Anthropic introduced Opus 5.5 on 22 September with two different price cuts. Rates per million tokens fall 20%, from $5 to $4 for input and from $25 to $20 for output. The company also estimates 40% lower running costs than Opus 5 on typical workloads at default settings, combining those rates with fewer tokens per task. It further reports text generation more than 30% faster. None of these figures means every assignment automatically costs 40% less or finishes 30% sooner. Anthropic’s announcement.
That distinction matters to anyone paying for completed work. A model can reply quickly and leave an afternoon of checking behind. It can also take longer to answer while eliminating several subsequent exchanges. The commercial unit is the token, a unit used to process text; a team's unit of value may be a usable report, an approved repair or a correctly resolved complaint. Confusing the two encourages purchases that look excellent on an invoice and considerably less impressive in a calendar.
Savings depend on what is being counted
Imagine an assignment costing one hundred units in total. Forty pay for model usage; sixty cover preparing information, checking results and incorporating them into a process. If the first component falls by 40% while everything else stays unchanged, the total becomes eighty-four. Overall savings are 16%, not 40%. This is a hypothetical arithmetic example, not an estimate for Opus. It shows how a genuine reduction can become an exaggerated promise without anybody changing the original percentage.
Review does not have a fixed price either. An answer identifying where a figure came from makes checking easier. One mixing evidence and conjecture requires someone to reconstruct the path. When a system proposes code changes, a reviewer needs to distinguish an improvement in intended behaviour from an alteration that merely passes the test. A more capable model may create value precisely here: reducing the work required to establish whether its output is useful.
An independent research project illustrates the difficulty of measuring that effect. In February 2026, METR warned that its new productivity experiment suffered from selection effects and difficulties measuring work with concurrent agents. It considered greater acceleration than in 2025 likely, but evidence of its magnitude weak. That research does not evaluate Opus 5.5. METR’s update.
More output can overwhelm the next desk
A team can multiply drafts without multiplying its ability to accept, reject or correct them. If review becomes an expanding queue, cheaper generation purchases accumulation. If review is skipped to clear that queue, costs may return when an error has already left the office and reached somebody else.
An alternative requires examining where attention is worth spending. A small, reversible change covered by tests can pass through different controls from a decision affecting a customer. Reviewing everything identically wastes resources; reviewing only what the system labels doubtful assumes it can recognise its own errors. Sampling apparently ordinary results helps test that assumption. The sampling need not duplicate all the work, but it must reach cases the model itself has not selected for attention.
A useful comparison would preserve a representative set of assignments, record necessary corrections and count elapsed time until an acceptable result is obtained. Acceptance criteria should exist before anyone sees which model wins. Redefining them to accommodate the most impressive-looking output makes the evaluation reward persuasion. Nor should a trial be confined to tasks the team already knows the tool can perform: difficult cases reveal where another person or process will still be needed.
A price cut can deliver substantial benefits. It makes previously unaffordable assignments viable and permits more alternatives to be explored. Capturing that benefit nevertheless requires deciding what to do with the released capacity: expand production, improve checking or reduce expenditure. These are different outcomes, with different measures of success. A supplier can reduce the price of generating answers; an organisation still has to turn those answers into work that somebody can use.