• Pictures no longer fail to save when they finish while your device is locked. An image from OpenAI, Grok, Gemini or on-device generation that arrived during that time was lost with a "You don't have permission to save the file" error. It now saves.

• Retry Failed Prompts, in the conversation's cog menu. When a batch loses a lot of prompts to one problem, they can now all run again at once with the model you have selected, in a queue of their own with progress, pause and stop. Cancelled prompts are offered as a separate choice, including whatever a stopped batch hadn't reached yet.

• Pins and archives now reach your other devices while Ari is open. A chat pinned or archived on one device used to show up on the others only after Ari was relaunched, and not always then. It now arrives straight away when your devices share a network, and shortly after you open Ari otherwise. An older copy of a chat can no longer undo a pin or archive you have just made.

• Every picture Ari generates carries its invisible mark now, whatever made it. Only pictures from Draw Things were being marked — the on-device models, OpenAI, Gemini, Grok, Shortcuts and the agentic image tool all missed it. Every one of those marks now. • Pictures you already have get the mark on their way out. Exporting, sharing or dragging a generated picture out of Ari marks that copy as it leaves. What is stored is never rewritten: the conversation's copy stays the record of what the model actually produced. • The mark is invisible where it would otherwise show. Ari writes only into the parts of a picture that have enough detail to hide a mark, and leaves flat colour untouched, byte for byte. An earlier build spread it evenly and left visible blotching across areas of solid colour; that is gone. • The mark survives cropping and re-compression, and now proves itself. It carries a checksum, so a reading either verifies or is discarded — it can no longer report a wrong identifier. Checking a picture tells you whether it has been cropped since it was made, and still recognises marks written by earlier versions of Ari. Resizing a picture removes it. • The slideshow no longer strands you with no controls. Reading the sensitive-content warning for a while and then tapping Reveal All used to leave a slideshow with no Share and no Close, because the controls' idle countdown had been running behind the shield the whole time. Revealing now restarts it, the way every other slideshow control already did.

• Pin and archive conversations. Pin a chat and it sits right below the one you have open; archive one and it moves into a collapsed Archived section at the bottom of the list. Press and hold a conversation to pin, unpin, archive or restore it. Regular and private conversations are organized separately, and your pins sync between your devices. • Batches pick up where they left off. If Ari closes in the middle of a templated or repeated prompt, the conversation now offers to resume from the first prompt that had not finished, instead of the run vanishing. Prompts you had stacked while the model was answering come back too. • Long prompts are read in full. The on-device Stable Diffusion models used to stop listening after about 75 tokens (roughly 55 words), so the second half of a paragraph was silently ignored. Ari now reads the whole prompt and the whole negative prompt. The downloaded Core ML models read long prompts too, a little slower per step. • A Negative tab beside Prompt in the composer, for the models that use one. It is remembered for each conversation, shown under every message that used it (tap to expand), and restored when you retry. • Auto-export pictures. Choose a folder for a conversation and every picture is also saved there as it is made, optionally into a subfolder. Folders that are missing, even ones you delete later, are created again. Private conversations export as ordinary files, and the screen says so. • The model can look at an image file on your device with a new view_image tool, instead of having to open it in the browser first. • Templated prompts confirm as an interactive tree: switch any option off and the permutation count updates live, and only what is left runs. • The slideshow keeps your place when new pictures arrive while it is open. Its controls no longer pop back on the next slide, show only when you tap the middle of the picture, and hide after 10 seconds.

• Every private conversation is listed again. Private conversations are the last thing Ari asks iCloud for, and it used to stop waiting while they were still on their way — so a device that had only just synced showed one of them, or none, instead of all. Ari now waits for the whole set at once, the way iCloud actually delivers it.

• Remote terminals. Turn on "Mirror terminals to my other devices" on your Mac and its Code Session terminals appear here over Remote Dispatch, on your LAN first and through iCloud elsewhere. Watch a build, a dev server or an agent run from your iPhone or iPad, and type into it from there. • A real terminal on the Mac. Code Session terminal tabs now run your actual login shell in a pseudo-terminal with a full xterm-256color emulator, so vim, htop, REPLs and other full-screen programs work, and Homebrew and other host tools are found the way your Terminal app finds them. • The model can drive that terminal with terminal_open, terminal_send, terminal_key, terminal_read and terminal_list, for dev servers, REPLs and anything interactive. • Browser tools grew up: the model can read a page as an accessibility tree with reference ids, find elements, fill forms, hover, press keys, inspect network requests and responses, manage tabs, batch actions and zoom into a region.

• Mi Speaks voices are back. With Mi Speaks installed, Ari's voice picker lists its Studio voices again and reads chat replies and film narration with them. Studio voices are part of Mi Speaks Premium, an in-app purchase inside the Mi Speaks app. • Describe a look, and Luma applies it. The new Shortcuts action "Describe a Look for Luma" asks your current model to turn words like "moody film, cool shadows, a little grain" into Luma's own sliders, then opens Luma's editor with every slider live, ready to finish by hand or keep as a preset. Ari writes the sliders from your words; it does not look at the photo. Needs Luma 2.0.4 or later. • "Write a Luma Recipe" returns the same recipe as text for your own Shortcuts, and the link ari://luma-look?look=… does it from any app that can open a link. • Open a folder in a Code Session from another app. The Shortcuts action "Open Folder in Code Session" (or the link ari://code-session?folder=…) starts a new Code Session with that folder as its workspace, named after the folder. Enter Space's file browser offers it as "Open in Ari Code Session". A folder on an external drive or the Enter Space volume is allowed once when asked, and a folder Ari cannot read is refused with the reason instead of a session that fails on its first file.

New: • The cloud model lists match each provider's current lineup: GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna; Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5; Grok 4.7; Gemini 4 Argon and Nano Banana 2.1. Auto-select picks the current flagships; your own choice is left alone. • Gemini video is made by Gemini Omni Flash, which replaces the retired Veo 3.1 previews. A saved Veo choice moves over on its own. • AWS Bedrock lists the models your account can actually use, straight from Bedrock, and reaches the current Claude models (Opus 5.5, Sonnet 5.5, Haiku 5.5 and more) through Bedrock's Messages endpoint. • OpenAI Sora is no longer offered; OpenAI shut the Sora video service down in September. Films saved with Sora keep their place and follow your current model choice. Fixed: • GPT-6 models no longer reject requests over a temperature setting, and photos attached to a GPT-6 conversation are sent again. • GPT-5.5 Pro uses the right endpoint, and Code Session cost estimates match current prices.

Fixed: • Conversations kept in iCloud no longer go missing on iPhone while their files are still downloading; Ari now waits for each one. • Photos you attach are no longer sent to AI services larger than the original on high-resolution screens, and small pictures are no longer enlarged. • Shortcuts now read Latin-1 text files correctly instead of showing unrelated characters. • Images Ari generates keep the name of the model that made them, so the gallery can show it.

Fixed: • The Lock Screen "Look at Something" button now opens Ari's camera, so you can point it at something and ask. It used to only vibrate. • Settings › Voice no longer closes the whole Settings sheet a moment after it opens on iPhone.

Optimized for iPhone Duo. Open or folded, Ari fits the screen: chat, the photo and film studios, the live camera and the sidebar use the wide display when it is unfolded. Fixed: • Attached text files saved as Latin-1 now read correctly. • Grok and OpenAI-compatible replies no longer break at line endings, and Gemini replies keep accented and non-Latin characters. • A prompt template saved in Settings now shapes the model's replies, and short answers like "No" or "OK" are no longer replaced. • Asking for a picture while an image-only model is selected makes the picture instead of answering in words. • Removing a custom endpoint also clears the model selection that pointed at it.

Make pictures on this device. Settings › Voice & Media › On-Device Pictures downloads Stable Diffusion 2.1 or Stable Diffusion XL, converted for Core ML. Ask for a picture while Apple Intelligence or a downloaded model is selected, and it is made here — nothing leaves the device. XL appears where the device has the memory for it. Import Model… brings in a model you already have: an MLX language model folder, or Core ML Stable Diffusion resources. Ari checks the folder before copying anything and says what is missing or why it can't be used. New: import an SDXL checkpoint (.safetensors), including Pony, Illustrious and NoobAI models, and make pictures with it on devices with enough memory — or have another of your devices make them. Watch each picture form, with its steps and the time left. Browse Models offers Stable Diffusion models to download, each with its licence. "What each tool does" no longer freezes the app. Its film example ran every frame on the main thread; it works in the background now, and the pictures scroll with the rest of the page.

Repair damage no longer stops at a crossing. Where two scratches cross, the crossing is thicker than either of them — and Ari finds damage by how thin it is, so it saw picture rather than damage and left a speck behind in the middle of a line it had otherwise taken out. A mark is now followed a little way past the last point Ari is sure of: a spot that only half answers the test counts as damage where it continues damage, and not at all on its own. That carries a scratch through its crossings, and through the stretches where it runs over something as bright as itself. On a scratched photograph it now finds 95% of the damage, against 92% before.

Ari 7: it makes pictures and films now, and repairs the ones you already have. FILMS is a studio rather than a form. Write a film as a list of shots and render them one at a time, keeping the ones that work and redoing the rest. A player shows the cut as it currently stands, a timeline shows every shot with its own poster frame and real length, and a sidebar edits whichever shot you pick. Re-render shot four and the player updates without exporting anything first. A rendering film says how much longer. Each shot draws its own progress and time remaining on its own card in the timeline, and a bar across the bottom does the same for the whole run — counted from how long the shots that already finished actually took, rather than from a guess that never improves. PICTURES have an editor. Brush over a crease and heal it, paint a leash or a stranger out of the frame, or select nothing and treat the whole photograph: repair, denoise, sharpen camera shake, colorize, upscale four times, or describe what should be there and have it repainted. Every step is kept, so going back three edits is one click, and each one blends through the brush's own soft edge rather than landing as a pasted rectangle. LOOK AT SOMETHING — point the camera at whatever you are wondering about and ask. It goes through whichever model you have selected, on-device, self-hosted or cloud, with the answer captioned under the frame and spoken if you want it. Talk to it and it listens, answers, then listens again — and waits for you to finish your sentence before it asks anything. There is now a camera button beside the microphone, and you can reach it from the Lock Screen, Control Center, the Action Button or Siri. RESTORATION, all of it on the device, with models downloaded on first use rather than shipped in the app: four-times upscaling, damage repair for scratches, creases and spots, noise reduction, motion-blur sharpening, color for a black-and-white photograph, frame interpolation, and de-interlacing for footage that came off tape or a DVD. Stack them over a whole video in one pass with Process video. GENERATING — Ari now makes still pictures on its own, on the device, with no Draw Things needed. And when you do use Draw Things, Ari knows what each video model wants: sampler, schedule, guidance, size, clip length and frame rate come from the model itself rather than from whatever was last set for an image, which is why Wan and Flux produce video now instead of noise. A video generation shows a live color preview and a step count while it runs, and a long one is no longer mistaken for a dead one. MARKS — generated pictures can carry an invisible mark that lives in the image itself, not in metadata that any edit would strip. Check any picture from its menu, or at wemiller.com/apps/ari-helper/watermark/, which runs in your browser and uploads nothing. Both say plainly what a negative means: a mark that is found is reliable, a mark that is missing proves much less. Also: choose the voice for spoken replies, under Settings then Local Models. An Enhancement models list shows what is downloaded, its size, source and license, and lets you remove it again.

Ari 6: Apple Intelligence sees your images on OS 27, your devices talk over Wi-Fi, and remote dispatch waits its turn. - The Mac app is now Apple silicon only. If you are on an Intel Mac, Ari stays at version 5.12 — it keeps working, but this and later updates need Apple silicon. - On iOS 27 and macOS 27, the images you attach go straight to Apple's on-device foundation model, so it reasons over the picture itself rather than a description of it. Up to 8 per message, and nothing leaves the device. - On iOS 26 and macOS 26 — or wherever the model can't take images — nothing changes: Ari still describes each image on-device with Vision, and the attachment strip and Settings say which path is in use. - Your devices now talk to each other over your own Wi-Fi. When Ari is open on two of them, a conversation — and the pictures in it — arrive in about a second instead of waiting on iCloud. Private conversations stay encrypted the whole way, and iCloud still carries anything that misses the local link. - Pausing a queue of dispatched prompts and pressing play now resumes it, and stays resumed. - Dispatched queues keep running on a Mac with every window closed. - Dispatch Priority, a new setting under Devices & Dispatch, decides when this device runs prompts sent from your other devices: Immediately, When I'm not prompting (the default — your own prompts go first, and a dispatched prompt already running steps aside and re-queues), or Only while this Mac is idle, which waits until the Mac has seen no input for three minutes. - Conversations take turns again. A template or repeat batch running in one conversation no longer holds every other conversation's queue until it finishes — batches share the machine one prompt at a time, round-robin. - A draft no longer sticks to a conversation after you send it, and clearing the box clears the saved draft too. - API Settings is now Settings: a hub of pages — Cloud Providers, Self-Hosted & Network, Local Models, Privacy & Security, Devices & Dispatch — instead of one long scroll. Manage Local Models lives there as its own page. - On iPhone and iPad the sandboxed browser is now a sheet you open from the menu while a code session is active, with an address bar, back and forward, the Open Web switch and zoom — drag it to half height or pull it up full screen. - The tool-call log on iPhone and iPad is now a sheet you can search, with a page per call showing the parameters, output and any error. - Queued prompts from different conversations now take turns properly. A long templated or repeat run no longer holds the device until it finishes while another conversation waits. - Fixed a crash that could end the app while it was in the background waiting on iCloud, and stopped the background iCloud polling that caused it. - The home screen is translated end to end: the suggested prompts, the greeting and the "New Chat" title now appear in your language.

Bring your own server — and remote dispatch that works. Ari now talks to any local or self-hosted server that speaks the OpenAI API — LM Studio, vLLM, llama.cpp, LiteLLM, LocalAI, Jan, or your own gateway. Add as many as you like; each gets its own group in the model picker. - Add the address, tap Validate Connection, and the models your server is serving fill the picker. An API key is optional and stays in the Keychain. - Paste it however your server prints it. A LAN address, a bare hostname, a Bonjour ".local" name, a ".lan" or ".home.arpa" name from your router, or Tailscale all work over plain HTTP; anything off your own network needs HTTPS. - Fortress Mode judges each endpoint by where you pointed it, so a server on your own network stays available while a public one beside it stays blocked. - It shows up everywhere the other providers do — Shortcuts, Siri, the Apple Watch model list, and device-to-device dispatch. Remote dispatch is fixed. It had stopped working entirely, and the cause left no error behind: each device was silently reading an empty queue. Along with it: - Several dispatch queues aimed at the same image server now share one line, a prompt at a time, in the order the work arrived, instead of one conversation running several before another gets a turn. - Stop actually stops, and cancelled work no longer reappears moments later. - A prompt that fails now tells the device that sent it why, instead of going quiet. Private conversations sync across your devices again. The encryption passphrase was being kept on each device rather than in your iCloud Keychain, so every device held a different key. Nothing already encrypted is lost — each device's key is merged in. Also fixed: a crash, a new private conversation opening as a plain one, a private conversation that could not be deleted, duplicated replies, and a media gallery that reloaded and lost its zoom mid-reply.

Fixed: a crash on opening. Some installs quit the moment you tapped the icon — an iCloud check ran before the app had drawn anything, and when iCloud wasn't available on that device it took the whole app down. That check now happens after the app is on screen, and it fails quietly instead of fatally. If Ari wouldn't open for you, it will now. Feedback & Support, where you can find it: - "Support & Feedback" is now its own item in the main menu. Report an issue, suggest a feature, ask a question, rate Ari, or browse my other apps — one tap from the chat screen instead of buried at the bottom of API Settings. - On Mac, the same buttons are in Settings, which previously had none of them at all. Also fixed: - The Previous Chats drawer no longer looks washed out, and the delete button is red again rather than pale pink.

Now fully localized in 8 languages: Simplified Chinese, Japanese, German, French, Spanish, Korean, Brazilian Portuguese, and Italian. The entire app, from settings to system prompts, now reads naturally in your language. New Feedback & Support: - Contact the developer directly from the app to ask a question or report an issue. - Spotted an awkward translation? Send translation feedback in one tap and help make Ari better in your language. Rating improvements: - The rate-Ari prompt now appears at a sensible moment and no longer disables itself forever after one dismissal. Chat on your Claude subscription (Mac): - The new Claude Agent provider sends messages through Claude Code, so an existing Claude subscription pays for replies instead of a metered API key. Ari never sees your Claude credentials, and the provider stays hidden until you set it up yourself. Remote dispatch, live: - Dispatch a prompt to another device and watch the answer arrive as it is written, end-to-end encrypted. A locked device simply waits rather than showing a fragment. Local models, shared fairly: - Run several conversations against the same local model at once. Ollama, Draw Things, and on-device models take turns fairly, and the queue shows your place in line. Draw Things, upgraded: - Live previews and per-step progress, image-to-image and inpainting from an attached photo, video generation, LoRA and ControlNet support, and cancellation. Apple Intelligence: - Keeps its tools across conversation turns, and can use Ari's own tools: reading and writing files, searching, running commands, and inspecting git, on device or through Private Cloud Compute. Also fixed: - Text-to-speech no longer repeats itself, and the audio meter no longer competes with speech recognition. - Tailscale connections are supported, and long runs can be cancelled cleanly. - iCloud and XPC work moved off the main thread, so the app stays responsive during heavy syncs.

SuperGrok sign-in for Grok models — use your SuperGrok / X Premium+ subscription instead of (or alongside) a console API key. Prefer SuperGrok-only billing when you want, with automatic session refresh via iCloud Keychain. Code Session tooling is much more reliable: Grok tool calls are parsed and forced when needed, shell commands work better under App Sandbox (Shell Helper setup), and the Mac sidebar now tabs Browser, Terminal, and Activity so they don’t all fight for space. Open multiple browser tabs and terminal sessions side by side. Session diffs, expandable activity rows, and multi-agent task settings round out the agent workflow. Also in this release: image attachments with Apple Intelligence via on-device Vision; Grok 4.5 agentic coding; Local Inference API; OpenAI GPT-5.6; and animated-image fixes.

OpenAI's new GPT-5.6 models are here. Pick GPT-5.6 Sol (flagship), Terra (balanced), or Luna (fast and cost-efficient) in the model picker and in Shortcuts — and when Ari auto-selects a model for you, complex reasoning, coding, and research now go to Sol while everyday chat uses Terra. Fixed: animated images (GIF, APNG, WebP) streamed from a personal media server with a self-signed certificate no longer show a black screen in the full-screen viewer.

Lighter, faster install. Ari no longer bundles a one-size-fits-all AI model, so the download is dramatically smaller and installs much quicker. Apple Intelligence is now the default. On supported devices (iOS 26 / macOS 26) Ari chats with Apple's private on-device model right out of the box — no setup, no API key, and nothing leaves your device. Optional on-device vision, sized to your device. On first launch Ari offers to download Google's Gemma 4 model so you can send photos and images to a fully on-device assistant that works offline. Ari automatically picks the best-fitting Gemma version for your hardware instead of shipping one heavy copy for everyone — and you can always skip it. Manage or switch models anytime in Settings. Fixed: a downloaded model could occasionally appear twice in your model list. Downloads are now de-duplicated correctly.

Lighter, faster install. Ari no longer bundles a one-size-fits-all AI model, so the download is dramatically smaller and installs much quicker. Apple Intelligence is now the default. On supported devices (iOS 26 / macOS 26) Ari chats with Apple's private on-device model right out of the box — no setup, no API key, and nothing leaves your device. Optional on-device vision, sized to your device. On first launch Ari offers to download Google's Gemma 4 model so you can send photos and images to a fully on-device assistant that works offline. Ari automatically picks the best-fitting Gemma version for your hardware instead of shipping one heavy copy for everyone — and you can always skip it. Manage or switch models anytime in Settings. Fixed: a downloaded model could occasionally appear twice in your model list. Downloads are now de-duplicated correctly.

Ari now brings Apple Intelligence chat. On supported devices Ari talks to Apple's on-device foundation model directly in the app — streaming answers, full conversation history, and text-to-speech, with zero data leaving your device and no API key. If Apple Intelligence is off or still downloading, Ari explains why and falls back to its built-in local model automatically. The "Ask Ari" Siri shortcut uses it too when available. Find any chat with Spotlight — even by what's in its pictures. Regular conversations now appear in system Spotlight, and Ari can index the contents of their images entirely on-device: text in screenshots and photos becomes searchable, and recognized objects and scenes become keywords. It's controlled by a new "Index Image Contents" setting, and private conversations, their media, and Fortress Mode sessions are never indexed. New sensitive-content guards: when Sensitive Content Warning is enabled in iOS Settings, Ari checks every chat image — attached, AI-generated, or incoming — on-device before showing it, and blurs anything flagged behind a "Show Anyway" interstitial. Adults who verify their age with Apple's privacy-preserving Declared Age Range check can optionally skip the warnings. Ari is now a first-class Shortcuts citizen. Every action takes a dynamic AI Model parameter — Apple Intelligence, any downloaded local model, or any configured cloud provider — resolved live. New actions: Start Conversation, Generate Image, Edit Image, Analyze Documents, Speak Response, Get/Set Ari Setting, Run Code Session Task on macOS, and Dispatch Task to Mac. Private conversations stay invisible to Shortcuts, and Fortress Mode applies everywhere. And Ari is on your wrist. The new Apple Watch app is a remote control for Ari on your iPhone: dictate or scribble a question and the answer streams back, generated by whatever model you have selected — including fully-offline local models. Add the "Ask Ari" complication or Smart Stack widget, have replies spoken aloud, hand long answers off to your iPhone, and let questions queue automatically when your iPhone is out of reach. Private conversations are never accessible from the watch.

5.4.0 is a major step up for agentic coding. The AI can now drive a sandboxed in-app browser, take screenshots of your local dev server, edit files between turns in a TreeSitter-powered editor, and run real git operations via libgit2 — including push and pull with your saved Personal Access Token, in-memory anonymous remotes so no token ever lands in your .git/config, plus stash, amend, and force push. Local model downloads also got a polish pass. Multi-shard MLX repos finally show smooth byte-accurate progress instead of the old staircase. Add a Hugging Face token to unlock gated models like Gemma. Cancel + retry buttons on every in-flight download, with proper partial-file cleanup.