Benchmarks
Same models. Better results.
We run the same business task with frontier models twice: once on their own, and once inside a Thunk.AI workflow. The models don’t change. The platform around them does, and so do the results.
Drug safety case intake · October 2026
Every model improved, and the gap between them closed
Three frontier models triaged 75 adverse-event reports. Each model ran twice: once as a single AI step given the SOP and the documents, and once as a seven-step Thunk.AI workflow given exactly the same inputs. One run per configuration, scored against the expected outcome for every report.
Model
On its own
On Thunk.AI
Change
gpt-5.1
little or no reasoning
88.1%
97.9%
+9.8 pts
gemini-3.5-flash
little or no reasoning
92.1%
98.5%
+6.4 pts
muse-spark-1.3
default (medium) reasoning
95.4%
99.8%
+4.4 pts
Weighted Score. Two of the models, running with little or no reasoning, scored higher on Thunk.AI than the best model managed on its own.
37 of 37
The reports that couldn’t be missed
Some adverse-event reports must reach a regulator on an expedited timeline. The test set had 37 of them. On their own, the three models missed 9, 4 and 3. On Thunk.AI, every model flagged all 37.
7.3 → 1.9
The model matters less
On their own, the best and worst models were 7.3 points apart. On Thunk.AI the gap shrank to 1.9 points, so a faster, cheaper model lands close to a stronger one. The structure of the workflow does work the model would otherwise have to do.
What it costs
Reliability isn’t free. The seven-step workflow costs 2.5 to 3.8 times as much per report as a single AI step, $0.11 to $0.42, and takes three to five and a half minutes per report. Set against the cost of a missed regulatory deadline, that trade is easy to make.
Earlier benchmarks
The same pattern, in other kinds of work
99%
IT service management
AI Reliability score, February 2026
97.3%
Enterprise document workflows
AI Fidelity score, September 2025