Changelog

Gemini 4 Hydrogen joins the model picker, Auto sends more of your work to it, and memory commands answer faster.

Gemini 4 Hydrogen

Gemini 4 Hydrogen joins the model picker, between Gemini 4 Carbon and Gemini 4 Argon. It’s Google’s fast multimodal model for reasoning, coding and agentic work.

• Reasoning through X-High. Choose Low, Medium, High or X-High. Gemini 4 Hydrogen has no fully-off setting, so with reasoning off it runs at Low.

• Reads images, PDFs, audio and video directly, with a context window of about 1 million tokens.

• Uses less of your API usage budget per token than Gemini 4 Carbon and Gemini 4 Argon.

Compare current models.

Auto uses Gemini 4 Hydrogen

We tested Gemini 4 Hydrogen in Auto against the models it would replace, and it now handles more of your work.

• Math and logic of moderate difficulty now go to Gemini 4 Hydrogen, which answered them more accurately in testing.

• Writing of moderate difficulty, such as emails, posts and short documents, now goes to Gemini 4 Hydrogen, which wrote better in testing and uses less of your API usage budget. These replies can take a few seconds longer to start.

• Coding of moderate difficulty, from writing code and building front ends to fixing bugs, now goes to Gemini 4 Hydrogen, which did as well or better in testing and uses less of your API usage budget.

• Research and planning of moderate difficulty now go to Gemini 4 Hydrogen, which was rated as well as before for about a quarter of the API usage budget. Plans can take longer to start.

• Backup choices. When Auto’s first choice for a message isn’t available, more kinds of work now fall back to Gemini 4 Hydrogen instead of Gemini 4 Argon.

Nothing changes in how you use Auto. Learn how Auto works.

Also in this release

• Memory commands answer faster. A “remember…” or “forget…” command typed in memory settings is now read by Gemini 4 Hydrogen, which answered every ordinary command in testing within 8 seconds, where some took Gemini 4 Argon too long to answer.

We tested how Auto handles each kind of task against the newest models, comparing answer quality, speed and cost. Auto now sends more of your work to the models that came out ahead.

• Hard problems go to Claude Fable 5.2. Most demanding requests, from research, planning and math to science and coding, now go to Claude Fable 5.2, which answered them best in testing.

• Demanding writing and database work. Long or careful writing now goes to Claude Sonnet 5.5, and complex database design to GPT-6.1 Sol. Both answered these best in testing, and Claude Sonnet 5.5 uses less of your API usage budget for writing than before.

• Science, finance and legal questions start on Claude Fable 5.2, even when they aren’t the hardest kind, because it answered them best. These replies use more of your API usage budget than before.

• Everyday work uses less of your budget. Most conversations, questions, summaries, translations, SQL queries and data extraction now go to GPT-6.1 Sol, which answered them better in testing and uses less of your API usage budget.

• Quick factual questions are more accurate. Short questions of fact now go to GPT-6.1 Terra, which gets more of them right.

• Code review and configuration. Code reviews now go to GPT-6.1 Terra, and server, deployment and configuration work to Gemini 4 Carbon. Both scored higher in testing and use less of your budget.

Nothing changes in how you use Auto. Choose Auto in the model picker as before, and check Auto · Models used: below the composer to see which models answered. When your budget runs low, Auto still favors models that use less of it.

Learn how Auto works.

Faster artifacts

Documents, code and sheets in Canvas are now written by GPT-6.1 Terra, a fast writing model, from the request your model makes. The model you chose, or the one Auto picked, still decides what goes in them. This applies to every model.

• Much quicker. In testing, a typical artifact arrived in under a minute, against two to six minutes before.

• Large builds finish. Writing now has up to about six and a half minutes, up from three, and large app builds that used to run out of time finished in testing.

• Less of your API usage budget. Writing an artifact counts at GPT-6.1 Terra’s rate.

• Visual artifacts are unchanged. Babbily draws them as before. Use Canvas and artifacts.

Claude Fable 5.2, Gemini 4 Carbon and Muse Ember join the model picker, DeepSeek V4.1 Flash 1017 gets an even quicker speed, and video generation moves to a new model that adds sound.

Claude Fable 5.2

Claude Fable 5.2 takes Claude Fable 5.1’s place in the picker. It’s built for agentic coding, knowledge work and research, and it always thinks before it answers, adjusting how much to the task.

• Reasoning through Max. Choose Adaptive, Low, Medium, High, X-High or Max. Claude Fable 5.2 has no fully-off setting, so with reasoning off it runs at Low.

• Fast mode. Select the lightning control near the composer for quicker replies. Fast mode uses more of your API usage budget.

• A 1-million-token context window, and it reads images and PDFs directly.

• Claude Fable 5.1 is now Sunsetting. It stays selectable for now. Choose Claude Fable 5.2 for new work.

Compare current models.

Gemini 4 Carbon and Muse Ember

• Gemini 4 Carbon is Google’s reasoning model for complex software engineering, sustained agentic work and cybersecurity defense. It reads images, PDFs, audio and video directly, supports Low, Medium and High reasoning effort, and sits above Gemini 4 Argon in the picker.

• Muse Ember is Meta’s multimodal model for conversation, reasoning and agentic work, with a Reasoning toggle. It sits below Muse Wonder and uses little of your API usage budget.

Both have a context window of about 1 million tokens.

Ultrafast

DeepSeek V4.1 Flash 1017 now has a third speed. Select the lightning control once for Fast mode, and again for Ultrafast, its quickest setting. Select it a third time to return to standard speed.

• Know the cost before you send. When you choose Ultrafast, Babbily shows how many times more of your included usage it uses than standard speed.

• Attachments still work. In Ultrafast, the model reads images itself. Babbily reads PDFs and other documents first and passes along what it found.

• Like Fast mode, it’s a setting on the model. It’s hidden while Auto is selected, and switching to a different model returns to standard speed.

Use Fast mode and Ultrafast.

Video generation on Gemini Omni 1.2 Flash

The Video generation tool now uses Gemini Omni 1.2 Flash. You reach it the same way from any chat model: ask in Auto tool mode, choose Manual > Video generation, or type /create-video.

• Videos with sound. Choose a length from 4 to 8 seconds and a landscape (16:9) or portrait (9:16) shape. Videos are 720p and include sound. Square videos are no longer offered.

• Animate an image. Attach one image and describe how it should move.

• Edit a video up to 10 seconds long. Attach one MP4 or MOV video and describe the change. The edit keeps the video’s length and framing.

Create images and videos.

Also in this release

• Selected tools work with Muse models. A tool you choose in the composer, such as Manual > Web search or /create-image, now works with Muse Wonder and Muse Ember.

• Notifications in your language. When a chat, image or video finishes in the background, its notification now appears in the language you chose under Language in your account menu. Manage notifications.

You can now talk with Babbily out loud and hear it answer, and its replies can show charts, diagrams, maps and interactive explainers right where they help. Memory, long chats and Auto get better too.

Talk with Babbily

Select Use Voice Mode (the sound-wave button) in the composer to start a spoken conversation. Babbily listens, answers out loud, and can look things up while you talk.

• Speak naturally. Babbily lets you finish before it answers. To cut in, just start talking: it stops and listens.

• Answers that look things up. Ask about the weather, the news, a stock price or anything that needs a search. Babbily hands lookups and detailed questions to your chat model, then tells you the answer. On Auto, the call shows which model it picked for lookups.

• Your call becomes a chat. The transcript appears as you talk, with the word being spoken highlighted. When you select End, the conversation opens as a regular chat you can keep going by typing.

• 31 voices. Choose from voices with North American, British, Irish, Australian, Southern US and Filipino accents, plus native voices for German, French, Brazilian Portuguese, Hindi, Japanese and Korean.

• Talk about what you’re reading. Start voice from a News article or a Finance page to ask about it. Anything that needs your approval, such as a Plugin action, shows a card during the call.

• Calls that look after themselves. If the connection drops, select Continue to pick up where you left off. If nobody has spoken for three minutes, Babbily asks Are you still there? and ends the call 30 seconds later unless you speak or select I’m here. You can mute your microphone at any time.

Voice Mode is available on paid plans. Calls use your API usage budget, and a call stops if your budget runs out. Have a voice conversation.

A live interpreter in Translate

On the Translate page, select Interpret by voice next to the microphone to hold a spoken conversation across two languages. Each person speaks their own language and hears the other in theirs.

• Choose the languages or let Babbily detect them. Set either side, or both, to Detect language.

• Switch languages mid-call. The call moves to the new languages as soon as you pick them.

• A voice that speaks the language. Before you start, Babbily tells you which language’s voice it will use.

• Nothing is kept. You see what was said and its translation during the call, and it’s gone when you end.

Use the live interpreter.

Visuals in your replies

Babbily can now answer with visuals that appear right in the reply, next to the paragraphs they explain.

• Ready-made visuals for everyday answers. Bar, line, area, pie and scatter charts, key figures, timelines, step-by-step guides, comparisons with pros and cons, checklists you can tick and copy, quizzes, calculators, maps of places and routes, weather and QR codes.

• Custom drawings. When an explanation needs a picture, like how a heat pump works or how a process flows, Babbily draws a diagram or a small interactive explainer.

• On by default for explanations and plans. Visuals is one of the composer’s tools. For text-only replies, switch tools to Manual and leave Visuals off.

• Interactive, and kept apart from your account. Answer a quiz, pick an option to send it as your reply, or move around a map. Each visual runs in its own isolated space, and Babbily asks before a visual opens any link.

• Take it with you. From a visual’s … menu, choose Copy as image, Download PNG, Download SVG, Download HTML or Save as artifact. The options depend on the visual.

• Math looks like math. Equations in replies now appear as formatted math.

Weather, places and stats answers now use the new visuals too. Get visual replies.

Artifacts gets its own page

Select Artifacts in the sidebar to see everything you’ve made in one place. It replaces the artifacts tab in Library.

• Find things fast. Filter by Visuals, Documents, Code or Sheets, search by name, and switch between grid and list views. Visual artifacts show a picture of themselves.

• Make something new. Start a Docs, Design, Dashboard or Sheet artifact in one click. Design builds an interactive app, tool or mockup, and Dashboard turns your data into charts and figures.

• Share an artifact on its own. Select Share to share an artifact with your team or publish it on a public link. Anyone with the link sees the version you published, signed in or not, and you can choose Update to latest version later. Published artifacts show how many views they have.

• Interactive artifacts are visuals now. Asking for an app, tool or interactive mockup creates a visual artifact. Older Generated UI artifacts open as text you can read.

Manage your artifacts.

A memory that understands you better

• Every message uses what Babbily knows. When Memory is on, each message now brings in your most relevant memories.

• Cleaner memories. Babbily saves each fact on its own, in your language, and keeps everyday details like your home city, languages and hobbies. It never saves passwords or keys.

• Forget in plain words. Say “forget that I live in Denver” and Babbily removes just that memory, however it was worded when it was saved.

• Newer facts win. When two memories disagree, Babbily goes with the one that changed most recently.

• Import your memory. In Settings > Memory, use Import memory to bring over what ChatGPT, Claude, Gemini or another assistant remembers about you. Copy the prompt into that assistant, paste its answer, and choose which facts to keep. Sensitive items stay unchecked until you select them.

• See your team’s shared memory. Settings > Team now lists your team’s shared memory and who saved each entry. Authors and the team owner can delete entries.

• No surprise memory notes. Replies mention saving or forgetting only when you ask. If Memory is off and you ask Babbily to remember something, it tells you it wasn’t saved.

Manage memory.

Long chats keep going

• Better summaries. When a chat gets long, Babbily summarizes the earlier part and keeps your own words, so the details you gave stay accurate.

• Long chats stay sendable. Very long chats no longer stop working after they’ve been summarized. If a chat truly can’t continue, Babbily says This chat is too long to continue so you can start a new one.

• Tool-heavy replies stay on track. In replies that use many tools, Babbily sets aside old tool results it no longer needs.

• Switching models keeps your context. Changing models mid-chat no longer forces a new summary.

Learn how long chats work.

A sharper Auto

• Better routing for every message. Auto recognizes more precisely what each message asks for and sends it to a model suited to that kind of task.

• Choose a tool, keep any model. A tool you choose in the composer now runs on every chat model, including Auto.

• Auto on the Free plan. On the Free plan, Auto picks the model for every message. Choosing a specific model, Projects, Scheduled Tasks, video generation, Deep Research and Voice Mode are part of the paid plans.

Learn how Auto works.

Also in this release

• New models. GPT-6.1 Terra joins the picker. Gemini 4 Argon, Muse Wonder, Kimi K3.1 Pro and DeepSeek V4.1 Flash 1017 replace Gemini 3.8 Flash, Muse Spark 1.3, Kimi K3 and DeepSeek V4.1 Flash. Gemini 3.8 Flash and Kimi K3 stay selectable as Sunsetting for now. Compare current models.

• Finance. An earnings calendar, richer company pages with the day’s change on every range, consensus and peers, and answers that use a verified live quote.

• News. Every Discover tab now fills up like Top.

• Places. A new Places tool finds places and shows them on a map.

• Deleting removes files. Deleting a chat, a message or a Library item now also removes the files stored with it.

• Security and reliability improvements across the app.

Babbily can now pick the right model for each message, the whole app has a cleaner design, and two new models join the picker.

Auto picks the model for each message

Choose Auto at the top of the model picker. For every message, Auto judges what kind of task it is and how demanding it is, then picks a model and reasoning level to match. A quick question gets a quick answer, and a hard problem can get a stronger model with more reasoning.

• New chats start on Auto when it is available for your account. Existing chats keep the model they were using.

• See which models answered. Auto · Models used, under the composer, lists every model that replied in the chat.

• Choosing the model is free. It does not count toward your API usage budget, and when your budget runs low, Auto favors models that use less of it.

• Named models are one click away. Pick one when you want the same model every time, or want to set Fast mode or reasoning effort yourself. Those controls are hidden while Auto is selected.

Auto is rolling out to accounts gradually on the web. It is not offered in Translate or Scheduled Tasks.

A cleaner, calmer design

Babbily has a quieter, more consistent look across chat, the sidebar, Settings, Library, Projects, News, Finance, and sign-in.

• A new typeface. Babbily now uses Geist throughout the app, and PDF and Word exports match.

• Stop lives in the composer. While Babbily writes, the send button becomes a stop button. You can also press Esc to stop.

• A tidier thread. Your messages sit in soft rounded bubbles, replies start closer to your question, and Copy, Good response, Bad response, Save to library, and Retry are compact icon buttons that stay out of the way on older replies.

• One header across every page. The title and Share line up the same way everywhere. Thread actions in the sidebar now live in a … menu, or you can right-click a chat. On a touch screen, press and hold.

• Built for every screen. Tablets get an always-visible icon rail, phones get a sliding drawer and bottom-sheet menus, and touch targets are larger.

• More accessible, and in more languages. A clearer focus ring, higher-contrast text, reduced-motion support, and pause buttons for the News and Finance tickers. Much more of Babbily now appears in all 10 app languages.

• A quieter notification offer. It is now a small card with Turn on notifications that you can close.

New models: Claude Sonnet 5.5 and GPT-6.1 Sol

Claude Sonnet 5.5 is built for focused coding tasks, polished documents, and clear everyday work, with reasoning effort through Max. It is also available in Translate.

GPT-6.1 Sol replaces GPT-6 Sol for demanding coding, reasoning, and complex professional workflows. It supports Fast mode and reasoning effort through Max.

A simpler model list

The picker now has 14 named models. Older models, including every model that was marked Sunsetting, have been removed.

• Images and videos come from built-in tools. Image and video models are no longer separate picker choices. GPT Image 2.5 Flare and Grok Imagine Video now power the Image generation and Video generation tools, which you can use from any chat model to create new media, edit an image, animate an image, or edit a video.

• Removed models switch automatically. If a chat used a removed model, Babbily selects Auto or another available model and tells you. Nothing is resent; choose Retry if your last message did not get a reply.

• Translate has a new lineup. Choose Claude Sonnet 5.5, GPT-6 Luna (the new default), or Gemini 3.8 Flash. Translate uses each model’s lightest reasoning setting so results start quickly.

Also in this release

• Reasoning off uses the lightest setting. When reasoning is off, Babbily asks each model for its lightest setting. Some models stop reasoning entirely; others, such as Claude Opus 5.5, still reason briefly at their lowest level. Gemini 3.8 Flash adds Low and High reasoning effort.

• Fewer unexpected sign-outs. A dropped connection or a temporary service problem no longer signs you out. If your account details do not load, Babbily shows Retry and Reload, and you stay signed in.

• Safer team invitations. If you open an invite link while signed out, Babbily asks you to confirm the email address the invitation was sent to. Team invitations no longer expire.

• Plugin approvals are clearer. Approval cards appear right away, typing “yes” in chat does not approve an action, and unanswered cards expire after 24 hours.

• Attach files in News and Finance. Add files to your question from an article or a Finance page.

• Memory understands your language. The memory box in Settings > Memory understands all 10 app languages.

• Leaner error reports. The reports we use to fix problems are now filtered to remove sign-in details, the messages you send, AI prompts and responses, and chat titles. Filtering is not a guarantee, so keep secrets out of chats and titles.

Four new chat models are available in the model picker. Claude Opus 5.5 is suited to sustained coding and knowledge work. GPT-6 Sol handles demanding reasoning and coding, while GPT-6 Luna is a lighter option for focused work. Grok 4.7 supports coding and long-context research.

Choose Fast mode when speed matters. Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna have a Fast mode control. The selected model stays visible, and Fast mode can use more of your API usage budget.

Check older favorites before starting new work. Claude Opus 5, GPT-5.6 Sol, Terra, and Luna, and Grok 4.6 now show Sunsetting in the picker. They remain selectable during the transition.

Use GPT-6 Sol in Translate. It replaces GPT-5.6 Sol in the Translate picker and becomes the default when your previous selection is unavailable. Claude Haiku 4.5 and Gemini 3.8 Flash remain available there.

Attachments no longer depend on the model you picked. Every chat model now accepts documents, text and code, images, audio, video, and ZIP archives. When the selected model cannot read a file itself, Babbily reads the file first and gives the model what it found, so you do not have to switch models before attaching something.

More formats are supported. Excel (.xlsx) and PowerPoint (.pptx) files, RTF, HWPX, HEIC and HEIF photos from an iPhone, and a wider range of audio and video formats now work alongside PDF, Word, text, code, and images.

ZIP archives work as one attachment. Attach a ZIP and Babbily reads the supported files inside it, up to 10 files per archive, and tells you which ones it skipped. Encrypted archives still need their password removed first.

Each file is read once. Follow-up questions, message edits, and retries reuse what Babbily already read, so they are faster and do not spend your API usage budget on the same file twice.

You can see what Babbily is doing. While a file is being read, the activity line shows Reading with the file name instead of a bare spinner.

Clearer answers when a file cannot be used. Babbily names the file and the reason before the message is sent, and points to the fix: save a legacy .doc as .docx, remove a ZIP password, or send a smaller file. An old attachment that can no longer be read no longer blocks the rest of the thread.

Long files fit better. File content now counts toward the context indicator as the text the model actually receives. When several files would not fit, Babbily shows the most recent ones in full and shortens the rest rather than failing your message.

Image, video, and character models are unchanged. Files you attach to them are source media for that model, so they keep their own narrower rules.

Chat replies now stream into the thread as Babbily produces them. Before, a thread filled in steps while it waited for the next update, so the opening words took longer to land.

In our testing across more than 30 responses, replies began arriving about 47% sooner. How much faster your own chats feel depends on the model, the prompt, the tools a reply uses, and your connection.

Interrupted delivery recovers on its own. If the stream stalls or your connection drops, Babbily reconnects to the reply already in progress rather than starting a second one, so you do not get a duplicate answer.

Stop keeps what you already have. Choose Stop and the partial reply stays in the thread, including the text, reasoning, citations, and media produced so far. Stop also works right after you send, before any text appears.

A stopped or stale thread cannot disturb a newer one. Stopping a reply, or a late update from a thread you have moved on from, no longer affects the conversation you are in now.

Replies keep running when you leave. Babbily generates replies on its servers rather than in your browser tab, so closing the tab does not cancel them. That behavior is unchanged and is now documented.

Image and video generation, Scheduled Tasks, and Plugin follow-up steps keep their existing behavior. If a thread still looks unfinished after you reopen it, refresh the page.

Web search now adds pictures only when they fit the question. For the first question in a chat, Babbily can include up to three pictures of a subject you can see, such as a place, product, or landmark. In follow-up questions, Babbily adds search pictures only when you ask to see them, for example “Show me photos of the Eiffel Tower at night.”

Search pictures now appear after the written answer starts instead of before it, and Babbily leaves out logos, icons, ads, tiny thumbnails, and pictures that do not match your question. Deep Research reports do not include search pictures.

GPT-6 Astra is available for complex reasoning, coding, and extended research, with Fast mode and reasoning effort through Max.

GPT Image 2.5 Flare is available for image generation and editing. The chat Image generation tool also uses it for new images from text prompts. GPT Image 2 remains selectable as Sunsetting.

Organize existing threads. Use Add to project, Move to project, or Remove from project on your own eligible threads within their current workspace. You can also create a project from the thread actions.

Review what Babbily remembers. Older items may appear under Needs review, with individual and bulk approve or forget controls. Memory save feedback makes outcomes clearer, and forgetting a memory leaves the original chat history intact.

Find earlier discussions and product help. In eligible chats, Auto mode can look up earlier conversations when memory is on. Babbily can also look up help-center instructions with Babbily Docs.

Long-chat context handling has also been updated. Summaries help keep work moving but can omit details, so restate exact requirements when needed.