Cortex Discover 1.6 reduces the cost of large AI workspaces, long conversations, and streamed responses while preserving the browser performance users expect.
A browser should not become progressively heavier as its intelligence becomes more useful.
That principle has guided our performance work on Cortex Discover from the beginning. In Making Agentic AI Stay Fast with Cortex Discover, we moved beyond short benchmarks and asked whether an integrated AI experience could remain responsive through sustained use. With Cortex Discover 1.5, we showed that faster starts, responsive browsing, and near memory parity could coexist with an agentic workspace.
Cortex Discover 1.6 takes the next step.
The focus is no longer only whether AI can remain available without accumulating uncontrolled overhead. The new question is what happens as the work itself becomes larger: conversations grow, histories become richer, files and page context move through the workspace, and streamed responses update continuously.
In controlled testing, Cortex Discover 1.6 opened a large saved workspace 47.1% faster, reduced settled panel heap by 56.0%, and completed a streamed-response rendering workload 48.6% faster. The amount of context data retained after the tested workflow fell by 99.2%, while inactive session connections fell by 90.0%.
These are scoped results from defined workloads, not universal claims about every device or browsing pattern. Together, however, they demonstrate a clear direction: persistent AI can become more capable without allowing its interface and retained state to grow without bounds.
From Staying Fast to Becoming Lighter
Our earlier work established an important baseline.
Cortex Discover could keep AI close to the user while maintaining competitive browser performance. Repeated panel use remained healthy, long sessions completed without retained process growth, and the integrated experience stayed responsive after the first interaction.
Version 1.6 builds on that foundation by reducing the work required inside the AI surface itself.
The latest performance program examined several pressures that become more important during everyday agentic work:
- Larger saved conversations and workspaces
- More page context and file-related state
- Continuous streamed-response updates
- Repeated navigation and interface refreshes
- Background sessions that are available but inactive
The objective was not to remove capabilities or shorten the work being performed. It was to keep the experience useful while making its resource boundaries more deliberate.
That distinction matters. An AI workspace can feel responsive when it is empty and still slow down as users build history. It can release visible interface elements while retaining unnecessary data behind the scenes. It can render a response correctly while doing more work than the user can see.
Cortex Discover 1.6 addresses those longer-term costs.
Large Workspaces Open 47% Faster
Persistent AI becomes more valuable as it remembers more of the work around it.
Users may return to earlier research, security reviews, technical investigations, files, or extended conversations. That continuity is useful, but it also creates a performance challenge: opening a large saved workspace should not require loading every historical detail into the active interface at once.
In the qualified large-profile workload, Cortex Discover 1.6 reduced opening time at the 95th percentile by 47.1%. The associated stored footprint for the test profile was also reduced by 80.6%.
For users, the result is straightforward. Returning to substantial prior work requires less waiting and less unnecessary active state.
The improvement preserves durable content and continuity. It is not achieved by silently dropping completed work. Cortex Discover restores the information needed for the active experience while keeping the amount of immediately retained data bounded.
That balance is central to persistent AI. History should remain useful without forcing every prior interaction to become permanent live overhead.
56% Lower Settled Panel Heap
Long-running AI interfaces accumulate more than visible text.
They manage context, attachments, response state, navigation history, and the information needed to continue a workflow. Without clear lifecycle boundaries, those elements can remain active longer than necessary.
In the Cortex Discover 1.6 qualification workload, settled panel heap fell by 56.0%. Retained context and attachment data after the tested workflow fell by 99.2%.
The goal is not to make active intelligence free. Analyzing a document, working with a file, reasoning over page content, or carrying out a connected task naturally requires resources while the work is happening.
The improvement concerns what remains afterward.
Completed work should not leave large retry payloads, temporary context, or inactive state permanently attached to the live interface. Cortex Discover 1.6 keeps the information needed for continuity while releasing data that no longer needs to remain active.
This creates a more predictable workspace for users who keep the browser open throughout the day and move repeatedly between browsing and AI-assisted work.
Streamed Responses Render with Less Work
Streaming is one of the most visible parts of an AI experience.
Users see an answer develop in real time while the interface updates, preserves interaction, and prepares the completed result. A response can arrive quickly from a provider and still feel slow if the client repeatedly performs unnecessary rendering work.
Cortex Discover 1.6 reduced wall time in the qualified streamed-response rendering workload by 48.6%. Time spent in long tasks fell by 51.0%.
The result is not simply a faster animation. Reducing long-running interface work helps preserve responsiveness while a response is being assembled, particularly when conversations and surrounding context are already substantial.
This is part of a broader principle in Cortex Discover: update what changed, retain what is still needed, and avoid repeatedly rebuilding work the user has already seen.
Quieter When Intelligence Is Waiting
An agentic browser should be ready when users need it without treating every inactive session as active work.
Cortex Discover 1.6 reduced inactive session connections by 90.0% in the qualification workload. It also reduced unnecessary navigation-history work by 74.0% and interface update activity by 50.0% in the corresponding controlled tests.
These results reflect a simple product expectation: background availability should not mean background churn.
Users often move between a page and Cortex many times during a task. They may leave a conversation open while reading documentation, reviewing a repository, or comparing information across tabs. The interface should remain ready to continue without continuously repeating work while attention is elsewhere.
Version 1.6 makes that waiting state more disciplined while preserving active and in-progress sessions.
Browser Performance Remains Within Budget
Optimizing the AI surface is only useful if the browser around it continues to perform normally.
The signed Cortex Discover 1.6 candidate was evaluated in a controlled browser-wide comparison with Chrome 156 on Apple Silicon. Both browsers completed the measured workloads.
In that comparison:
- Cortex Discover warm launch was 3.2% faster.
- Active CPU was 5.8% lower in the measured workload.
- Background readiness was 28.5% faster at p95 as a favorable point estimate.
- Rendering remained at 60 FPS for both browsers.
- Workload throughput remained at parity.
- Both browsers completed video playback with zero dropped frames.
Browser-wide active memory for Cortex Discover measured 0.9% above the Chrome 156 control, within the predeclared 2% practical-equivalence budget. We treat that result as near parity, not as a memory improvement.
This framing is intentional. The strongest Cortex Discover 1.6 gains are inside the persistent AI workflow: opening substantial saved work, controlling retained context, rendering streamed responses, and quieting inactive state. The browser-wide comparison serves as the guardrail showing that those changes preserve the broader browsing experience.
The comparison is with Chrome 156 under the stated test conditions.
Tested Beyond the First Interaction
Short tests are useful, but persistent software also needs evidence that its behavior holds over time.
The signed Cortex Discover 1.6 candidate completed nine measured 10-minute sessions across panel-closed, panel-open, and repeated native-cycle conditions after warm-up runs. The campaign recorded:
- 9 of 9 measured sessions completed
- Zero workload errors
- Zero sampling errors
- 903 of 903 healthy cycle samples
- Zero retained process growth
- 25.9 ms render readiness at p95 against a 100 ms limit
- No additional network transfer after the workload settled
The testing was performed natively on Apple Silicon. The results apply to the signed candidate, host, controls, and workloads described in the Version 4 performance report. They are not claims of identical behavior across every device, architecture, workload, or operating environment.
That scope does not diminish the result. It makes the result useful. Performance evidence should describe exactly what was tested and avoid turning favorable measurements into claims the evidence cannot support.
Why This Matters for Agentic Browsing
The browser is becoming more than a place to retrieve information.
Developers move between documentation, repositories, development tools, and running systems. Security teams move between evidence, analysis, and remediation. Researchers and enterprise users build context over time rather than beginning every task from an empty conversation.
An agentic browser has to support that continuity without making users choose between useful intelligence and a responsive workspace.
Cortex Discover 1.6 advances that goal in three ways:
- Large work resumes faster. Substantial saved workspaces open with less delay.
- Completed work leaves less active weight. Context and temporary data remain bounded after use.
- Continuous AI interaction does less interface work. Streamed responses and repeated updates remain responsive as conversations grow.
These improvements are more meaningful together than any single benchmark number. They describe an AI surface designed for continued use rather than a demonstration optimized for its first response.
Performance Is Still a Product Contract
Our earlier Cortex Discover research described performance as a product contract: new intelligence should preserve responsiveness, bounded resource use, and sustained stability.
Cortex Discover 1.6 strengthens that contract.
The earlier question was whether agentic AI could stay fast after the benchmark ended. The new evidence addresses the next challenge: whether the experience can remain light as histories, context, files, and streamed interactions become more substantial.
For the tested Cortex Discover 1.6 candidate, the answer is yes.
Large workspaces open faster. Settled interface memory is lower. Stream rendering requires less time. Inactive sessions do less work. Browser-wide performance remains within the declared release budgets.
The result is not intelligence without cost. It is intelligence with clearer boundaries.
Cortex Discover 1.6. Lighter persistent AI. Faster return to meaningful work. Performance designed to hold as the workspace grows.
Read the Performance Research
The Cortex Discover Performance and Resource Efficiency White Paper – Version 4 documents the workloads, controls, qualification gates, limitations, and complete measured results behind this update.
Results are specific to the named candidate and controlled test conditions.
