Claude 5.5: cost fell, in two different ways

Anthropic announced two models six days apart: Claude Opus 5.5 on 22 September and Claude Sonnet 5.5 on 28 September. Both launches use the word "cheaper" — but in two different senses, and the difference matters for businesses.

With Opus 5.5 the list price genuinely fell: from $5 to $4 per million input tokens, and from $25 to $20 per million output tokens. That is roughly a 20% cut, and the first time an Opus model has ever gotten cheaper. Combined with efficiency gains, Anthropic puts the total cost to run it at 40% below Opus 5.

With Sonnet 5.5 the list price did not change at all — $2 input and $10 output per million tokens, the same as Sonnet 5. It is still described as "up to 30% less for most work." The reason: the model runs 30% faster and finishes the same job with fewer tokens and fewer tool calls. The unit price is flat; consumption per task is what falls.

What the numbers say

The most striking result is on Terminal-Bench 4.0. This test measures whether a model can carry a real, multi-step workflow through to completion in a command-line interface — it asks "did it finish the job?", not "did it answer correctly?". Sonnet 5.5 scores 70.6% here; Opus 5.5, twice its price, lands at 66.4%. The previous Sonnet 5 scored 10.3% — a gap that is less an improvement than a change of category.

The real message in the chart: on this test, the cheaper model beats the expensive one. But reading that as "Opus is no longer needed" would be wrong. In Anthropic's own words, Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment. A benchmark measures one face of a job.

Across the other tests the picture is more even:

Opus 5.5 still leads most rows, but the margins are thin — 1846 against 1844 on knowledge work. A twofold price difference does not buy a twofold difference in output for most jobs.

The overlooked variable: effort setting

The thing that moves your bill most does not appear in the price table at all: the effort setting. The same model can answer the same question at low or high effort, and cost per task differs by several times between them. In Anthropic's Terminal-Bench chart, Sonnet 5.5's lowest and highest settings differ by several times in both score and cost.

By the company's own note, at the medium effort that is the default in the Claude apps, Sonnet 5.5 exceeds Sonnet 5's best score for less than a tenth of the cost per task. In practice: picking the model is half the decision, and matching the effort setting to the job is the other half. Running everything at maximum effort usually inflates the bill without improving the result.

Two different kinds of "cheaper"

This distinction is not a technical footnote — it changes how you budget. A list-price cut is predictable: hold your volume steady and your bill drops by that much. Efficiency savings depend on the kind of work — pronounced on long, multi-step, tool-using jobs, and barely noticeable on short one-off requests.

So how much of that "30% cheaper" applies to you depends on what you are actually asking the model to do.

Why this matters for your business

Most work that could be automated with AI was postponed not because of a technical barrier, but because the arithmetic did not hold. An assistant answering customer questions, a flow generating product descriptions in bulk, a system reading and classifying incoming invoices — all of that was technically possible a year ago too. The problem was that per-operation cost added up to more than the job was worth.

When cost drops on two models inside the same quarter, that calculation deserves to be run again. The break-even point may have moved.

What to actually do

If you have an idea from this year that was discussed as "let's do it with AI" and then shelved, reopen the file and ask one question: did we drop it because of cost?

If yes, run the same calculation at today's prices. That is the only input that changed; the work itself, its volume, and the expected benefit stay the same.

If no — if you dropped it over data quality, security, or process uncertainty — these announcements mean nothing for you. Cheaper models solve none of those problems.

One caution

"Cheaper" and "costs you less" are not the same thing. As cost falls, usage usually rises: a faster, cheaper model starts getting tried on work nobody attempted before. Track your budget on the monthly total, not the unit price — otherwise you will overspend without noticing the discount.