Analysis · 22 September 2026 · 7 min read

Claude Opus 5.5: 20 hours of work in 3

A codebase audit that took Opus 5 over 20 hours took Opus 5.5 under 3. What faster long running work means for jobs, with Anthropic usage data by occupation.

Buried in the Claude Opus 5.5 announcement is a comparison that says more about work than any leaderboard. Anthropic reports that Opus 5.5 audited and fixed a 200,000 line codebase in under three hours. The same job took Opus 5 more than twenty.

A second example from the same page: one tester completed a 680,000 line code migration in less than a day.

These are vendor examples, chosen by the vendor, on the tasks where the model shines. Read them as the ceiling, not the average. But the direction they point is consistent with everything else published this year, and it is the direction that matters for jobs: models are getting dramatically faster at finishing long pieces of work without a person holding the thread.

From one release to the next

Between Opus 5 and Opus 5.5, that one audit went from a working week of machine time to an afternoon. Terminal-Bench 4.0, which tests exactly this kind of sustained multi step work, moved from 52.3 percent to 66.4 percent across the same release.

Speed is the least discussed part of any launch because it sounds dull next to a benchmark score. Economically it is one of the loudest. A job that finishes in an afternoon instead of a week can be handed over, checked and corrected inside a normal working day, and that turnaround, together with cost against a wage, is what decides what gets handed over. Our post on the new pricing covers the other half of that equation.

What real usage already shows

Vendor examples are one thing. Measured usage is another, and Anthropic publishes that too. In its June 2026 Economic Index data release, Anthropic estimates for each occupation how long its tasks would take a person without AI, and how long with it. Across the 700 occupations in its API data, the median task takes about 44 minutes without AI and about 5 minutes with it, roughly a tenth of the time.

For the occupations most people search for, the gap is wider:

Estimates from Anthropic's June 2026 release, API traffic, global. They describe the tasks people actually bring to Claude in those occupations, which skew toward the parts of the job AI handles well, so they overstate the saving for the job as a whole.

Faster tasks are not fewer jobs

A task that shrinks from hours to minutes changes a job. It does not, by itself, remove it. Anthropic's own labour market research found no systematic rise in unemployment among the most exposed workers so far. What it did find is an early slowdown in hiring at the entry level, which is where the compressed tasks concentrate: juniors do the long, well specified, checkable work first, and that is exactly the work that got faster.

So the practical question is not whether AI can do your job. It is how much of your week is long, well specified work that a model can now finish in a fraction of the time, and how much is judgment, relationships and decisions that no benchmark measures. Computer Programmers (59) and Software Developers (47) are the clearest pair in our index: same field, different share of each.

Measure your own share

Our index scores 961 occupations from published research, but an occupation score describes the typical version of a job. Our free assessment scores yours: your seniority, your sector and how routine your week is, against the research baseline for your role. It takes about two minutes and needs no signup.

For the full benchmark results, 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.