Median savings
25.3%
input-token reduction per session, combined across 1 dataset
Lens lays out a lot of signal at once, so it is built for the desktop. Open it on a laptop or a larger display and everything will be right where you left it.
Loading this view
Loading this view
Lens lays out a lot of signal at once, so it is built for the desktop. Open it on a laptop or a larger display and everything will be right where you left it.
Lens brings activity, project and developer cost, source coverage, anomalies, policies, approvals, and the evidence behind each decision into one operating view. Start with one source and follow each material change from evidence to an accountable decision.
One time window and scope follow every step, so the evidence you notice on Home is the evidence you investigate, discuss, approve, and audit.
Connect your stream
Start with one read-only source. Lens turns the first real signal into cost, activity, anomaly, policy, and source-coverage evidence.
Teams that opt in can also find removable prompt residue and verify measured savings without making token reduction the product's operating model.
30-second demo
Paste one log to see the evidence Lens extracts and the optional token-removal opportunity it can identify on this sample (). No install, no account, no data retained.
The comparison above runs one prompt. To run the full benchmark, the playground replays a full dataset and compares savings and parity across model and toolset configurations. Open the playground.
Ready to see and govern your AI operations?
Sample data lives entirely in your browser.
Public benchmark proof
Published June 18, 2026.
25.3%
input-token reduction per session, combined across 1 dataset
No regression detected
26.2% parity, p=0.21
the share of replayed pairs the judge scored at full parity with the baseline
42replayed pairs in this run
each pair runs the same prompt with and without lens optimization
No regression detected is the honest read here. The parity figure is the share of replayed pairs the judge scored at full parity with the baseline, and at this sample size the difference from baseline is not statistically significant (Wilcoxon p=0.21). A non-significant difference is not proof of equivalence, so the measured story is no regression detected while the savings land, not a quality regression and not proven parity.
View the published numbers | View run history | Read the methodology