Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens, with cached input at $0.50 and cache writes at $6.25 for the five-minute window or $10 for the hour, read from the Anthropic pricing page on 16 August 2026. That is half of Claude Fable 5 and two and a half times Claude Sonnet 5.
It is also, to the cent, what Opus 4.8, 4.7, 4.6 and 4.5 cost, so upgrading through the Opus line has never had a price.
What Claude Opus 5 costs, line by line
Anthropic bills on five lines rather than two. The two cache write columns are the ones most calculators leave out, and they decide whether keeping a long system prompt alive is cheap or expensive.
| Billing line | Opus 5, per million tokens |
|---|---|
| Input | $5.00 |
| Cache write, 5 minute window | $6.25 |
| Cache write, 1 hour window | $10.00 |
| Cached input read | $0.50 |
| Output | $25.00 |
Opus 5 at $5 in and $25 out per million, read 16 August 2026
The three cache columns are fixed multiples of the input price: 1.25x to write a five-minute entry, 2x to write a one-hour entry, and 0.1x to read either. Nothing about Opus 5 is special there, which is useful, because it means the arithmetic you work out once carries to almost every other tier. There are two exceptions: on Claude Fable 5.1 and Claude Mythos 5.1 a cache read costs 0.025x rather than 0.1x, which is $0.25 per million, and on Claude Opus 5.5 it costs 0.05x, which is $0.20 per million, per the Anthropic pricing page read on 29 September 2026. Claude Fable 5.2 and Mythos 5.2, listed on 9 November 2026 at $8 input, keep the 0.025x read, also $0.20 per million.
Five generations, one price
This is the fact worth knowing before anything else. Every Opus from 4.5 to 5 is billed identically, unlike the tier above Opus 5, which follows a different pricing structure. However, the newer Claude model represents a shift in strategy. The Claude Opus 5.5 pricing breakdown shows it at $4.00 input and $20.00 output with cache reads at $0.20 per million, twenty percent under Opus 5 on input and output and sixty percent under it on a cache read, according to the Anthropic pricing page as of 29 September 2026.
| Model | Input / cached / output, per million |
|---|---|
| Claude Opus 5 | $5.00 / $0.50 / $25.00 |
| Claude Opus 4.8 | $5.00 / $0.50 / $25.00 |
| Claude Opus 4.7 | $5.00 / $0.50 / $25.00 |
| Claude Opus 4.6 | $5.00 / $0.50 / $25.00 |
| Claude Opus 4.5 | $5.00 / $0.50 / $25.00 |
So if you are still calling Opus 4.5 because a migration was never scheduled, the money argument for staying does not exist. The retired models are the exception and they run the other way: Opus 4.1 and Opus 4, still reachable on Bedrock and Google Cloud, are billed at $15 and $75, three times the current rate.
The oldest Opus you can still call is the most expensive one. Newer is cheaper here, which is the reverse of how hardware works and catches people out.
Three prices for the same model
The grid above is the standard rate. Anthropic sells the same Opus 5 two other ways, and what separates them is not quality but when the answer arrives.
The Batch API takes fifty percent off input and output alike, putting Opus 5 at $2.50 and $12.50. The trade is that batch work is asynchronous, so it only suits jobs nobody is waiting on: overnight enrichment, backfills, evaluation runs.
Fast mode runs the other way. In research preview, it serves Opus 5 and Opus 4.8 (Anthropic retires the Opus 4.8 option on 30 November 2026) at $10 and $50 for significantly faster output. That grid is worth staring at, because it is exactly what Claude Fable 5 costs at standard speed. At that budget you are choosing between a different model and the same model answering faster, which is a much more interesting question than it first looks. Fast mode is first party only and cannot be combined with the Batch API.
What a month actually costs
Three shapes of work on the standard grid, with ninety percent of input served from cache, which is realistic for anything carrying a stable system prompt.
| Monthly workload | Opus 5 / Sonnet 5 / Haiku 4.5 |
|---|---|
| Assistant: 2M in, 0.5M out | $14.40 / $5.76 / $2.88 |
| Same volume, no caching at all | $22.50 / $9.00 / $4.50 |
| Agent: 20M in, 4M out | $119.00 / $47.60 / $23.80 |
Two things stand out. Caching removes more from the bill than dropping a whole tier does on input-heavy work, and the ratio between Opus 5 and Sonnet 5 stays at exactly 2.5x whatever the shape of the month, because the entire grid is scaled rather than the individual lines. Our LLM cost calculator runs the same volume across every provider we track.
When Opus 5 is the right tier
Opus 5 is the tier most production work should sit on when the output matters and the volume is not enormous. Below it, Sonnet 5 costs 2.5 times less and handles the large majority of tasks; above it, Fable 5 costs twice as much and earns that only where a wrong answer is expensive.
The honest test is not which model is best in the abstract, it is whether your pipeline retries. If you run two Sonnet 5 passes and a judge to get one usable result, you have already spent Opus 5 money and added latency. Count the retries before you compare the grids. And if the failure you keep hitting is speed rather than quality, fast mode changes one line of configuration where a model swap changes your prompts.
How you buy it changes the price
Everything above is the first-party rate. Routers add a margin on top of the provider grid, and the managed clouds run their own grids entirely: Bedrock and Google Cloud price Claude separately, with a ten percent premium on regional and multi-region endpoints. On the first-party API, pinning inference to the United States through the data residency parameter applies a 1.1x multiplier across every line, cache included. Our full comparison of what every major model charges puts Opus 5 against the rest of the market.
Frequently asked questions
How much does Claude Opus 5 cost per million tokens?
$5.00 for input and $25.00 for output, with cached input at $0.50 and cache writes at $6.25 for five minutes or $10.00 for an hour. Read from the Anthropic pricing page on 16 August 2026.
Is Opus 5 more expensive than Opus 4.8?
No. They cost exactly the same on all five billing lines, as do Opus 4.7, 4.6 and 4.5. There is no price reason to stay on an older Opus.
What is the cheapest way to run Opus 5?
The Batch API, at $2.50 input and $12.50 output, a flat fifty percent off. It stacks with prompt caching, which takes ninety percent off any input you resend. The cost is that batch requests are asynchronous.
Why does fast mode cost the same as Fable 5?
We do not know why, only that it does: fast mode on Opus 5 and Opus 4.8 (retiring on 30 November 2026) is billed at $10 and $50 per million, identical to the Fable 5 standard grid. It makes the two a direct choice at the same budget.
Is Opus 5 cheaper than GPT or Gemini at the same tier?
It depends entirely on the shape of your work, because output-to-input ratios differ sharply between providers. We compare them side by side on the same reading date in our cross-provider page.
What we verified and what we did not
Verified: all five billing lines for Opus 5, 4.8, 4.7, 4.6 and 4.5, the retired Opus 4.1 and Opus 4 rates, the fifty percent Batch API discount, the fast mode grid, and the cache multipliers, all read from the Anthropic pricing page on 16 August 2026, with the Opus 4.8 fast mode retirement read in the release notes on 3 November 2026.
Not verified, and therefore not asserted: rates on Bedrock or Google Cloud, which publish their own grids; any benchmark or quality claim, since this page prices models rather than ranking them; and what fast mode delivers in practice, which Anthropic describes as significantly faster output without publishing a figure.
Prices in this category move. Everything here carries the date it was read, and the grid is worth checking at the source before you commit a budget to it.