Introducing Claude Opus 5
# The effort dial and the price that didn't change
Anthropic shipped Claude Opus 5 on Thursday, July 24. The pricing is the first thing to notice, because it is exactly the same as the model it replaces: $5 in, $25 out, per million tokens. Opus 4.8's pricing, unchanged, on a model that lands 61 on the Artificial Analysis Intelligence Index and 43.3% on Frontier-Bench, both new highs for that tier. SWE-bench Pro at 79.2%. Positioned by Anthropic as Fable-5-class intelligence at half the Fable 5 price.
Same-price, better-model is the vendor's promise every couple of years, and mostly the numbers do not survive independent evaluation. So take those first two paragraphs down as marketing until Artificial Analysis, Terminal-Bench, and your own eval set say otherwise. What is objectively new, and the reason to write about the release the day it lands, is the effort dial.
Opus 5 exposes a low / medium / high setting on the API, plus a max tier, plus a separate fast mode that runs about 2.5x faster at twice the price. Four price and quality points on the same model, chosen per request by the caller. Not four different models. Not a routing decision made by Anthropic's server. A dial the caller turns.
If you run a routing table, you have felt the pain this fixes. Two weeks ago in Brief 63 I described the exercise of rebuilding a client's routing table after new pricing landed: about half the calls I had parked on a frontier tier moved down to a workhorse at roughly a tenth of the cost with no measurable quality loss on their own eval set. That exercise was expensive to run, because it required re-benching every candidate model against the customer's task, and then re-doing the work every time a new price appeared. An effort dial inside one model reframes it. You are no longer picking between families for the middle eighty percent; you are picking a setting inside one family and paying the difference.
The routing table for the near future then looks like three rows, not seven or eight. A Fable 5 or unreleased-frontier row for the very small share of calls where cost does not matter and correctness is everything. An Opus 5 row for the fat middle, with the effort setting exposed as a parameter the application chooses per request. And a heuristic or small open-weight row for the tail that never needed an LLM. The interesting engineering work moves out of 'which model is best' and into 'which effort setting is the cheapest that still passes the customer's eval.' That is a much more tractable question, and one you can ship an experiment for on a Wednesday afternoon.
Two second-order effects worth watching. First, evals get more useful, because you can sweep the effort dimension against the same underlying model without needing to hold a hundred other variables constant. If you have a task eval set, running it across low / medium / high / max on Opus 5 is a one-day exercise that produces a real cost-per-completed-task curve. That curve, in turn, is a much better artifact for a portfolio conversation than 'we're on the flagship' or 'we're on the workhorse.'
Second, the same-price part of the announcement is doing a lot of work. Anthropic held the flagship line at $5 / $25 while everyone else, notably xAI two weeks ago with Grok 4.5, has been racing the sticker price down. The floor is the customer's cost-per-completed-task, and if Opus 5 finishes a task in fewer tokens than Opus 4.8, the effective per-task price fell without the sticker moving. Watch for the independent replications on token efficiency; if they hold, this is a bigger cut than any headline number.
The choice this reopens for anyone with a serious Anthropic bill is Opus 5 versus Fable 5. Fable is still the top of the pyramid, and for the calls where you were already paying it, keep paying it. But at half the Fable price with the ability to dial effort up on hard cases, Opus 5 becomes the reasonable default with a Fable escalation path for the few percent that need it. If your production routing table was written before today, it is out of date.
One thing to check before Monday. Take one production route that currently pins to Fable 5. Run the same eval set against Opus 5 at high and at max effort. Log tokens used per completed task on both. If Opus 5 max ties Fable and uses fewer tokens on completion, that route pays for the effort dial and then some. If it does not, you know Fable is doing something real for that workload and can defend the line item next quarter. Either outcome is a better answer than the one you have today.
The rest is the usual. Available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, claude.ai, Claude Code and Cowork. New default on Claude Max. Strongest model on Claude Pro. If you re-run your routing table this weekend and want a template, reply and I will send the one I am using with three clients this week.