
Discover 1.7 and the Cortex 6.5 clients make substantial AI workspaces faster to resume, keep first-use paths lighter, and stay quiet when intelligence is not active.
The best performance improvement is not always a faster launch or a smaller package. It is less friction at the moment a user needs to continue meaningful work.
That is the focus of Cortex 6.5.
Earlier Cortex Discover performance programs answered three foundational questions. Could an agentic browser remain competitive with mainstream browsers? Could integrated AI stay healthy through sustained use? Could large conversations, rich context, and streamed responses become lighter as workspaces grew?
Discover 1.7 builds on those results, but this release tells a different story. Performance is no longer treated only as a property of the browser. The same discipline now spans the Cortex Browser Extension, Cortex VS Code Extension, Cortex Mobile, and Cortex CLI.
The result is a performance program organized around the way work moves: from browsing to coding, from desktop to mobile, and from a visual workspace to the command line.
This is not another browser ranking or a retelling of the Discover 1.6 optimization story. It is the first portfolio-level view of how Cortex 6.5 keeps intelligence responsive across distinct client experiences. Discover receives the deepest runtime qualification, while the other four clients contribute some of the release’s largest first-use and delivery improvements.
Four Results That Define the Release
The Cortex 6.5 performance dashboard starts with four different perspectives: returning to substantial work, minimizing first-use weight, preserving active responsiveness, and becoming quiet when hidden.
| Customer experience | Qualified result | What it means |
|---|---|---|
| Return to a large Discover workspace | 71.6% lower p95 than the pre-1.6 baseline | A 25 MiB, 100-thread profile hydrates in 363.5 ms p95, down from 1,278.8 ms |
| Reduce work before first use | Up to 99.9% less | Optional capability stays outside the initial path across Cortex clients |
| Stay responsive during active work | 16.80 ms p95 | Interaction remains within a single-frame-scale budget in the qualified workload |
| Become quiet when hidden | 0.281% of one CPU core | Discover recorded zero hidden network events and recovered 99.95% of lifecycle heap |
The 71.6% result represents the cumulative improvement across the 1.6 and 1.7 performance programs. Compared directly with the qualified Discover 1.6 result, Discover 1.7 lowers the same large-profile hydration measurement by another 46.3%, from 676.7 ms to 363.5 ms.
These numbers matter together. Fast active work is less valuable if an application stays busy after the user moves on. A small initial path is less valuable if the interface cannot respond smoothly. Cortex 6.5 treats those outcomes as one product experience.
What Changed: Performance Moved Beyond the Browser
The previous performance studies centered on Discover because the browser owns the broadest runtime surface. Cortex 6.5 expands the qualification across five customer touchpoints.
| Cortex client | Performance focus | Qualified outcome |
|---|---|---|
| Cortex Discover 1.7 | Large histories, active interaction, hidden-state efficiency, repeated lifecycle use | 363.5 ms large-profile hydration p95; 0.281% hidden CPU |
| Cortex Browser Extension | Delivery footprint and first-use side panel | 45.8% smaller package; 99.4% smaller side-panel entry |
| Cortex VS Code Extension | Editor-host footprint and focused chat experience | 46.0% smaller package; 64.3% smaller chat render |
| Cortex Mobile | Minimal startup reachability | 442-byte startup path; 99.9% lower than baseline |
| Cortex CLI | Fast entry, bounded idle use, large transcript formatting | 119.2 ms first nonblank frame p95; 46.0 MiB maximum idle RSS |
The workloads are intentionally different because each client serves a different moment in the workday. A mobile startup path and a command-line redraw are not interchangeable benchmarks. They are expressions of the same product principle: load what the current action needs, keep the active experience responsive, and avoid turning optional capability into permanent overhead.
That portfolio structure is what differentiates this release from the earlier Discover articles:
- The original study compared browser-level efficiency.
- The sustained qualification asked what happens after a benchmark ends.
- Discover 1.6 focused on large histories, retained state, and streaming work.
- Cortex 6.5 connects those principles across browser, editor, mobile, and terminal experiences.
Discover 1.7: Faster Return to Substantial Work
The defining Discover 1.7 result is not an empty first launch. It is the experience of returning to an established AI workspace.
The qualified profile contains 25 MiB of conversation data across 100 threads. It represents the kind of accumulated work that makes persistent AI valuable: prior research, technical discussions, decisions, and context that users expect to remain available.
| Large-workspace measure | Discover 1.7 result |
|---|---|
| Profile hydration p95 | 363.5 ms |
| Improvement from Discover 1.6 | 46.3% lower |
| Cumulative improvement from pre-1.6 baseline | 71.6% lower |
| 2,000-chat tree build p95 | 0.61 ms |
| Interaction p95 | 16.80 ms |
| Long-task maximum | 0 ms |
| Dropped-frame intervals | 0 |
For customers, this means continuity does not have to become a startup penalty. Discover can restore substantial local history, make the active conversation available even when a service is temporarily unreachable, and refresh connected state without forcing the user to wait for the entire history to be reconstructed remotely.
That distinction is especially important for long-running research and engineering work. A conversation is not only a sequence of messages. It can be a durable workspace that users reopen throughout the day. Discover 1.7 makes that return path faster while keeping the visible conversation responsive.
Responsive Across Interaction, Streaming, and Reopen
Perceived performance is shaped by repeated moments, not one launch event. Users select a conversation, send a request, watch a response stream, move back to a page, and reopen Cortex later.
Cortex Discover 1.7 keeps each of those moments within the qualification budget.
| Active experience | Observed p95 | Qualification limit |
|---|---|---|
| Interaction response | 16.80 ms | 100 ms |
| 64 KiB streaming update | 17.29 ms | 100 ms |
| Panel reopen | 28.79 ms | 100 ms |
| Optional model first use | 3.20 ms | 1,000 ms |
These results describe a workspace that remains responsive as information arrives and context grows. The benefit is functional as much as technical: users can continue scrolling, reviewing, and moving between the page and Cortex without the AI interface monopolizing the interaction loop.
Quiet When Attention Moves Elsewhere
Agentic AI is often persistent, but it should not behave as if it is permanently active.
Users pause to read. They switch tabs. They review a file in an editor. They leave a conversation open while thinking. In those moments, background activity becomes part of product quality because it competes with the pages, development tools, video calls, and enterprise applications sharing the same machine.
| Hidden or completed state | Qualified result |
|---|---|
| Hidden CPU | 0.281% of one core |
| Hidden network events | 0 |
| Completed attachment payload retained | 0 B |
| Added GPU processes after lifecycle | 0 |
| Lifecycle heap recovery | 99.95% |
This is not a claim that active AI has no resource cost. Models, document analysis, media, search, and connected actions naturally use resources while work is happening. The customer value is that the cost remains tied to the work rather than continuing simply because the interface still exists in the background.
Lighter First Use Across Cortex 6.5
The broadest change in Cortex 6.5 is how much capability stays outside the first useful path.
Modern AI clients can support documents, media, search, speech, local and hosted models, security workflows, connected tools, and workspace context. Loading every capability before the user can begin would make breadth feel like weight.
Cortex 6.5 narrows that first-use path across the portfolio.
| Client path | Baseline | Cortex 6.5 | Change |
|---|---|---|---|
| Browser Extension package | 81.6 MB | 44.2 MB | 45.8% smaller |
| Browser Extension side-panel entry | 1.35 MB | 7.9 KB | 99.4% smaller |
| VS Code Extension package | 12.8 MB | 6.9 MB | 46.0% smaller |
| VS Code chat render | 910 KB | 325 KB | 64.3% smaller |
| VS Code extension host | 918 KB | 409 KB | 55.5% smaller |
| Mobile startup reachability | 338,452 B | 442 B | 99.9% lower |
A smaller initial path does not automatically guarantee a faster launch on every device. It does reduce how much code can participate before the first useful action, which creates a more predictable foundation for responsiveness.
For users, the functional outcome is simple: the active conversation can appear first, while less common capabilities become available when requested.
Cortex Browser Extension: A Focused AI Surface in the Browser
The Cortex Browser Extension brings the AI workspace to customers who use an existing browser environment. Its performance challenge is different from Discover’s: the extension shares a host and should avoid making optional AI capability part of every panel opening.
The qualified production package falls from 81.6 MB to 44.2 MB. More importantly for first use, the side-panel entry falls from 1.35 MB to 7.9 KB.
That makes the extension experience more focused. The active panel can establish the conversation surface first, while document, media, and other specialized capabilities remain available without dominating the initial path.
The measured panel-open footprint is reported as a host-inclusive cost rather than being attributed entirely to extension code. This matters because transparent performance reporting should distinguish the customer-observed browser footprint from the portion controlled by an individual extension.
Cortex VS Code Extension: Intelligence That Respects the Editor
Developers rarely use an AI coding assistant in isolation. The editor is already managing repositories, language services, terminals, tests, debugging, source control, and other extensions.
Cortex 6.5 reduces the Cortex VS Code Extension package by 46.0%, the extension-host artifact by 55.5%, and the chat render by 64.3%.
The functional goal is not merely a smaller download. It is a narrower path from opening Cortex to continuing the current engineering conversation. Active work is prioritized, rendered history stays bounded, and optional capabilities do not need to remain active simply because VS Code remains open.
That allows Cortex to stay close to the repository without treating every editor session as a request to load the entire AI capability set.
Cortex Mobile: The Active Conversation First
Mobile performance is shaped by tighter memory, battery, and background-execution constraints. The customer usually wants one thing immediately: the active conversation.
The Cortex Mobile startup-reachable path falls from 338,452 bytes to 442 bytes, a 99.9% reduction. The full iOS JavaScript bundle remains essentially flat at 1.97 MB, so the claim is intentionally specific: the improvement is in what becomes reachable at startup, not a claim that the entire application became 99.9% smaller.
This separation matters. Cortex Mobile can preserve a broad capability set while keeping the initial route focused on resuming work. When the application backgrounds, network and media activity return toward idle rather than remaining active by default.
Cortex CLI: Fast Entry Without Artificial Size Claims
The Cortex CLI already had a small entry path, so Cortex 6.5 does not manufacture a reduction headline where one is not supported. The 28 KB fast entry remains within its target while larger commands and heavier workflows stay demand-loaded.
| Cortex CLI workload | Qualified result |
|---|---|
| Fast entry | 28,004 B |
| First nonblank frame p95 | 119.2 ms |
| Maximum idle RSS | 46.0 MiB |
| 2,000-message formatting p95 | 99.3 ms |
| Cached redraw | 0.62 ms |
For terminal users, this preserves the quality that matters most: Cortex can enter the workflow quickly, remain bounded while idle, and format a substantial conversation without turning routine navigation into a long wait.
Two Hours of Evidence, Not One Favorable Moment
Discover receives the deepest runtime qualification because it owns both the browser shell and the integrated AI surface. The Cortex 6.5 campaign therefore repeats panel-closed, panel-open, and rapid open-close conditions across a 121.4-minute Apple Silicon qualification.
| Long-session evidence | Result |
|---|---|
| Measured sessions completed | 9 of 9 |
| Workload errors | 0 |
| Native sampling errors | 0 |
| Render readiness p95 | 25.8 ms |
| Healthy repeated-cycle samples | 903 of 903 |
| Settled transfer growth | 0 B |
| Sessions with retained process growth | 0 |
The important result is not merely that the panel opened quickly once. It remained ready through repeated use, completed every measured workload, and did not retain additional processes after closure.
The memory analysis also separates the tested Discover experience from behavior attributable to the unchanged browser foundation. At the final comparison point, the campaign found no positive Discover-specific excess beyond the matched browser mechanism. That is a bounded qualification result, not a universal claim that memory can never grow under a different workload.
A Performance Contract for Work That Moves
The defining change in Cortex 6.5 is not one percentage. It is the extension of a shared performance contract across the places where customers use Cortex.
- In Discover, substantial history becomes ready quickly and stays responsive.
- In the Browser Extension, the active panel is no longer preceded by the weight of the entire capability set.
- In VS Code, Cortex remains close to the repository while respecting an already-busy editor host.
- On Mobile, the active conversation comes first within a tightly bounded startup path.
- In the CLI, a small fast entry remains small while large conversations stay workable.
This is a different stage of the Cortex performance story.
The first studies asked whether an agentic browser could be efficient. Later releases showed that it could remain healthy through sustained use and become lighter as histories grew. Cortex 6.5 now applies those lessons across a connected client portfolio.
The customer should not have to think about which capability was loaded, which interface owns the conversation, or whether AI is still consuming resources after attention moved elsewhere. The experience should simply be ready when needed, responsive during work, and quiet afterward.
Cortex 6.5 and Discover 1.7. Performance that follows the work, without following the user as background weight.
Read the Performance Research
The Cortex Clients: Performance and Resource Efficiency White Paper – Version 5 documents the qualified workloads, declared limits, long-session campaign, evidence boundaries, and complete measured results behind this update.
The published measurements apply to the stated Apple Silicon qualification environment and controlled workloads. Device hardware, operating-system policy, extensions, account history, and concurrent applications can change observed results.
Earlier performance articles:


