A score of 66.4% against 57.9% tells you nothing about what you will actually notice. What shows up in use is three things: how long a task takes, how many attempts it needs, and the bill at the end of the month. Here is what is reported on those three points, and where those gaps come from.
Last night we published the analysis of the release and the pricing. Here is the practical part.
The most telling reported case
One tester says they had a 200,000-line codebase audited and fixed in under three hours. The same task would reportedly have taken more than twenty hours on Opus 5, while consuming two and a half times as many tokens.
That figure comes from a user quoted by the company, on a task we cannot reproduce. It should therefore be read as an illustration, not as a measurement. But the order of magnitude is consistent with what is announced elsewhere, and it illustrates well why the gains compound.
Why the gains multiply instead of adding up
This is the point to grasp, and it explains the gap between a 20% cut in the rate and 40% off the real cost.
On an agentic task, three factors act at the same time.
The price per token falls by 20%.
The number of tokens consumed also falls, because the model reaches the result in fewer steps. An agent that finds the right lead on the third attempt rather than the tenth does not cost 20% less: it costs three times less.
The cached share is billed 60% less. Yet in an agentic loop, the re-read context — the codebase, the instructions, the history — accounts for most of the volume. This is the mechanism we detailed in our article on caching.
These three effects multiply. That is why conversational use will see a modest improvement, and why an agent running in a loop will see a considerable change.
An investment firm reports two things. The model passed its evaluation suite at the lowest effort setting, where it previously had to be turned up. And at a higher setting, it identified an error in the instructions of that evaluation suite, which no previous model had flagged. That second point is more significant than a score: spotting that a question is badly posed is a different skill from answering it well, and it is precisely what these systems lack, as we wrote about their tendency to validate whatever is put to them.
Speed, and what it unlocks
A generation more than 30% faster is not just a comfort.
In a single conversation, you barely notice. For an agent chaining thirty steps together, each shorter step compounds, and a task that used to take all night can fit into a morning. It is the difference between a tool you launch and forget, and a tool you work with.
We touched on this when we developed text diffusion models: latency decides what is possible, not just what is pleasant.
The change nobody will talk about
Anthropic says it has worked on the shape of responses: putting the important information up front, so it can be grasped at a glance.
That sounds cosmetic and is not. A response that buries its conclusion under three paragraphs of preamble forces you to read the whole thing to find out whether it is useful. Across dozens of interactions a day, the time lost far outweighs the gains in generation speed.
This is the kind of improvement that shows up in no benchmark and changes the actual experience. We noted the same phenomenon when writing about invisible settings: part of what you perceive as a model's quality comes from presentation choices.
What does not change
Three caveats, so as not to tell too smooth a story.
It is not more capable, it is cheaper. The model is announced at the level of Fable 5.1, not above it. If you were hitting a capability ceiling, it is probably still there.
Effort settings still matter just as much. The gap between the medium setting and the maximum remains considerable, as we explained in our article on that parameter. A test at the wrong setting will tell you nothing useful.
The gains depend entirely on your usage. The advertised 40% applies to a given load profile. On plain text processing without caching, the saving will be limited to the 20% on price.
What to do
The method, always the same, and it takes an hour.
Take your twenty reference cases. Run them on the old and the new model, at the same effort setting. Compare three things: the success rate, the number of tokens consumed, and the total time.
If you do not have those twenty cases, now is the time to build them. It is the only investment that pays off with every model release, and there will be more: Anthropic has announced Sonnet 5.5 and Haiku 5.5 in the coming weeks.
>>