Introduction
Anthropic published a study of scientific Claude use on November 5. Its main sample covers 207,049 consumer-plan research sessions from August 2 to September 30, 2026. Data analysis accounted for 33.7% of sessions on Claude and 48.1% on Claude Code. Those percentages describe activity, not measured hours saved. Study and methods
For a team deciding where to use AI, the useful question is what happens after that activity: does the full piece of work finish sooner, at the same standard?
Read the denominator before the headline
Anthropic used classifiers to identify scientific sessions and assign research tasks. It retained categorical labels for the analysis, rather than extracting open-ended research details. A separate survey covered 528 scientists, primarily active AI adopters in the life and health sciences. The report also models possible research acceleration under assumptions about adoption, capabilities and costs; those scenarios are not observed productivity gains. Methods and scenarios
Our reading is that the usage map helps identify tasks worth investigating. It cannot substitute for timing your own workflow. Session frequency, time spent, accepted output and end-to-end completion are different units. A task that generates many short conversations can dominate a usage chart while occupying little of a project's schedule.
A fictional 100-hour project
Here is a deliberately simple calculation, separate from Anthropic's model. Imagine a sequential research project with 30 hours of data analysis and 70 hours of other work. Assume an AI-assisted method halves the analysis time without changing quality, review effort or the other steps.
- Original workflow: 30 hours of analysis + 70 hours of other work = 100 hours.
- Changed workflow: 15 hours of analysis + 70 hours of other work = 85 hours.
- Time saved: 15 hours, or 15% of the original total.
- Overall speed: 100 ÷ 85, approximately 1.18 times the original rate for an equivalent project.
The analysis step is twice as fast. The whole project is not. The 30-hour share is an invented input for this example; it is not the study's task share, an araNET result or a measured Claude speedup.
The example also gives a useful boundary. Even eliminating those 30 analysis hours entirely would leave 70 hours, under the same assumptions. If a proposal promises to halve the full project's duration while changing only that step, it needs another explanation: perhaps other work also changes, the steps overlap, or the initial time breakdown is wrong.
Add review before claiming the saving
Now suppose the faster analysis needs four extra hours to inspect calculations, resolve a data mismatch and reproduce the final result. The changed workflow takes 89 hours. The saving falls to 11%, and the overall rate rises by about 1.12 times rather than 1.18 times.
These are local arithmetic checks on hypothetical inputs. We did not run Claude on a scientific dataset or test the report's forecasts.
For a real comparison, define the accepted deliverable before the experiment. For example: the same input dataset, a saved analysis script, reconciled totals and a figure another team member can regenerate. Count the time required to reach that state, including failed attempts and corrections. Comparing an unchecked draft with a reviewed final result would reward the wrong thing.
Measure a completed workflow
We would start with a small task record rather than a count of prompts:
- Record the input version and the question being answered.
- Separate active work, review and blocked time; retain the order in which steps depend on one another.
- Save the output and its acceptance check, along with the human corrections it required.
- State whether the work was usable and whether its dependent next step could begin.
The distinction between effort and elapsed time matters. If a saved hour occurs while the project is already waiting for access to equipment, the calendar finish may stay unchanged. Conversely, shortening a step that blocks every other step may move the finish even when its share of total labor is small. Our 100-hour calculation assumes a sequential chain; it should not be applied unchanged to overlapping work.
What changed, and what remains open
This is a research report, not a new model or access announcement. Its sample is Claude activity on the specified plans and dates, not a census of all scientific AI use. Primary report
Our takeaway for a small team is to use the reported activity patterns to choose a concrete experiment, then judge the complete deliverable. A convincing result should name the work, its acceptance standard and the time that remained after review. That makes a speed claim useful to the next person deciding whether the workflow fits their project.
