We began by challenging our own strongest result. What followed changed the benchmark, the architecture, and how we qualify performance.
Earlier this month, we published our first detailed performance study of Cortex Discover, looking across startup, memory, responsiveness, rendering, media workloads, and resource efficiency. That study gave us a useful baseline, but it also raised a more important question: what happens after the benchmark ends?
An agentic browser has a different performance profile from a conventional browser. It is not enough to launch quickly or perform well during a short test. AI remains available while users browse, research, switch tabs, continue conversations, review findings, work with files, and return to Cortex repeatedly throughout the day. The harder challenge is making sure that intelligence does not gradually turn into permanent overhead.
That became the focus of our next performance program.
We Challenged Our Strongest Result
We started by going back to the result that looked best in our original study.
The first campaign showed a substantial apparent memory advantage for Cortex Discover compared with Chrome. When we tightened the methodology and applied stronger controls, that headline changed. Under equivalent test conditions, the updated result showed Cortex Discover operating near memory parity with the controlled Chrome baseline.
We retired the earlier memory headline rather than continue using a number that stronger testing no longer supported.
That correction moved the project away from trying to win a browser ranking and toward a more meaningful engineering question: could Cortex Discover preserve competitive browser performance while keeping a persistent AI experience available throughout normal use?
The answer required more than another short benchmark. It required sustained qualification.
From Benchmarking to Sustained Qualification
The updated controlled testing remained strong. Cortex Discover started approximately 13% faster than Chrome for Testing, while memory reached near parity. Rendering and workload throughput remained effectively equivalent, and background readiness and interaction responsiveness were also close to parity.
We also revisited media performance. In a deterministic H.264 workload, Cortex Discover measured 5.6% lower CPU and CPU time, with equivalent decoded output and zero dropped frames.
But the more important part of this update was what happened over time.
AI interfaces naturally accumulate state. Conversations get longer. Users move repeatedly between the page and the assistant. More capabilities become available. The browser may stay open for hours rather than minutes.
A system can look lightweight during a short benchmark and still become progressively heavier during real use.
So the program evolved from browser comparison into sustained qualification.
Testing What Happens Over Time
The final qualification covered twelve measured 15 minute runs across different Cortex panel states and repeated usage cycles, with all twelve completing successfully.
Panel readiness measured 25.80 ms at p95. Across repeated native cycling, 3,343 out of 3,343 samples remained healthy, and no measured cycling run finished with an additional retained browser process.
The final median footprint difference after sustained testing was +2.11 MiB, remaining within the program’s qualification boundary.
That result is not a claim that AI has no cost. Models, document analysis, security workflows, tool use, and other active capabilities naturally consume resources while they are doing work.
What the qualification shows is narrower and more useful: keeping Cortex integrated into the browser, opening and closing it repeatedly, and returning to it over time does not have to make the baseline experience progressively heavier.
That is the behavior we care about in an agentic browser.
Fast When You Need It. Quiet When You Do Not.
As Cortex Discover grows, it will continue gaining new forms of intelligence. That creates a simple challenge: more capability cannot automatically mean more baseline overhead.
Our latest work focused on maintaining resource efficiency, responsiveness, and long session stability as one connected problem rather than as isolated benchmark targets.
The implementation details are covered more deeply in the updated whitepaper, but the product principle is straightforward:
The browser should remain lightweight when intelligence is waiting, and responsive when intelligence is needed.
That principle matters more as AI moves from occasional assistance into a persistent part of the workspace.
Why 3,343 / 3,343 Matters
One result from the qualification stands out because it captures the difference between a benchmark and real usage over time.
3,343 out of 3,343 native cycle samples remained healthy.
Opening an AI panel once is easy. Keeping it healthy as users repeatedly open it, close it, return to the web, continue a conversation, switch context, and come back again is a more meaningful test.
That is closer to how Cortex Discover is intended to be used. Developers move between technical documentation and Cortex during debugging. Security teams move between evidence and analysis. Enterprise users may keep the browser open throughout the day while conversations and tasks continue to evolve.
The AI surface has to remain responsive across that experience, not just on the first interaction.
The goal is not merely fast AI. It is AI that stays fast.
Performance as a Product Contract
The biggest change from our first study is not a single benchmark number. It is how we now treat performance itself.
The first study was primarily comparative. This second stage turns performance into an ongoing engineering requirement.
As Cortex Discover gains more models, tools, workflows, and user experiences, those additions have to coexist with browser responsiveness, bounded resource use, and sustained stability. Performance cannot be something evaluated only after features have been added.
For agentic systems, that distinction matters. More context and longer running workflows increase the importance of disciplined resource use.
Our approach is to make those costs measurable and bounded.
Fast Browser. Bounded Intelligence.
The updated work gives us a clearer performance story than the original benchmark alone.
The goal is not to claim that intelligence is free, or that Cortex Discover wins every possible browser comparison. The more important result is that an integrated AI experience can remain available without becoming uncontrolled background weight over time.
That is the direction behind this work and the standard we want Cortex Discover to continue meeting as it evolves.
Fast Browser. Bounded Intelligence. Performance that holds after the benchmark ends.
Read the Full Performance Research
The original Cortex Discover performance whitepaper is available at pervaziv.com/white-papers/, and the updated Version 2 report will be published there as well.
The full whitepaper goes deeper into the ten phase qualification program, measurement methodology, sustained testing, evidence boundaries, and the technical reasoning behind the updated results. Readers can also contact Pervaziv AI to request complete versions of both whitepapers.
