Changelog
Products
CruxAI: A major update for learning, creation, voice, and faster everyday chat
This update for CruxAI features numerous changes and updates to improve the experience for every day users.
CruxAI Summary
CruxAI introduces a School workspace for managing coursework, flashcard study tools, the faster CRX-4.1 model lineup with adjustable reasoning effort, improved memory controls, context-aware health notices, simplified web voice features, and performance optimizations making Fast mode up to 3.5x faster.
This update for CruxAI introduces School and study tools, the CRX-4.1 model lineup, contextual health notices, an adaptive intelligence determination model, a redesigned web voice experience, unified branding, broad accessibility and privacy improvements, and a dedicated low-latency path for Fast mode.
A more useful place for schoolwork
- CruxAI now includes a dedicated School workspace for organizing coursework and getting help with it.
- Create and manage classes with subjects, teachers, rooms, colors, and related details.
- Track assignments with due dates, status, priority, type, notes, search, and class filters.
- View upcoming, due-today, next-seven-days, overdue, completed, and all assignments.
- Open an assignment directly in School Help with its class and assignment context attached.
- Choose between Tutor and Answer styles.
- Tutor mode provides progressive hints and withholds the final answer until the student explicitly requests it.
- School Help supports questions, worksheets, photos, PDFs, essays, research, and homework.
- CruxAI can suggest assignment status or note changes, but no school record changes until the student confirms them.
Flashcards and study tools
CruxAI on the web now supports Flashcards and study sets.
- Generate flashcards from pasted notes, images, PDFs, text, Markdown, and CSV files.
- Choose a set name, card count, and difficulty.
- Save and reopen flashcard sets across sessions.
- Flip cards, shuffle a set, move between cards, track progress, and rate your confidence.
- Delete saved sets when they are no longer needed.
- Use the feature with keyboard controls as well as mouse and touch input.
Expanded, safer memory controls
CruxAI’s memory system now supports substantially more context while keeping the active prompt bounded.
- About You, Interests, and active-memory fields accept much larger entries.
- Memories are stored as structured records with categories, sources, timestamps, stable keys, and version history.
- Active, superseded, pending, and archived memories are separated clearly.
- Only active and relevant memories enter a response; archived and superseded history is excluded.
- Older line-based memories migrate into the structured system.
- New non-conflicting facts can be saved automatically.
- Corrections, conflicts, edits, and forgetting requests require confirmation before changing active memory.
- Settings now includes character counts, Active and History views, edit/archive controls, timestamps, and memory history in profile exports.
Meet CRX-4.1
The model lineup has been updated across CruxAI.
- CRX-4.1 Fast is now the everyday option for free, Plus, and guest users.
- CRX-4.1 Think is designed for questions that benefit from additional reasoning.
- CRX-4.1 Pro is intended for the most demanding work.
- Retired Fast generations are no longer offered as an accidental default to free users.
- Plus members retain access to supported model archives.
- Existing free Think and Pro usage limits remain in place.
- Site Builder’s visible model selection now matches the model actually used for free accounts.
Adjustable reasoning effort
CRX-4.1 Think and Pro now include a separate four-stop effort slider:
- Light
- Medium
- High
- Extra High
Medium is selected by default, and the choice is remembered separately for each conversation. The control is available through a compact, viewport-aware menu that remains usable on smaller screens. Light provides fast responses for straightforward tasks, while Extra High uses additional time and reasoning capacity to deliver deeper accuracy and nuance for complex problems, which may count more against your usage limits depending on the selected model.
More context-aware health notices
Health notices are no longer selected by simple keyword matching. CruxAI evaluates the meaning and context of each distinct request with a separate classifier.
- Text, images, PDFs, and other supported readable uploads can be evaluated.
- Medical content inside an uploaded image cannot bypass the notice system merely because the accompanying text is vague.
- Ordinary educational content, including non-medical statistics worksheets, is less likely to trigger an irrelevant notice.
- Notices are appended to the bottom of the completed response.
- The classifier runs after primary response generation, so it does not block the first streamed text.
- If classification fails, the assistant response still completes normally.
Adaptive Intelligence Determination
We've developed a new classification model that works in tandem with CRX-4.1 Fast to determine if your response needs a higher intelligence level.
- Responses that require higher intelligence like math, coding, or writing will automatically switch to the smartest available Think model.
- Users have the ability to disable this in settings and require CruxAI to ask permission before switching to a higher intelligence model.
- CruxAI respects your decision. If you switch back to Fast model after it automatically switches you to a higher intelligence model, it will continue to ask you if you want to switch to a smarter model, when necessary, for that specific conversation.
Web Live Mode replaced with focused voice controls
The web version no longer exposes Live Mode. The richer Live backend remains available for the iOS app.
On the web, Live Mode is replaced by two clearer controls:
- Dictation lets users speak into the composer and edit the transcribed text before sending it.
- Read aloud plays an assistant response on request.
Voice behavior is now more predictable: dictation never sends automatically, playback can be interrupted, only one response can play at a time, and browser speech provides a fallback when service audio is unavailable or a response is too long.
Faster Fast-mode responses
CRX-4.1 Fast is up to 3.5x faster than CRX-4 Fast and earlier. This new version now follows a dedicated low-latency request path for everyday chat.
- Fast uses a substantially smaller purpose-built instruction core while retaining safety, upload, freshness, tool, and health-notice requirements.
- Stable Fast instructions, tools, profile context, and conversation context are arranged for better prompt-cache reuse.
- Large tool catalogs can be discovered semantically instead of loading every definition into the initial request.
- Web search and relevant product knowledge remain available when needed.
- Short greetings and acknowledgements can finish with a smaller output ceiling, while substantive and upload-heavy requests retain larger limits.
- Irrelevant changing context, such as the exact local clock, is omitted unless it matters.
- Recently used memory rows may be served from a short-lived server cache without changing the durable memory source of truth.
- The first streamed text delta is rendered immediately.
- Streaming transport settings reduce avoidable buffering where the server and network support it.
- Privacy-safe latency timing helps measure first byte, first text, preparation, classifier, and total response time without recording prompts or answers.
These changes are intended to make simple Fast requests feel more immediate. Provider queueing, network distance, tools, uploads, and response length still affect actual response time.
A more consistent interface
The update includes a broad interface and interaction pass across chat, School, Site Builder, Flow, Library, Developer, Operations, and signed-out pages.
- The solid ivory three-point CruxAI mark is now used consistently in chat, assistant response icons, logged-out pages, Site Builder, developer and operations areas, the offline screen, favicons, PWA icons, and social previews.
- Sidebar organization is clearer, with improved spacing and scrolling on shorter desktop displays.
- Mode and effort menus position themselves within the available viewport instead of being clipped below the composer.
- Mobile layouts are optimized for narrow 320–360 px screens, rotation, safe areas, and coarse touch input.
- Primary and secondary touch targets are sized for reliable mobile use.
- Canvas panels are constrained to the viewport and recover correctly after resize or rotation.
- Empty Send states are explicit rather than ambiguous.
- Temporary Chat messaging accurately describes what is and is not saved.
- Custom Apps correctly show Fast mode when they force Fast mode internally.
- Dynamic app reviews, ratings, emoji controls, filters, tabs, and generated cards now match their intended styling and interaction behavior.
- Site Builder failed-template states render as proper error states instead of raw or unstyled markup.
Accessibility improvements
The web interface now has stronger support for keyboard users, screen readers, high-contrast modes, and reduced-motion preferences.
- Added skip navigation to authenticated, signed-out, and developer surfaces.
- Added or normalized accessible names for search, composer, form, dialog, iframe, and secondary controls.
- Corrected tab, tabpanel, filter, radio-group, pressed-state, and selected-state semantics.
- Added arrow-key navigation where tab interfaces require it.
- Removed inactive controls from inappropriate keyboard navigation paths.
- Added visible focus treatment across primary and secondary surfaces.
- Dialogs now isolate background content, trap focus, close with Escape, and restore focus after closing.
- Nested School, Site Builder, Artifacts, and Focus Mode dialogs receive the same focus behavior as top-level dialogs.
- Improved normal-text contrast across light and dark themes, marketing pages, developer tools, operations pages, reviews, offline views, and the main workspace.
- Added explicit support for forced-colors and reduced-motion preferences.
- Replaced clickable non-controls with proper buttons or links where appropriate.
These are technical accessibility improvements aligned with common WCAG, ADA, and Section 508 implementation practices. They are not a legal certification by themselves.
Privacy, security, and reliability improvements
- Removed unnecessary remote font requests from web surfaces.
- Microphone access remains limited to the same origin; unused camera, location, payment, and USB capabilities are disabled.
- Web Read aloud audio now uses private runtime storage rather than a public document directory.
- Audio responses use no-store behavior and short-lived private caching.
- Added configurable retention and cleanup for web speech cache data, logs, sign-in approval metadata, and security metadata.
- External new-tab links use opener and referrer isolation.
- Developer API keys are removed from the page when their dialog closes.
- Upload and request limits were tightened in privacy-sensitive speech paths.
Bug fixes and correctness improvements
- Fixed valid first-party model overrides being discarded before request processing.
- Fixed bounded conversation history being lost before health classification and completion handling.
- Fixed tool-continuation replay so provider content blocks retain their original order.
- Fixed deferred custom-App usage accounting from delaying visible response delivery.
- Fixed Site Builder’s free-account model selection mismatch.
- Fixed clipped response-mode and reasoning menus.
- Fixed misleading effort-selection states and replaced the checklist presentation with a slider.
- Fixed custom Apps showing Think controls while internally forcing Fast mode.
- Fixed incorrect Temporary Chat persistence wording.
- Fixed stale competing CruxAI logo markup and cache paths across signed-out pages and response previews.
- Fixed missing PWA service-worker registration and a broken installed-app shortcut route.
- Fixed inaccessible search and composer controls that depended on placeholder text instead of programmatic names.
- Fixed inconsistent tab and filter semantics, including inactive items remaining in the tab order.
- Fixed nested modal focus trapping and background isolation.
- Fixed unstyled generated review/rating controls and mouse-only review cards.
- Fixed an emoji-picker class mismatch.
- Fixed developer navigation links that lacked usable link targets.
- Fixed the Site Operations drawer’s missing modal behavior.
- Fixed untrusted Canvas previews from combining script execution with same-origin access.
This new update turns CruxAI into a more complete workspace for conversation, schoolwork, study, projects, and voice. It adds major user-facing capabilities while also addressing the less visible problems. Message speed, keyboard behavior, mobile layout, privacy boundaries, caching, and state consistency.
