Flagship effort caps
Improved
- Effort is capped at medium on
Claude Opus 5.5, GPT-6 Sol and Sol Pro to keep costs down
Arrivals stays
Improved
- The arrivals board now stays on the home screen
- Picking a row switches the model and the board stays in place
Arrivals
New
- The home card becomes an arrivals board for
Claude Opus 5.5 and GPT-6 Sol - Pick a row on the board to switch to that model
New flagships
New
Claude Opus 5.5 and GPT-6 Sol join the main model list- Grok 4.7 takes the xAI seat; GPT-6 Luna, MiMo V2.6 and DeepSeek V4.1 Flash arrive
- Command A+, Fugu Max and Ultra v2, Ember-1, Qwen3.8 Omni Flash, Ling VL and Solar Mini 4
- Space Bunny Alpha joins the free models
Fixed
- Mercury 2.5 runs on the released model instead of the withdrawn preview
- Conversations started on Opus 4.8 Fast keep working on Opus 4.8
Effort, delivered
Fixed
- Your effort level now reaches the model when code execution is switched on
- Top effort levels are no longer quietly lowered on models that run code
The pair
New
- The home card now presents Claude Fable 5.1 and GPT-6 Astra together
Improved
- Effort is capped at medium on the two flagship models to keep runs affordable
- The effort dial shows which levels a model will actually use
Astra
New
- GPT-6 Astra and Astra Pro join the roster; Astra takes the OpenAI seat in the picker
- Ling 3.0 Flash Sante, a health-focused model, joins the free tier
Fixed
- Qwen3.8 Max follows its upstream snapshot rename
- Two MiniMax free routes withdrawn upstream are locked, not removed
Tuned
New
- The Fable 5.1 notice now names your current model and carries the effort dial
Improved
- Effort defaults to medium; any level you chose is kept
- Usage shows four decimals below ten cents
Findable
New
- Every release now has its own page, for example /changelog/v15-413
- The home page describes the app properly to visitors who are not signed in
Improved
- Privacy and Terms live at one address; the old links land on the right section
- Docs pages carry their own titles, headings and share previews
Fable 5.1 takes over
Models
- Claude Fable 5 is retired in favour of 5.1; older chats still replay on it
New
- A short Fable 5.1 note on the home screen, dismissible
Fable 5.1 and a quieter memory
Models
- Added Claude Fable 5.1, now first in the picker
- Added Muse Spark 1.3, Gemini 3.8 Flash, Mercury 2.5, Hy4 Preview and Granite 4.2
KAT-Coder Air went offline upstream; its row stays for older conversations
Improved
- Memory now works quietly instead of announcing itself
- Spend on OpenRouter models is now booked at the exact amount OpenRouter charged
Fixed
- Conversations with memory and web search on used far more tokens than they needed to
Code execution everywhere
New
- Code execution now works on nearly every model, not only Claude
Search everywhere it works
Fixed
- Web search was greyed out on seventeen models that support it, including
GPT-5.6 Sol and the Sonar search models - Code execution now shows correctly on older Opus 4.8 conversations
One reasoning control
Improved
- Adaptive thinking is retired, so Effort is now the single reasoning control
Fixed
- Effort above High was quietly capped on Claude if Adaptive thinking was off
Effort on every model
Improved
- Reasoning effort can now be set on almost every model, not just Claude
- Code execution is available on Claude Haiku 4.5
Fixed
- Your effort choice was ignored outside Claude and every reply ran at high
Steadier under load
Fixed
- Fixed the app dropping out and asking you to sign in again during a busy session
- Memory is read less often, so busy conversations stay responsive
Memory writes
Fixed
- Fixed a reply vanishing after the assistant saved something to memory
- Saving to memory is much faster
Memory everywhere
New
- Memory now works on every model, not only Claude
Also
- The smallest models can read your saved facts but not add to them
Writing tools
Fixed
- Fixed the grammar and prompt-improvement tools returning no result
Community reopens
New
- Community threads are open again
Fairer limits
Improved
- Free models now have fair-use limits
- Shared pages open faster and large attachments start arriving sooner
Fixed
- Fixed attachments on shared pages showing placeholder text
- Fixed image generation stopping early on longer batches
Security
- Routine hardening across sign-in and shared links
Flagships up front
Models
- Model picker now leads with Claude Fable 5, Grok 4.6, GLM 5.3, Gemini 3.7 Flash and Qwen3.8-Max
- Previous versions remain under More models; a saved older model stays selected
Code execution and roster updates
Improved
- All Claude models now run through OpenRouter
- Code execution runs in a sandboxed Linux shell; Python is available, internet is not
Models
- Claude Fable 5 is now unlocked and selectable
Devstral 2 is back online at its provider- Added
MiniMax M3 and M2.7 (free; M3 with 1M context), Ling 3.0 Flash Fin (free), GLM 5.3 Flash and Qwen3.8 Flash - Ox Alpha and
GPT-OSS 20B went offline upstream; their rows remain for older conversations
Fixed
- Fixed code execution failing with a credit error
Also
- Files created by code runs are temporarily not attached; printed output still shows in full
Eleven more models
Models
- Added free models Ox Alpha (
1M context, vision, tools), Dots3 Note Preview (vision, 512K) and LFM2.5 - Added Muse Spark 1.2 Contributor, Muse Glimmer 30B,
DeepSeek V4 Flash Vision and Solar Pro 4 - Added Qwen3.8 27B, the Qwen3.8 flagship, ByteDance Seed 2.1 Turbo and Seed 2.0 Code
Fixed
- Enabled web search on Gemini 3.7 Flash, Grok 4.6, Sakana Namazu and Qwen3.8 Max
Rendering and search fixes
Fixed
- Fixed reply text rendering twice around parentheses and dashes
Improved
- Web search now works on every Claude model
Dark follows your system
New
- The app follows your system dark mode until you pick a theme yourself
Models
- Added Grok 4.6, Gemini 3.7 Flash, GLM 5.3, Qwen3.8 Max and Sakana Namazu; Muse Spark and DeepSeek V4 Pro moved to latest releases
Improved
- Claude models now run through OpenRouter with the rest of the roster
Fixed
DeepSeek V4 Flash and V4 Pro now bill at their catalogue ratesDevstral 2 was removed by its provider; its row remains for older conversations
Model availability
New
- Home page now offers free models when paid models are unavailable
Fixed
- Replaced Ling 3.0 Flash with Nemotron 3.5 Lightning (free,
1M context)
Model picker dropdown
Fixed
- Model picker opens as a compact dropdown and no longer overlaps the header
Claude is back
Models
Claude Opus 5, Opus 4.6, Sonnet 5 and Haiku 4.5 are available again
Reliability and accessibility
New
- Installable as an app from a phone home screen or the browser
Fixed
- Fixed duplicate sends from a fast double Enter
- Fixed files from one chat overwriting another when switching quickly
- Fixed Stop corrupting the saved partial reply
- Fixed the composer sitting behind the browser toolbar on phones
Improved
- Exported transcripts now name every model in the roster
- Improved keyboard and screen-reader support across chats, dialogs and the composer
Stone theme
New
- New Stone theme replaces Glacier; Glacier users are moved over automatically
Improved
- Restyled the See what’s new button; the changelog back link returns to the page you came from
Documentation
New
- Nine documentation pages are now available, linked from the sidebar and account menus
Improved
- Tightened the Settings panes and corrected an outdated theme count
Gray default theme
New
- Gray is now the default theme; a theme you pick yourself still sticks
Improved
- See what’s new button on Gray now matches Paper’s quieter ink style
Fixed
- Fixed a white flash on page load in the dark theme
Sitewide announcement strip
New
- Announcement strip now appears on every page, themed to match each palette
- Privacy, Terms and changelog links moved into the strip, replacing the footer and home cards
- LATER hides the strip for this visit; NEVER keeps it hidden until the next release
Improved
- Home particle field is off by default; enable it from the panel rail or Settings
- Shared conversation pages no longer show the strip
Fixed
- Fixed touch targets for the strip’s controls on phones and tablets
Theme lineup and contrast
Improved
- Theme family settles at four: Paper, Gray, Glacier and Slate
- Paper restored to its exact source palette
- Re-measured contrast across all four themes and fixed failing pairs
Fixed
- Fixed unreadable hover states on the strip’s LATER and NEVER buttons in light themes
Answer formatting
Improved
- Answers now render with structured headings, links, inline code and quotes
Fixed
- Bold text no longer appears darker than headings
- Fixed headings falling back to body style on shared and read-only pages
Accurate usage figures
Fixed
- Fixed the Messages figure counting billable actions instead of messages sent
Improved
- Messages is now computed from stored records and counts only messages you sent
- Each headline figure states the definition it was measured under
Also
- Token and spend figures are unchanged
Chats name themselves
New
- Chats are now titled from their opening exchange; Arabic prompts get Arabic titles
Also
- Naming runs on a small fast model and doesn’t count as a message
Announcement banner and default theme
New
- Announcement banner at the top of the app, dismissible per visit or permanently in Settings
- Paper is now the default theme; existing accounts move over once
Improved
- Banner renders consistently on every theme; on phones it truncates and dismiss opens on tap
Fixed
- Fixed banner height and a dismiss control shown when signed out
Sign-in refinements
Improved
- Redesigned sign-in with a single-line email field and unified copy
Fixed
- Fixed page zoom on iPhone when focusing the email field
Particle controls
New
- Home particles can be turned off, from the home page or in Settings
Improved
- Restyled home particles to a finer, calmer look with distinct button icons
Model search
New
- Search in the model menu; typing filters the full roster
Home page particle field
New
- Interactive particle field on the home page, responding to mouse movement and clicks
- Glass message box with a placeholder drawn from the particle field on hover
Improved
- Home animation pauses when the tab is hidden or while typing
- Deepened six themes: Sora, Glacier, Warm, Paper, Night and Ink
- Simplified the docs shortcuts on the home page to two rows
Fixed
- Fixed the model list running off screen and faint stars on night themes
Translate and settings redesign
New
- Translate drafts to English, Arabic, Spanish, French or Italian from the composer
- A + menu on the composer gathers starter prompts, Improve prompt, grammar and translate tools
- Copy button on your own messages
- New themes: Paper, Ink, Mist and Sora
Improved
- Settings redesigned on desktop; phones get full per-section settings
- Fix grammar sticks to correcting; grammar and Translate run on a faster model
Fixed
- Fixed mid-stream chat deletion and low-contrast colours on dimmer themes
Models
Kimi K3 is the default for new chats, with web search on from the startDeepSeek V4 Flash added to the main model list
Also
- Retired the Sky, Sage, Void, Hush and Dark themes; existing users migrate automatically
Theme readability
Improved
- Labels, toggles and gradient text retuned for readability on Glacier and Hush
- Focus outlines, text selection and scrollbars now follow the active theme
Fixed
- Fixed unreadable code in research traces on Night and Void
- Fixed faint error messages and dim secondary text on the dark themes
- Dialogs on Night and Void darken the page behind them again
Every theme redrawn
Improved
- All nine themes rebuilt so each is visually distinct and purpose-built
- Glacier is now deep violet; Hush is dimmer and cooler
- Theme buttons describe each theme on hover
Fixed
- Fixed faint delete and warning colours on the Dark theme
- File drop zone now has its own background in every theme
New themes and instant loads
New
- Two light themes, Sage and Hush
- Privacy Policy and Terms rewritten in plain language on a single page
Improved
- Chats, shared conversations and community threads now open instantly
- Code execution keeps its sandbox for the whole conversation; files and packages persist
- Redesigned Changelog, Privacy and Terms, and Community around a pinned contents index
- Clearer headings and more readable reply text
Fixed
- Fixed code execution failing when web search ran in the same turn
- The assistant no longer contradicts its own earlier work in a conversation
Also
- Collapsed sidebar adds quick links to Settings, Community, Changelog and legal pages
Structured reasoning traces, stream recovery
New
- Reasoning traces now stream in titled sections and collapse into an expandable outline
Improved
- Improved syntax highlighting and a redesigned artifact card with live write status
- Dropped connections now retry automatically
- Retuned code execution and web search across frontier models
- Cleaner sign-in with inline errors and code entry
Fixed
- Stopped responses keep all streamed content, and Continue resumes from it
- Extended thinking now displays on
Claude Opus 5 - Fixed the free Poolside coder route; saved conversations work again
Models
- Added
Qwen3.7 Flash (1M context, low cost) - Added
GPT-5.6 Terra Pro and Sol Pro alongside Luna Pro Gemini 3.5 Flash Lite replaces the 3.1 tier
Voice input, seven themes, 100 templates
New
- Voice input in English and Arabic (Egyptian, Syrian and Gulf variants)
- Transcripts are cleaned up for review before sending; nothing sends automatically
- Grammar button fixes spelling, punctuation and grammar without changing meaning
- Seven themes, switchable from Settings, the artifact panel and shared chats
- Mic has a language switch;
Ctrl+Shift+M starts listening and Esc stops
Improved
- Template library rewritten and expanded from 43 to 100
- Better contrast on every theme
- Redrawn sidebar with grouped navigation
- Voice capture no longer drops the last phrase
Models
Kimi K3 moved to the top of the model picker- Every image model now edits attachments
- Image batch generation in sets of 1, 2, 4 or 10
Improve Prompt, six new models
New
- Improve Prompt wand rewrites a rough prompt in any language and returns it for editing
Improved
- Redrawn composer and a larger, more readable reasoning trace
- Sending scrolls straight to the reply
- Templates reorganized into scrollable categories, opened from the lightbulb
Fixed
- Long conversations no longer drop replies on large contexts
- Code view wraps long lines
- Usage meter shows the exact enforced limit
Models
- Added
Inkling, Thinking Machines’ debut model - Added
Kimi K3, GPT-5.6 Sol, GLM 5.2, MiniMax M3 and KAT-Coder Air - Free tier refreshed with
Tencent Hy3, Nemotron 3 Ultra (1M context) and GPT-OSS 20B; retired routes replaced
25 models, image editing, Fusion
New
- Edit generated images by asking in chat
- OpenRouter Fusion queries several models and blends the strongest parts
- Global Radio: live stations on a spinning globe
- One-tap starter prompts on home, each pinned to its best model
Improved
- Faster model switching and a tidier mobile chat layout
Models
- Added 25 models across reasoning, coding, vision and web research, several free
Also
- Assorted small fixes throughout
Coding and reasoning models, security pass
Security
- Hardened file downloads and shared-link previews
- Broader new-location sign-in alerting
- More accurate usage limits
Models
- Added
Kimi K2 Thinking, Codestral and Devstral 2 for multi-step coding and reasoning
Per-chat links, reliability fixes
New
- Every conversation has its own URL that can be bookmarked and shared
Fixed
- Replies that hit the length limit say so and can continue
- Memory works reliably across Claude models
- Stopping a response cancels it cleanly
- Image and PDF uploads reach more models
- Fixed code-viewer stalls on large files and wrong model labels on shared conversations
- Steadier mobile layout and quieter theming on legal and changelog pages
Security
- Security and reliability hardening across the app
GPT-5.6, Grok 4.5, doubled limit
Improved
- Monthly limit doubled to $10
- Web search and extended thinking are off by default for new chats
- Instant usage checks
- Code viewer and phone files-panel refinements
Fixed
- Fixed web search on models where it had stopped working
Models
- Added the
GPT-5.6 family: Sol, Terra, Luna - Added
Grok 4.5, Gemini 3.1 Pro, Mistral Medium 3.5, Hunyuan 3, Laguna XS 2.1 and free North Mini Code - Free models available after the monthly cap
Stats, handoffs, file viewer
New
- Summarize any conversation into a shareable
HANDOFF.md brief - Active Users stat on the platform card
Improved
- Usage and platform stats load instantly
- Files panel redesigned: line numbers, markdown previews and large-file support
Sonnet 4.6, image generation, page previews
New
- Generate images inside a chat
- Web pages built by models appear as previewable files, openable in the browser
Improved
- Attachments are saved with the chat and included when shared
- Conversation files easier to find in the files panel
- Dark-mode and sign-in refinements
Models
- Added
Claude Sonnet 4.6 - Wider model lineup, several free
- Web search on more models, with sources in answers
Also
- Reliability and polish fixes throughout
HTML preview and model picker
Improved
- HTML preview matches real browser rendering, including interactivity and animations
- Artifacts panel on all screen sizes and shared chat links; auto-opens on phones during streaming
- Model dropdown rebuilt:
Opus 4.7, Sonnet 4.5, Gemini 3.1 Pro, GPT-5.5 and Grok 4.3, with more in a flyout - Research mode moved into the model dropdown
- Empty chat screen simplified to just the input bar
- Files being written show a shimmer on their card
Community launch, Qwen 3-Max
New
- Community: share threads, comment and follow discussions from the sidebar
- Skills, attachable knowledge libraries, rolling out soon
Improved
- Raised AI credit allowances across the local model roster
Models
- Added
Qwen 3-Max for multilingual reasoning and math - Cross-conversation memory now limited to
Opus 4.7 and Sonnet 4.5
Web search on by default
Improved
- Web search on by default for
Opus 4.7, Sonnet 4.5, GPT-5.5 and GLM-4.6, with cited sources