a saving of $7.10 [8]. The same post warns that gateway pricing is volatile [8].
Worth noting where all this comes from: a dev.to writeup that attributes the rates to Anthropic's pricing page and newsroom alongside several aggregator sites, and that repeatedly tells readers to check the official page rather than trust the article [13]. The $5 and $25 are firm enough to plan against. The batch and cache rates are not numbers yet.
