Not all AIs think alike. ambientChat connects you to models from multiple leading AI vendors and lets you switch between them at any time. It's like having a panel of experts on call — each one sees the same context about you, so switching doesn't mean starting over, it means getting a different perspective.

#What It Does

The model selector lets you pick which AI answers your next question. Whichever model you choose gets the same context — your inventory, documents, recent conversation, location signals — so you can swap mid-conversation without repeating yourself.

#The vendors

ambientChat conversation routes to models from the three frontier AI labs:

  • OpenAI — a broad, reliable all-rounder family.
  • Anthropic — known for nuanced reasoning and careful long-form analysis.
  • Google — strong multimodal (image + text) capabilities.

Web search and source-grounded answers are handled by Perplexity under the hood — any of the three frontier models can call it as a tool whenever a question needs real-time web information. You don't have to pick Perplexity directly; the model you're talking to will fetch web results on its own when needed, and cite them in the answer.

We keep the specific model versions up to date as vendors ship new ones, so this guide sticks to vendors and tiers rather than naming particular models — those change.

#The four tiers

Models are grouped into four cost tiers. Picking the right tier usually matters more than picking the right vendor.

Tier Cost What it's for
Everyday $ Fast, low-cost models for the majority of daily questions — quick lookups, short summaries, light reasoning. Marked Recommended because most questions don't need more than this.
Thoughtful $$ Mid-cost models for harder reasoning — comparing options, working through a decision, longer-form writing.
Premium $$$ The strongest general-purpose model each vendor offers — deep reasoning, complex analysis, web-grounded research.
Frontier $$$$ A vendor's newest, largest model, sold at a premium above their Premium tier. Worth it for genuinely hard problems — long multi-step work, dense technical material, code and analysis that a Premium model gets nearly right. For everything else it costs several times as much to reach the same answer.

#What $×N means

Next to each model you'll see a badge like $×1, $×3 or $×10. It answers one question: roughly how many times more does this model cost than the cheapest one on your list?

  • $×1 is the cheapest model available to you. Everything else is measured against it.
  • $×3 costs about three times as much per question.
  • $×10 costs about ten times as much.

The tiers above tell you what a model is for; the multiplier tells you what it costs. They don't line up as neatly as you might expect — two models can share a tier and still differ by three times in price, which is exactly why the number is shown.

A few things worth knowing:

  • It's an approximation, deliberately. We round to friendly steps (×1, ×2, ×3, ×5, ×10) rather than showing something like ×2.7. The point is the order of magnitude, not false precision.
  • It's weighted towards what you send, not what you get back. Most of the cost of an answer is the context that goes in — your documents, your inventory, the conversation so far — rather than the words that come back. The multiplier reflects that.
  • It updates itself. The number comes from our current price list, so when a vendor changes their prices the badge changes with them. Nothing is hardcoded.
  • A missing badge means we don't have a published price for that model yet — not that it's free.

Does a bigger number mean a better answer? Not automatically. A Premium model is genuinely better at hard reasoning, but for a quick lookup it costs ten times as much to tell you the same thing. Match the model to the question, not to the price.

#The capability icons

Each model row shows small icons telling you what that model can do:

Icon Meaning
⚡ Quick reply Tuned for speed — answers fast. Shown on Everyday-tier models.
🧠 Thinks aloud Shows its reasoning before answering, useful for hard problems.
👁️ Sees images Can read images you attach or that come from your inventory photos.
📚 Cites sources Returns citations to the sources it used — helpful when facts matter.

Not every icon appears on every model. The icons reflect what each model supports today; as models improve, icons can change.

#All paid access is equal

Every paid subscription can use every model on the list. Subscription tiers control how many tokens you can spend per month, not which models you're allowed to pick. Premium-tier models simply cost more tokens per question, so your monthly allowance goes less far if you use them heavily.

#How to Use It

The model selector is available in three places. Your selection is remembered per-device.

#iOS app

  1. Open Chat.
  2. Tap the current model name near the top of the screen.
  3. A bottom sheet slides up showing tiers (Everyday / Thoughtful / Premium) with each model's capability icons.
  4. Tap the model you want. The next message uses it.

#Web

  1. Open the Chat page.
  2. Click the model button in the top-right (shows the current model).
  3. A dropdown lists the tiers and models with the same capability icons.
  4. Click a model to switch; your choice is remembered on this device.

#MCP (Claude Desktop, OpenClaw, and other AI agents)

When you ask ambientChat via MCP — for example from Claude Desktop — the MCP client is already running its own model (that's how you're talking to it). ambientChat's role in that case is to supply context, not to pick a model. So there's no separate model selector in MCP clients. You choose the model inside your MCP client (e.g., pick a Claude model in Claude Desktop); that client then pulls context from ambientChat via tools.

#Your default model — including your own Mac

Your default model is the one chat starts with. It belongs to your account, so setting it on the phone also sets it on the web, and the other way round.

  • iOS app: Control → AI Model Selection → tap a model under Default model.
  • Web: Control → AI Model Selection → pick from the Default model list.

A model you pick in chat stays picked. Choosing a model in chat's model picker (or swiping to it on the phone) doesn't change your default, but chat keeps your pick: coming back to chat, the app going to the background, a restart or a web reload all leave it selected. It changes only when you pick another model, or when your default changes — then chat opens on the new default. Each device keeps its own pick.

If you've set up Desktop AI (a model running on your own Mac), your Mac appears in both places, above the cloud models, as your Mac's name and the model it runs — for example "Stephen's Mac – Qwen 3.6 35B". You can make it your default. If the Mac is asleep or offline it stays listed and shows why. You can still keep it as your default: chat then tells you it can't reach the Mac. It never quietly switches you to a cloud model instead. If the Mac is connected but its model fails before it starts answering, chat says straight away that your Mac couldn't answer, instead of leaving an empty reply. The same goes for a failure partway through, after it has looked something up for you.

#Choosing which models appear

Under Enabled models (Control → AI Model Selection, on the phone and on the web) every model has a switch, your Mac included. Switch one off to hide it from the chat model picker. The choice belongs to your account, so a model you hide on the phone is hidden on the web too, and AI agents connected to ambientChat can see which models you've switched off. Two things you can't do: switch off your current default (pick a different default first), or switch off every model.

New models are switched on when they arrive. If a model you switched off is replaced by a newer version, the newer one stays off.

#Setting up Desktop AI on your Mac (Alpha)

Desktop AI runs an AI on your own computer. Answers from your Mac use no credits.

It is an alpha. Answers can be slow — often slower than the cloud models — and it may change while we work on it. Only one model is offered, Qwen 3.6 35B.

What you need. A Mac with Apple silicon (M1 or newer), at least 24 GB of memory and about 20 GB of free disk space. The first setup downloads about 17 GB. On a Mac that falls short, ambientChat Desktop AI (and the Terminal command) says why and stops before downloading anything.

Turning it on. Go to Control → Desktop AI — on the phone or the web — and tap or click Try Desktop AI. Show me Anyone can switch it on — there is no invitation or waiting list. To turn it off again, use Turn off Desktop AI at the bottom of the same page: your Macs leave the model pickers, nothing on your Mac is removed, and you can turn it back on any time. You can also ask chat to "turn on Desktop AI".

Setting up your Mac. Setup runs on the Mac itself: the phone, or a browser on a PC, can turn Desktop AI on and send you the link, but can't install it. On the phone, Control → Desktop AI says how and has Email me the setup link, which sends the link to your account email so you can open it on your Mac (a PC's web page offers the same). Then, on the Mac you want to use, open that link — or ambientChat on the web → Control → Desktop AI.

  1. Click Download ambientChat Desktop AI, open the download and drag ambientChat Desktop AI to Applications, then open it. macOS says it was downloaded from the Internet and that Apple checked it — click Open.
  2. Back on the Control page, click Connect this Mac. Your browser asks to open ambientChat Desktop AI — allow it. The app asks you to confirm, then downloads the AI model, Qwen 3.6 (about 17 GB; 15–30 minutes on home broadband) with a progress bar. Your Mac is named after you (for example "Steve's Mac").
  3. When it says Ready, your Mac becomes your default model, on the phone and the web, so your next question goes to it. Chat says so once, on each device: "Your Mac is now your default AI". If you already have a Mac connected, a new one doesn't take over — pick your Mac's name in the model list instead.

Until your Mac connects, Control → Desktop AI on the Mac says Waiting for your Mac… and updates by itself; you can leave it open. If something goes wrong, the page says what:

  • Nothing opened after Connect this Mac. Open ambientChat Desktop AI from Applications yourself, then click Open ambientChat Desktop AI on the page.
  • "Open this page on your Mac to install Desktop AI." You opened the web page in a phone's or a PC's browser. Click Email me the setup link and open the email on your Mac. (In the phone app, Control → Desktop AI has the same button.) If the email can't be sent, the page shows the link to open on your Mac instead; "Couldn't send the link. Please try again." means the request itself failed, so try again.
  • "Couldn't turn on Desktop AI. Please try again." Tap or click Try Desktop AI again.
  • "Couldn't reach ambientChat. Check your connection and try again." This appears on the Mac after Connect this Mac or Get the setup command. Check the Mac's connection and click Try again.
  • "Desktop AI is off for this account." Go to Control → Desktop AI, tap or click Try Desktop AI, then try again.
  • "That code expired before your Mac came online." Each setup code works once and expires in 30 minutes. Click Connect this Mac again (or Get a new command if you're using Terminal).
  • The Mac doesn't meet the requirements. The app or the command says why and stops before downloading anything.

ambientChat Desktop AI lives in the menu bar: it shows whether Desktop AI is ready, answering, downloading or offline, and starts by itself when you log in. Quit ambientChat Desktop AI stops Desktop AI completely — nothing keeps running in the background — until you open the app again or next log in. The version is at the foot of the menu; Advanced… shows the app and Desktop AI versions and has Copy diagnostics for support (versions and states only, never what you asked), and Advanced… → Uninstall removes it and every model it downloaded. To install a newer copy, just open it: it offers to replace the running version and keeps your connection and models. Nothing you ask is stored on the Mac.

Watching it answer. While Desktop AI works on a question, a short word appears beside its menu-bar icon (Gathering, Reading, Searching, Writing) and the icon breathes slowly; with Reduce Motion on it keeps still. Open the menu to follow each step, live, without reopening it. The menu uses the same words as chat on your phone and the web. ambientChat first finds and gathers the context your question needs ("Finding relevant contextual information…", then "Gathering your context: 1,387 items, 3 documents"), with the detail underneath ("about 9,200 tokens"). Then your Mac reads it ("Your Mac is reading your context · 61%" — after a lookup, "Your Desktop AI is reading what it found"; on a follow-up it hasn't kept, "Your Mac is catching up on this conversation"), saying how much was already in its memory from the last question (warm cache), which makes follow-ups faster. Any lookup ambientChat runs for it shows up as it happens, worded as chat words it ("Searching your inventory…"). If a step needs your permission, the menu says Waiting for your OK… until you choose Allow in ambientChat. Last, your Mac writes the answer ("Writing the answer…"), with its speed underneath ("34 tokens/s"). The menu then says how long the answer took and that it was saved to your history. Afterwards the line under Ready keeps the last answer's time and how much of its context came from warm cache. A question after the model has been asleep shows Starting the AI on this Mac — up to a minute if it's been idle first. Chat on your phone and on the web follows the same steps, in one line under the answer that changes in place instead of a spinner: Gathering your context: inventory, 3 documents, Study, then Your Mac is reading your context with a thin bar that fills as it reads — Your Mac is catching up on this conversation on a follow-up — and, on a long read, roughly how long is left (— about 30 seconds). A lookup says what it is looking for, and the read after it says Your Desktop AI is reading what it found. Then Writing the answer… until the first word. After a rest it says Starting the AI on Stephen's Mac — up to a minute if it's been idle. A step only appears if it lasts more than a moment, so a quick answer goes straight from the dots to the words, and nothing is slowed down to show it. Cloud models show the same gathering and lookup steps, without the reading bar. A Mac on an older Desktop AI shows the steps without the bar until it updates. The Mac starts waking as soon as ambientChat begins gathering your context, so the wait is a little shorter than before. The menu shows counts and steps only, never what you asked.

Keep the model in memory. A switch in Advanced…, on by default. On, Desktop AI holds the model in your Mac's memory (about 15 GB) all the time, so answers start straight away. Off, it frees that memory after 15 minutes without a question — better on a Mac you also use for other work — and the next answer then takes up to a minute longer to start while the model loads; the menu says Sleeping to save memory meanwhile. Flipping it takes effect as soon as Desktop AI has finished any answer in progress, never in the middle of one. The choice belongs to that Mac: it's kept when you connect the Mac again or click Set up again, and it can only be changed on the Mac. If Advanced says Desktop AI needs an update first, update it from the Desktop AI menu (Check for Updates…, or Update ready — Restart to finish once one has downloaded). When it's off, the model pickers and Control → Desktop AI on your phone and the web show that Mac as Online · Frees its memory when idle, so a first answer can take a minute.

The development app and the main app on one Mac. Each has its own Mac app: ambientChat Desktop AI for the main ambientChat app and ambientChat Desktop AI Dev for the development app, each with its own connection and its own copy of the model. Only one can answer at a time, because one model already fills most of a Mac's memory. While ambientChat Desktop AI is running, ambientChat Desktop AI Dev waits: opening it says so, and your Mac shows as offline in the development app until you quit ambientChat Desktop AI from its menu-bar icon.

"Setup stopped — macOS wouldn't start Desktop AI in the background." macOS refused to run the part of Desktop AI that answers your questions, so the Mac shows as offline even though the menu-bar app is open. The app tries once more by itself first. If it still can't, the menu offers Try starting it again and Open Login Items…. Try again first. If it stops again, open Login Items and check ambientChat Desktop AI (or ambientChat Desktop AI Dev) is switched on under Allow in the Background, then try again.

Installed it when it was called ambientChat Home? Choose Quit ambientChat Home in its menu, drag ambientChat Home from Applications to the Trash (don't use Advanced… → Uninstall, which deletes your models), then download ambientChat Desktop AI from Control → Desktop AI and open it. Your Mac stays connected and keeps its downloaded models — nothing is set up or downloaded again. If macOS asks, allow it in Login Items so Desktop AI keeps running in the background.

Updates. Desktop AI keeps itself up to date. Fixes to the part that answers your questions arrive on their own and take effect between answers, never in the middle of one; if a new version fails to start or to connect, your Mac goes back to the version that worked. When a new version of the ambientChat Desktop AI app is ready, a small red dot appears on its menu-bar icon and the menu says Update ready — Restart to finish — click it. Check for Updates… checks any time: a small window shows the check and then either the new version or You're up to date, and the bottom of the menu keeps the answer (for example Up to date · checked 6:41 AM). If that line says Couldn't check for updates, the Mac was probably offline — Desktop AI tries again by itself every day. New models are not updates: if a better one is offered, the menu says A better model is available, and nothing is downloaded until you choose it. A Mac set up before October 2026 needs one last manual update: download ambientChat Desktop AI again (or run the Terminal command again).

Which model runs on your Mac. In the model picker, in Control → AI Model Selection and, on the web, on the model button at the top right of chat, your Mac's model is named beside it, for example "Stephen's Mac – Qwen 3.6 35B", so you can always see which model you picked. Under Control → Desktop AI → Your Macs, on the phone and the web, each Mac lists the model it runs and how much memory it has, then whether it is online — for example "Qwen 3.6 35B · 64 GB memory". A Mac with an older ambientChat Desktop AI shows only the model until it updates. While a Mac downloads a new model, that line names the model still answering and the line below says what is downloading. This is for reading only. Desktop AI runs Qwen 3.6 35B — in our tests it handled ambientChat's tasks nearly as well as the cloud models. ambientChat picks the model for your Mac's memory: a Mac with 48 GB of memory or more gets Qwen 3.6 35B (64k), the same model with space for conversations twice as long; a smaller Mac gets Qwen 3.6 35B. A new Mac gets the model for its memory the moment it connects, whether you use the Desktop AI app or the Terminal command. A Mac set up in place of one with more memory gets a model its own memory can hold. Set up again keeps the model your Mac runs — reinstalling, or starting over after a setup code expired, never moves a Mac back to the standard model — unless that model no longer fits the Mac's memory. Macs set up before October 2026 may run Gemma 4 26B, and keep it when set up again. When you click Set up again (or connect a Mac from the Desktop AI app), the web says which model your Mac will get and why — for example "Kept Qwen 3.6 35B (64k), the model this Mac already runs." — and so does chat. When a Mac's model does change, it downloads the new model while the old one keeps answering. Until the download finishes, Control → Desktop AI and the model pickers say so — for example "Downloading Qwen 3.6 35B · Gemma 4 26B answers meanwhile — 21% · about 20 min left" (the percentage and time left appear once your Mac has the latest ambientChat Home; Control → Desktop AI updates them about every 30 seconds after you click Set up again, and the pickers show the latest each time you open them) — and Control says the Mac is ready, naming the new model, only once the new model is the one answering. The old model stays on your Mac — Desktop AI never deletes a model by itself — so going back to it later doesn't mean downloading it again. Each model takes 16–18 GB of disk.

Facts about the world are checked on the web. A model small enough to run on a Mac can misremember people, history and the news. So when you ask Desktop AI about the world — "What about Mark Carney?", "When did the Berlin Wall fall?" — ambientChat runs a web search while it gathers your context, and your Mac answers from the results, citing their sources. If the search is still going once your context is gathered, the line under the answer says Searching the web… (unless your Mac is still waking its model, which takes longer). Questions about your own things (your inventory, documents and places) are not searched. The search sends your question, and your previous question when it is a follow-up, to the web-search service. If you have blocked web_search in your tool permissions, it never runs. And if any model, on your Mac or in the cloud, says "Let me search…" without searching, ambientChat runs that search and the answer continues from the results, starting Checked against a web search:.

When an answer was not checked, it says so. If Desktop AI states a fact about the world and no web result reached it (you have blocked web search, or no search ran or it failed), the line under the answer says Answered from memory, not checked on the web. Treat that answer as a recollection: names, dates, prices and addresses in it may be out of date or wrong. For one you can rely on, pick a cloud model in the model picker and ask again, or allow the web search. The same line appears wherever that answer goes: in Chat and Q&A on the phone and the web, on a shared link and in its email, in a printed copy, and to an AI app reading your conversations over MCP. Answers about your own things, greetings, rewriting and opinions never carry it.

Removing a Mac. An old Mac, or a duplicate left behind when you reinstalled, can be removed for good: tap or click Remove under it in Control → AI Model Selection (on the phone and the web), or next to it under Control → Desktop AI → Your Macs on the web. You can also ask chat to "remove my old Mac". It disappears from your models on every device, and that Mac can no longer connect as your Desktop AI. If it was your default, chat goes back to the standard default until you pick another. Removing can't be undone — to use that Mac again, set it up again. (Switching a Mac off under Enabled models only hides it; removing deletes it.)

#Sharing your Mac with family and friends (Alpha)

Your Mac can answer for up to 5 people besides you: family, friends, anyone in one of your groups. Answers from it still cost nothing.

Who can share. Sharing is rolling out, so for now only some accounts see Share Access. Anyone can use a Mac that's shared with them, whatever their plan, as long as its owner can share.

Sharing it. In Control → Desktop AI → Your Macs, tap or click Share Access next to the Mac. It opens the same share sheet as a document or a folder: Person to share with someone by email, or Group to pick one of your groups, make a new one, or add people to it. Everyone it reaches can use the Mac straight away, up to 5 seats across everyone you share it with (one person in two groups takes one seat). Sharing that would make more than 5 is refused, with the reason, before anything is shared. Someone without an account yet gets an email, and takes their seat when they join. A Mac stays shared until you stop sharing it. You can also ask chat to "share my Mac with Family" or "share my Mac with ana@example.com".

Who can use it. The same sheet lists everyone: Can use your Mac, or Waiting for your OK for someone another member added to the group after you shared it. They get no seat until you tap Let them use it, and ambientChat asks you when they join ("Alex joined Family. Let Alex use your Mac?"). Anyone you add, by email or into a group the Mac is shared with, gets a seat at once while one is free; if the Mac is full, ambientChat tells you. Remove frees someone's seat without taking them out of the group. Stop sharing ends a group's access at once. Someone who leaves the group loses the Mac on their next question.

What you see, and what you don't. For each person you see how often they used your Mac this month, by kind (questions, web checks, images, expert opinions), and when. You never see what they asked or what your Mac answered, and nothing is stored on the Mac. While someone else's question is being answered, the menu bar says Answering for Sam and nothing about their things: no counts, no zones, no lookups.

Using a Mac someone shared with you. It appears in your model pickers as their Mac ("Ana's Mac – Qwen 3.6 35B · Shared by Ana") and under Control → Desktop AI → Shared with you, with which group it came through and whether it's online. Until Ana says yes it shows Waiting for Ana's OK. The first time you use it (when you pick it, open a chat on it, or make it your default), ambientChat tells you: "Your questions are answered on Ana's Mac. We never store them there, but only use a Mac whose owner you trust." You see this once, on whichever device you use first. On the free plan you're offered, once: Use Ana's Mac as your default? A shared Mac is yours to use, never to set up, remove or share.

Who pays for what. Questions the Mac answers are free. Anything else a guest's question uses (a web check, an image, an expert's opinion) comes out of the guest's own allowance, never the Mac owner's. When a free guest's allowance runs out, the Mac still answers, but without checking the web. The answer says Answered from memory, not checked on the web, links to an upgrade, and says when web checks come back.

Prefer Terminal? Advanced: install with Terminal gives you a one-line command that sets up the same thing. If there's no Download button yet, that command is what the page offers.

Why "Desktop AI" and not "Private"? Your questions are answered on your Mac, but a few things around them still use cloud services today — the chat's title, for example. "Private" is kept for a stricter mode that will keep everything away from outside AI services.

#Switching mid-conversation

Switching models doesn't clear your conversation. All models read the same recent history, so you can start a question with a fast Everyday model, switch to a Thoughtful or Premium model to go deeper, and the new model picks up where the last one left off.

#Tips & Tricks

  • Stay on Everyday until you need more. Most questions — "is anything expiring this week?", "which beacon am I near?", "summarise that receipt" — don't need a Premium model. Save the expensive ones for the hard ones.
  • Match the model to the job. Use a 🧠 Thinks aloud model when you want it to reason through a decision. Use a 👁️ Sees images model when you're attaching photos or asking about photos on your items. Use a 📚 Cites sources model when accuracy of facts matters.
  • Compare answers. If a model's response feels off, switch vendors and ask again — different vendors genuinely reason differently. Same context, different framing.
  • Watch your token usage, not your model list. The usage indicator shows how much of your monthly budget you've spent. If it's climbing fast, drop to a cheaper tier for the routine questions.

#Options

Setting What it does Default
Selected model The model used for your next message on this device. Changed via the model selector. Anthropic, Everyday tier
Persistence Your choice is remembered per-device (web: browser local storage; iOS: device storage). —
Default model The model chat starts with. Saved to your account, so it's the same on the phone and the web. Can be your own Mac (Desktop AI). Anthropic, Everyday tier
Enabled models Which models, your Mac included, appear in the chat model picker. Saved to your account, so it's the same on the phone and the web. All on

#Known Limitations

  • Model catalogue is curated. We only expose models that pass our quality and cost checks. When vendors ship new models, they're added after vetting; older ones are retired. The list you see is always the current set.
  • Not every model supports every capability. A model without the 👁️ icon can't interpret images; a model without 📚 can't return citations. The icons tell you before you pick.
  • Context windows vary. Each model has a limit on how much history + context it can consider at once. Very long conversations may drop the oldest turns on models with smaller windows. Starting a fresh conversation resets this.
  • A second opinion is one model at a time. Switch models mid-conversation and ask, or ask a different model on its own for an independent answer — see Second Opinions. Nothing sends your question to several models at once.
  • Answers will differ between models. This is the point, not a bug. If two models disagree, that's signal worth noticing.

#Version History

Version Date What Changed
61 2026-11-12 A Desktop AI answer that states a fact about the world with no web result says "Answered from memory, not checked on the web" under it, on the phone, the web, shared links, email and print; a guest's "No web check" line now reads the same AMB-3255
60 2026-11-03 A Mac whose owner's account is turned off now says "Account paused — Desktop AI is off" in its menu bar, and one whose account was deleted says "Account deleted — Desktop AI stopped", instead of "Offline — check internet" AMB-2812
59 2026-10-29 A Mac running Desktop AI stops answering within a minute once its owner's account is disabled or deleted, and comes back by itself, with no new pairing, when the account is enabled again AMB-2809
58 2026-10-27 Setting up Desktop AI now says what to do when it goes wrong: nothing opens after Connect this Mac, the page is open on a phone or PC, turning it on or reaching ambientChat fails, the setup code expires, or the Mac falls short. It also says plainly that the phone and a PC can send the link but cannot install it. On the web, your Mac's model shows on the model button at the top right of chat AMB-2584
57 2026-10-27 A Show me link at the first step of Desktop AI set-up opens the right screen in the app (on the public help site it opens the same screen on the web) AMB-2584
56 2026-10-21 A model you pick in chat stays picked across a return to chat, a restart and a web reload, until you pick another or your default changes (AMB-2290)
55 2026-10-19 Desktop AI now waits up to three minutes for your Mac on a long answer instead of one, and if it still runs out of time it says your Mac is still working on it, not that the connection is down AMB-2165
54 2026-10-19 Desktop AI comes to the main ambientChat app for accounts in the alpha: its own Mac app, ambientChat Desktop AI, downloaded from Control → Desktop AI. The development app's Mac app is now ambientChat Desktop AI Dev, and only one of them answers at a time on a Mac (AMB-2134)
53 2026-10-15 Google's Flash model is now Gemini 3.8 Flash, replacing Gemini 3.7 Flash, which Google has retired. It is the same speed and does the same things, including reading scanned PDFs. If you had picked Gemini 3.7 Flash, or an older Flash, you move to 3.8 automatically. Until 31 December 2026 it costs half what 3.7 Flash did; from 1 January 2027 Google doubles its price, so it goes back to costing what 3.7 Flash costs today (AMB-2076)
52 2026-10-14 While Desktop AI answers, its menu bar shows each step live: ambientChat gathering your context, your Mac reading it (with how much came from warm cache) and writing the answer, any lookups, and the answer saved; a short word and a slow pulse beside the icon show it working. When your Mac's model was asleep, chat on the phone and the web says your Mac is waking it, and it starts waking sooner (AMB-2067)
51 2026-10-14 OpenAI's Premium model is now GPT-6.1 Sol, replacing GPT-6 Sol at the same price with cheaper repeated context; GPT-5.6 Terra is retired. If you had picked GPT-6 Sol or GPT-5.6 Terra, you move to GPT-6.1 Sol automatically (AMB-2066)
50 2026-10-13 Google's Premium model is now Gemini 4 Argon, replacing Gemini 3.1 Pro at a lower output price; if you had picked Gemini 3.1 Pro, you move to Gemini 4 Argon automatically, and past answers keep the name of the model that wrote them (AMB-2052)
49 2026-10-09 The Everyday Anthropic model is now Claude Haiku 5.5, and it is the default for everyone: about as accurate as Haiku 4.5 in our tests, faster on slow answers, and far cheaper per question. If you had picked Haiku 4.5, you move to 5.5 automatically — and if you had switched Haiku 4.5 off, Haiku 5.5 stays off until you switch it back on in AI settings (AMB-1922)
48 2026-10-08 OpenAI's Premium and Everyday models are now GPT-6 Sol and GPT-6 Luna, replacing GPT-5.6 Sol and Luna at lower prices; if you had picked a GPT-5.6 model, you move to its GPT-6 version automatically (AMB-1921)
47 2026-10-08 Settings → AI Model Selection, and on the web the model button above the chat box, name the model running on your Mac beside its name too, e.g. "Stephen's Mac – Qwen 3.6 35B"; the line under it says Desktop AI (AMB-1918)
46 2026-10-08 The Mac app is now called ambientChat Desktop AI. If you have the old ambientChat Home, quit it, drag it to the Trash and download the new one: your Mac stays connected and keeps its models (AMB-1900)
45 2026-10-08 If your Mac's Desktop AI model fails after looking something up for you, the answer now says your Mac couldn't answer instead of a generic "something went wrong" AMB-1873
44 2026-10-07 Home AI is now called Desktop AI: in the model list ("your Mac's name · Desktop AI"), in Settings → Desktop AI, and on answers your Mac gave. Nothing else changes, and the Mac app is still called ambientChat Desktop AI (AMB-1899)
43 2026-10-07 Each Mac in the model picker now shows $0 on its own row, where cloud models show their $×N, on iOS and the web AMB-1898
42 2026-10-07 The model picker now shows $0 beside Desktop AI, where the cloud models show $ and $$: questions your own Mac answers use no credits, on iOS and the web AMB-1888
41 2026-10-07 Old or duplicate Macs can be removed from Desktop AI for good — Remove in Settings on the web and the phone, or ask chat (AMB-1887)
40 2026-10-07 Switching your Mac to another Desktop AI model keeps the old one on the Mac instead of deleting it, so switching back needs no new download AMB-1868
39 2026-10-07 Desktop AI on your Mac updates itself: a fix reaches the Mac without reinstalling, it restarts only between answers, and if a new version fails to start or connect it goes back to the one that worked. The ambientChat Desktop AI menu is easier to read, Quit stops everything, opening a newer copy replaces the running one, and the version is shown in the menu AMB-1796
38 2026-10-07 When your Mac's Desktop AI model fails before it starts answering, chat now says your Mac couldn't answer straight away, instead of showing an empty reply about a minute later AMB-1712
37 2026-11-05 Desktop AI picks the model for your Mac's memory: 48 GB or more gets Qwen 3.6 35B (64k), a smaller Mac Qwen 3.6 35B — from the moment a Mac connects — and Set up again keeps the model your Mac runs instead of moving it back to the standard one (AMB-2882, AMB-2967)
35 2026-10-25 Web search answers come from the model you picked. For GPT, Gemini and Desktop AI, a web search now hands your model ranked results to read and cite as [1], [2]; no second AI writes an answer underneath, and sources still appear under the answer. A web search for a narration, and a barcode the catalogs don't know, now use your credits like any other web search (AMB-2509)
33 2026-10-23 While sharing rolls out, only some accounts can share a Mac, but anyone they share it with can use it (AMB-2375)
32 2026-10-23 Share your Mac with up to 5 people in your groups: you approve each person and see how often they use it, never what they ask; guests find it under Shared with you, read a one-time notice before their first question, and pay for their own web checks (AMB-2375)
31 2026-10-21 The Desktop AI menu bar now says each step in the same words as chat on your phone and the web ("Finding relevant contextual information…", "Your Mac is reading your context", "Writing the answer…") (AMB-2314)
30 2026-10-21 Desktop AI checks facts about people, history and the news with a web search before it answers; an answer that says "Let me search…" without searching now gets that search (AMB-2289)
29 2026-10-21 Chat's progress line on the phone and the web, in everyday words: a long read on your Mac says roughly how long is left, a follow-up says it's catching up on the conversation, and a lookup says what it's looking for (AMB-2299)
28 2026-10-21 The model you pick in chat stays picked — returning to chat or restarting the app no longer puts your default back over it; a new default still takes over (AMB-2290)
27 2026-10-20 Desktop AI is open to every account in the main app too: Try Desktop AI in Control → Desktop AI, no invitation (AMB-2234)
26 2026-10-19 Chat on the phone and the web shows what an answer is doing in one line — gathering your context (naming what it used), your Mac reading it with a real progress bar, lookups, writing — instead of a bare spinner (AMB-2171)
25 2026-10-19 In the chat model picker the Alpha tag sits once, after "Answered by a model on your own Mac.", instead of beside each Mac (AMB-2160)
24 2026-10-18 Email me the setup link on the phone (and on a PC's web page) sends the Desktop AI setup link to your account email (AMB-2133)
23 2026-10-18 Try Desktop AI: any account can switch the Desktop AI alpha on (and off) in Control → Desktop AI, on the phone and the web — development app first. The phone gets a Control → Desktop AI page; the hardware needed is stated (24 GB memory, about 20 GB free disk) (AMB-2133)
22 2026-10-18 Desktop AI wears a small Alpha tag wherever it appears: Control, and beside each Mac in the model pickers on the phone and the web (AMB-2133)
21 2026-10-16 The Settings tab is now called Control, on the phone and the web; the web address is /control (old /settings links still work) (AMB-2120)
20 2026-10-13 When Desktop AI is too old to offer Keep the model in memory, Advanced… says it needs an update instead of showing nothing (AMB-2047)
19 2026-10-13 Keep the model in memory in ambientChat Desktop AI's Advanced…: off frees about 15 GB after 15 idle minutes, at the cost of a slower next answer; the choice stays with the Mac (AMB-1959)
18 2026-10-09 The Panel of Experts toggle is gone; second opinions are switch-and-ask or an independent expert, see Second Opinions (AMB-1930)
17 2026-10-08 The chat model picker names the model running on your Mac beside its name, e.g. "Stephen's Mac – Qwen 3.6 35B"
16 2026-10-07 While a new model downloads, Settings and the pickers say how far it has got and roughly how long is left (AMB-1896)
15 2026-10-07 After Set up again, Desktop AI no longer says "ready" while the new model is still downloading: Settings and the pickers say which model is downloading and which one answers meanwhile (AMB-1885)
14 2026-10-06 Qwen 3.6 is the Desktop AI model for new Macs; Set up again moves a Mac to it, and Gemma 4 stays available (AMB-1829)
13 2026-10-06 Private AI is now called Home AI; Settings → Home AI leads with one Download ambientChat Home button, says what macOS asks on first open, no longer asks you to name the Mac, and keeps the Terminal command under Advanced (AMB-1782)
12 2026-10-06 Chat tells you when your Mac became your default, once on each device, so you don't just find it selected (AMB-1792)
11 2026-10-05 Your first Mac becomes your default the moment it's ready, so your next question isn't answered by a cloud model because you forgot to switch. It happens once: pick a cloud model afterwards and it stays your choice (AMB-1786)
10 2026-10-05 Enabled models: the on/off switches are now on the web too and belong to your account, so phone and web always match (AMB-1772)
9 2026-10-03 Added "Setting up Private AI on your Mac": the ambientChat Home app — Download, then Connect this Mac (AMB-1764)
8 2026-10-03 Added "Your default model — including your own Mac": the default is now settable on the web, and Private AI appears in AI Model Selection on both surfaces (AMB-1666)
6 2026-07-23 Corrected Panel of Experts limitation wording — sequential switch-and-re-ask, not "ask several models at once"; see Panel of Experts v3
5 2026-06-08 Corrected stale limitation: Panel of Experts is live (was "not live yet")
4 2026-06-07 Terminology: web model-picker section retitled "Web" (the web interface)
3 2026-04-22 Demoted Perplexity from the primary model selector. Conversation is now handled by the three frontier labs (OpenAI, Anthropic, Google); Perplexity is used under the hood for web-grounded answers and cited sources, invoked automatically as a tool by whichever primary model you pick. Cleaner product story, consistent capabilities across all primaries.
2 2026-04-19 Rewrote to match current behavior: three-tier structure (Everyday / Thoughtful / Premium), capability icon legend, clarified that all paid subscriptions can use all models (token budget is the limit, not access), added web + MCP platforms, removed references to specific obsolete model names.
1 2026-03-01 Initial guide