Analysis · 22 September 2026 · 6 min read

Claude Opus 5.5 pricing and the cost of work

Opus 5.5 costs about 40 percent less to run than Opus 5 at roughly Fable level. Why price, not capability, decides which work gets handed to AI.

The most important number in the Claude Opus 5.5 release is not a benchmark. It is the price.

Anthropic's announcement cut Opus pricing to 4 dollars per million input tokens and 20 per million output, down from 5 and 25 for Opus 5. Cached reads fell from 50 cents to 20. Anthropic says the model costs about 40 percent less to run than Opus 5 on typical workloads, generates output more than 30 percent faster, and performs at roughly the level of Claude Fable 5.1 on most tasks.

Fable level work at a lower price than the previous Opus. That is the change worth understanding if you are wondering what AI will do to your job.

Why price, not capability, sets the clock

Every model launch invites the question of whether AI can now do some piece of work. For most office tasks the answer has been yes for a while. The question that actually decides whether a task moves from a person to a machine is whether the machine is cheaper per finished piece of work, once someone has written the request down and checked the result.

An MIT study of computer vision automation put a number on how much this matters: at the costs prevailing when it was run, only about a quarter of the wages attached to technically automatable vision tasks would have been economical to automate. Three quarters of what was already possible stayed with people, purely on price.

Every price cut redraws that line. A cheaper model does not make any new task possible. It makes more of the already possible tasks worth handing over.

How much cheaper, against the competition

Anthropic's headline comparison against OpenAI's GPT-6 Astra is on Terminal-Bench 4.0, a test of sustained terminal work, where it reports Opus 5.5 matching Astra's performance at roughly 40 percent of the cost.

That is a vendor claim and the direction matters more than the exact ratio. Three weeks ago we wrote that Astra finishes a benchmark task for between $0.63 and $2.57 depending on effort. If Anthropic's ratio holds in independent testing, the same kind of task now costs a fraction of that.

What happens when a finished task costs cents

The work that moves first is not the hardest or the most valuable. It is the work where the saving multiplies and the risk stays small:

• high volume, so a small saving per task adds up
• well specified, so the request fits in a paragraph
• cheap to check, so a wrong answer is caught before it costs anything
• already on a screen, so no physical step blocks it

That list is not theoretical. It describes the four functions that took most of the 205,000 AI related US job cuts reported through August 2026: customer service, data operations, entry level software and finance back offices. Falling prices push the same line further into the same kind of work.

In our index, the occupations built on that profile already carry the highest scores. Customer Service Representatives (68) and Accountants and Auditors (54) are worth comparing: both text bound, both high volume, with the gap between them explained by how much of the work is judgment rather than processing.

Cheaper for whom

A lower model price does not mean your employer will replace anyone next quarter. Procurement, integration, liability and retraining all sit between a price cut and a changed role, and they move at the speed of the slowest department. What a price cut does is shorten the list of reasons not to.

The useful question is how much of your own week fits the four bullets above. That share is what our assessment measures, across the same 961 occupation baseline this site is built on, adjusted for your seniority, your sector and how routine your work is.

Take the free assessment. It takes about two minutes and needs no signup or card. For the benchmark results themselves, see Claude Opus 5.5 benchmarks and your job.

Figures come from the AI Job Risk Index and were current when this was published. Scores change as new research is ingested, so the index is always the live version. See how scoring works. Informational guidance based on published research, not professional career or financial advice.