Changelog

1.5.0

New

• ●Name a genre or a drum style in your own language. Typing "musique classique", "klassische Musik", "テクノ" or "레게" now reaches the same guidance as typing the English name, on the Melodic tab and the Drums tab both. Around 1,700 genre names and 2,000 drum style names, covering all sixteen genres and thirty-six drum styles in every language they have a name in. English is unchanged, and the dropdowns still override whatever you type.
• ●Aether answers in your language. Write to it in French, Japanese or Portuguese and the reply, the composition plan and the chat come back in that language instead of English. What it writes is unaffected: the key, the scale and the drum piece names stay in their standard form, so the notes still land where they should.
• ●Drop in a Drum Rack instead of retyping it. Drag an Ableton .adg from the browser onto the Drums tab and Aether reads the rack: every pad, its note and its name. Names it recognises land on the standard pieces, the rest become pads of their own, and nothing is thrown away. The mapper opens with the rows filled in and the rack's name in the box, so you can correct anything before you save. A large multi-mic kit reads in the background, so the window does not freeze while it works, and dropping the same rack again edits the map it made last time rather than making a second one.
• ●Your own pads, named your way. A drum map can now hold pieces the sixteen standard names do not cover. Type a name like PML_MBH_Perc_00 or Cajón into the map editor and it stays, instead of snapping back to Unassigned. A name written in Japanese, Cyrillic or Chinese stays as you typed it too. Aether tells the model those pads exist, so it can write for them, and their hits land on your notes instead of being discarded. A rack with three pads sharing a name gets all three, numbered.
• ●A note per pad. Each row in the map editor has a pencil that opens a small field, so you can record what an opaque pad actually is. Aether passes it on with the pad's name. Optional, and empty by default.
• ●Eight more drum styles. Alt-Pop, Afrobeats, Country, Hardstyle, Jersey Club, Phonk, R&B / Neo-Soul and Reggae each have their own kit, groove and fills instead of falling through to a neighbour.

Improved

• ●The first-run tutorial covers Drums, model refresh and custom prompts. The walkthrough now switches to the Drums tab and points at the real controls instead of skipping the tab entirely: what the tab is, the column that shapes the next beat, and the MIDI Map, Edit map and Sound pickers. Two new Settings steps cover refreshing the model list from your provider, and the System Prompt tab with its separate active prompts for Melodic and Drums. Fourteen steps, up from nine. It also opens Quick Params, Library and Drum Controls when it points at them, rather than highlighting a collapsed rail and describing controls you cannot see. The Settings steps open Settings on the right tab with the control highlighted, then put your own tab back and close the panel again if you had it closed.

Fixed

• ●What you type reaches the style you meant. Style detection picks the longest matching keyword now, instead of the first family that happened to contain one. "Amapiano house" was answered by House, "indie pop" by Rock, and "garage rock", "industrial metal", "80s music", "liquid" and "radio" each pulled a style out of a family they had nothing to do with.
• ●Afrobeats and Afrobeat are different genres. Both spellings reached the Afrobeat module, which is a 1970s live band: ghost-heavy, displaced snare, congas, around 112 BPM. Afrobeats is 2010s West African pop, programmed and sparse, and now has a module of its own. "Fela" and "highlife" still reach Afrobeat, which is correct.
• ●Drum and bass comes back as the two-step. Asking for drum and bass now puts snares on beats 2 and 4 of the bar at around 174, with the kick on 1 and the "and" of 3. It used to write a single snare on beat 3, which is half time and reads as a different style at half the speed. Ask for "half time" to get that pattern instead. Jungle and UK garage are unaffected.
• ●A BPM you set wins over the style's own tempo. Only the Pieces tag was binding, so a style's reference tempo could quietly replace the tempo you asked for. Setting a BPM while leaving Bars on Auto also sent "the tempo is yours to choose" in the same request. Both now say your tag wins.
• ●A beat fills the bars you asked for. The clip length was a field with no rule attached, so a beat could stop short and leave the last bar empty. The first hit belongs in bar 1, the last in the final bar, and the length has to match the bar count.
• ●Trap and big room can write 32nds. The beat grid listed eighths, sixteenths and triplet eighths, then forbade working out any other position, so a trap hat burst and a big room fill accelerating into 32nds could not be written as asked. Thirty-seconds and triplet sixteenths are on the grid now.
• ●The export shape selector is on screen. Combined, Multi-track and Single piece had room reserved in the Drums controls column but were never added to it, so the selector has been invisible since it shipped and the shape stayed on whatever your settings file held, which for almost everyone was Combined. The buttons and their caption are there now, and clicking one takes effect and is remembered.
• ●The drum map card appears in the chat. It was built on every drum generation and then thrown away before it reached the screen, so the legend showing which note each piece landed on has never been visible on the Drums tab.
• ●The drum ruler counts beats. It showed bar numbers alone at the usual window size, with nothing between them. It now reads 1.1, 1.2, 1.3, 1.4, 2.1 like the Melodic roll, and the grid tells bars, beats and sixteenths apart.
• ●A map saved by an older build reads "Kick", not "kick". Maps written before pads carried labels showed the internal name in the lane gutter and on the chips. They read properly now, and a pad you named yourself is left exactly as you typed it.
• ●A dropped loop that carries no tempo now gets a kick and a snare. Ableton's Export MIDI Clip writes no tempo, and Aether used to refuse to measure timing without one, so it proposed the hat and left every other pad for you. It now measures against the session tempo, and the import dialog says when it did. A file that declares its own tempo is unaffected.

1.4.0

New

• ●A Drums tab. Own chat, controls, clip and provider. Switching tabs keeps both. A finished beat never yanks you over.
• ●Generate is now Melodic. Same tab, new name, now that Drums sits next to it.
• ●Beats land on the right pads. Aether assigns every drum pitch. The model only names pieces. Exports on channel 10.
• ●Kit lanes instead of a piano roll. One lane per piece, kick at the bottom, solo and mute on the lane, velocity as opacity.
• ●About thirty styles. Boom bap, trap, drill, house, techno, jungle, UK garage, afrobeat, amapiano and more. Or just describe it.
• ●Live controls, no generation. Quantize, swing, humanize and ghost depth edit the beat you already have and bake into the drag-out.
• ●Three preview kits. Live Kit, Electronic Kit, Hip Hop Kit. Sound is what you hear in Aether. MIDI Map is what leaves for the DAW.
• ●Batch on Drums. 2 to 5 beats from one prompt, told apart in drum terms so they diverge. Numbered buttons switch between them.
• ●Drum map in the reply. Note, piece and pitch names, plus Copy map. The Note names chip picks Live or scientific conventions.
• ●Three export shapes. Combined (one track), Multi-track (one track per piece), Single piece (one lane, C3, for a one-shot sampler).
• ●Drop a MIDI beat in. It becomes the clip and the next-prompt reference. Tempo and meter fill in. Live controls work on imports.
• ●Any pad layout, not just GM. Import opens a mapper with a guess and hit counts. Save the map. The next loop from that rack comes in filled.
• ●Unmapped pieces still play. Missing entries fall back to GM and say so. Changing a map re-reads the beat on screen. No regenerate.
• ●Sections on Drums. Named parts with bar counts (Intro 8, Build 8, Drop 8). Suggestions follow the style.
• ●Beats go to the Library. Typed as Drums or Melodic, filtered by the tab you are on. A batch is one card. Add to chat uses a beat as the next reference. A saved beat reopens on the map it was written under, even after you rename or delete that map.
• ●Separate system prompt for drums. Settings has Active prompt for Melodic and for Drums. Drum starts empty so an old melodic prompt does not steer beats.
• ●Pro Mode on Drums. Plan, then write. One toggle for both tabs. With Batch each variation is planned on its own, so five variations is ten calls.

Improved

• ●Quick Params, Chat and Library are resizable. Drag the seam. Widths persist. Chat is shared by both tabs.

Fixed

• ●ChatGPT and Gemini now show the same Thought-panel plan as Claude, Grok and DeepSeek.
• ●Library cards keep ratings and action buttons at any panel width.
• ●Library clips, imports and older takes loop at the last bar, not through trailing silence.
• ●The first batch variation appears as soon as it lands.
• ●Standalone keeps settings changed during the session. Window and audio device state now live in their own file.
• ●Melodic layers are no longer auto-assigned to channel 10. Manual routing there still works.
• ●A batch is one undo step, starts off after a restart, and unlocks the model picker when you cancel.
• ●Gemini generations include genre again.
• ●Imports fill Key, BPM and Time Sig from every drop route, not only the chat box.

1.3.2

New

• ●Importing your own MIDI now sets the key and tempo. Aether fills the Key and BPM controls from the file and highlights them so you can see where the values came from, and the chat says which ones it took. Both were previously described to the AI but never pinned, so about one variation in three drifted out of key or tempo. Files with no clear key are left alone rather than pinned to a guess.
• ●Timing checks work on imported material. An empty BPM box used to switch off the check that compares a finished clip against the beat it was asked for. A tempo read from your import turns it back on.
• ●Three MIDI references instead of two. Aether also tells you which files did not fit when you drop more.
• ●Multi-track MIDI keeps its parts. Tracks arrive named instead of being merged into one block of notes, so "an alternative guitar idea over the existing harmony" now has something to work with. A multi-track file counts as one reference, so a four-part arrangement does not use up your three slots.
• ●Clear MIDI. A button beside New and Redo, shown whenever files are attached, that detaches them. Removing the wrong file no longer means starting a new chat.
• ●A batch is one card in the library, not one per variation. Expand it to see the variations, each of which plays on its own. None is promoted above the others, because none supersedes anything. Batches saved before this update stay as they are.
• ●A play button on every clip in the library. Every row has one, including the earlier versions under an idea, whether or not the row is selected. Double-clicking still auditions a clip; nothing on the row ever said so.

Fixed

• ●The local model picker no longer lists models you do not have. Ollama offered six popular model names and LM Studio offered one called Default Loaded Model, none of which came from your machine, so choosing one could fail against a model you never installed. The picker now asks you to click Refresh Models, then lists what your server actually reports and remembers it between visits.
• ●Local AI models could not connect from inside a DAW. Ollama and LM Studio worked in the standalone app but reported the server as unreachable when Aether ran as a plugin, even with that server running. Test Connection and Refresh Models now go through Aether's network helper, the way generation already did, instead of opening a connection your DAW blocks. Neither freezes the window while it waits, and if the helper is the missing piece the message says so.
• ●Picking a local model from the refreshed list did nothing. After a refresh, choosing anything below the first entry left the previous model in place. The list is now wired to the models it is showing.
• ●Batch generations were saving nothing at all. Each batch created a folder named after your prompt and left it empty, because the filename Aether built held the model's entire response and was far too long for the filesystem to accept, and the failure went unreported. Single generations were never affected.
• ●Genre, Mood and Texture can be cleared again. All three menus now have a "None" row.
• ●The library tells you which version you are looking at. Expanding an idea with two takes showed a single row labelled v1, which read as though the newer take was gone. It was not: that row was the newer one, and the only one that never said which version it was.
• ●Dropping files one at a time adds them instead of replacing. Each drop used to clear whatever was already attached, so three parts dropped in turn left only the last.
• ●The attachment tag names every file you attached, not just the first.

1.3.1

New

• ●New models arrive with their controls already working. The reasoning-effort setting was only offered for models Aether had been told about by name, so it quietly disappeared every time a provider released something new. Aether now assumes a new model supports it and stops sending it if the provider refuses, so a model released tomorrow gets its controls today.
• ●New OpenAI reasoning models appear in the picker on their own. The model list recognised the o-series one generation at a time, so a new one would not have shown up at all: not missing a feature, missing from the menu, with nothing to explain why.
• ●Every layer has its own channel selector. Output routing lived in Settings and had exactly three rows (Bass, Harmony and Lead), so an arrangement with a pad, an ostinato and a horn section could only have three of its layers pointed anywhere deliberately. Each row in the Layers panel now carries its own channel, and the remaining layers are given free channels around whatever you set, so nothing ever quietly shares one. Your choices are saved with the project and follow the layer, so re-voicing a part keeps it where you put it.
• ●The channel shown next to a layer is the one it exports on. It used to be the preview channel, which stopped matching the file once your routing settings started being applied to it.
• ●Live MIDI output follows your channels too. Monitoring a layered clip out to another instrument or to hardware sent the entire arrangement on a single channel, whatever you had set, because the channel was being dropped before the notes left the piano roll. Every layer now goes out on the channel it is actually assigned, so what you hear while monitoring matches what you get when you drag the clip out.

Changed

• ●Aether remembers which model you use with each provider. Switching from ChatGPT to Grok and back reset you to whatever model was current when the plugin was built, throwing away the choice you made. Each provider now keeps its own, and the built-in default only applies the first time you use that provider.
• ●Grok's model list is current again. xAI retired ten of the models Aether offered, including the one it picked by default, so anyone whose list had not refreshed was choosing from mostly dead options. The list now covers Grok 4.6, 4.5, 4.3 and 4.20, and the default is Grok 4.6.
• ●Cancel in Settings discards the provider and model too. Changing them applied immediately, so "Discard changes and close" kept the very thing you were trying to undo.
• ●Dragging a clip in can no longer change a sound you already chose. Exported files carried a General MIDI patch number at the top of every track. That is an instruction rather than a note, and the instrument already sitting on the track you dropped onto would obey it and switch its own preset, which looks, from where you are sitting, like the plugin reaching into your session and altering your instrument. Aether no longer writes them. If you are exporting for a General MIDI player and want the patch numbers back, there is a new switch in Settings → MIDI under Exported Files.

Fixed

• ●GPT-4 and GPT-4 Turbo failed every single generation. Aether asked for more output than either model allows and the API refused outright. Both work again.
• ●GPT-5 models could stop mid-clip. They were given a quarter of the output budget they support, and because thinking is charged against the same budget, a longer arrangement could run out of room before the music was finished.
• ●Reasoning effort works on the Pro models. GPT-5 Pro and GPT-5.5 Pro were sent the setting in a form Aether could not correct if it was refused, which could leave the model unusable with no control left to switch off.
• ●Changing provider mid-generation no longer throws the result away. A finished, paid-for generation was read by the newly selected provider instead of the one that produced it, and reported to you as an empty response.
• ●Voice, coding and research models are out of the picker. Grok's speech and coding models and OpenAI's deep-research models were all listed and none of them can write MIDI, so picking one failed every time.
• ●A model that gains reasoning support gets it back. When a provider refuses the effort setting Aether remembers and stops sending it, and that memory was permanent, so a model that later gained support would never have offered the control again. It now rechecks periodically.
• ●Your MIDI channel settings now reach the file you drag into your DAW. Part Output Routing was only ever applied to Aether's live MIDI output, while the file the drag hands your DAW numbered its channels by the order the layers came back in, so setting every part to channel 1, or to anything else, changed nothing about what actually landed in your project. Parts now follow the channels you set, and extra layers the model invents, like an ostinato or a named horn section, take the next free channel instead of quietly landing on top of a part you had already routed.
• ●Clips dragged out in Combined mode are the length you asked for. A region ended a quarter note past the last note instead of on the bar the clip was written to, so an eight-bar idea whose final chord stopped early arrived as an awkward eight-and-a-bit. Multi-track drags and clips saved to the library were already correct; this was the one path still guessing the length from wherever the notes happened to stop. Dragging a single layer out on its own was affected the same way.
• ●Combined mode flattens the channels, not just the tracks. The setting reads "Combined (flatten to Output Channel)" and only ever delivered half of that: you got a single track, but the notes inside it kept the separate channels Aether uses internally to tell bass, harmony and lead apart. An instrument listening on one channel heard a fraction of the arrangement, and the only way through was to renumber every item by hand once it was in the DAW. Combined now puts every note on the Output Channel you picked.

1.3.0

Highlights

• ●Clips land in time. This was several separate faults, not one. The section editor was silently dropping the timing reference Aether sends. Some styles asked every part to sit behind the beat, which just makes a clip late. Quantize offered straight and swung grids at once, so tightening a note could change its rhythm. And Aether trusted the tempo written into each response: a clip written correctly at your project tempo but labelled with the genre's usual tempo got stretched to match the label, sometimes to twice its length.
• ●Aether checks its own work. Every generation is measured against your project's bars and beats before you see it. If the notes and the reported tempo disagree, the notes win and the label is corrected. If a clip cannot be verified, Aether says so instead of handing it over quietly. Deliberate feel is safe: a bass on the beat with keys leaning behind it is a pocket, not an error. Styles with no fixed pulse, like ambient and cinematic, are exempt.
• ●The sustain pedal works. Aether has been writing pedal markings all along and every one was discarded before it reached playback, the piano roll or an exported file. Ballads, ambient, classical and gospel lost the sustain the writing assumed. It now survives everywhere, including both kinds of drag-out.

Fixed

• ●Sections stay in time from the first bar to the last. With the section editor on, longer forms drifted further out with every section. Each section now gets its own start and end time instead of adding bar counts up.
• ●Laid-back styles feel laid-back instead of late. R&B / Neo-Soul, lo-fi and ballads put every part behind the beat at once, and nothing anchors a groove if everything moves. The bass now holds the beat while the keys and lead lean back against it.
• ●Dragging out gives you what you were hearing. Dragging a multi-part clip as separate tracks rebuilt it from the untouched original, so your Clean or Enhance settings were missing. Dragging the same clip as one track included them.
• ●Asking for a bassline no longer means asking for only a bassline. Naming a layer anywhere in your message was read as a restriction, so "drum and bass, rolling sub bassline and atmospheric pads" gave you a bass part and nothing else. "Just bass", "melody only" and "chords only" still mean what they say.
• ●Aether remembers the style you asked for. Naming a genre set it for that message only, so following up with "make the bass busier" lost the style. It now carries through a conversation until you change it.
• ●Clips are the length you asked for, and loops loop. A clip whose last note stopped before the final bar arrived in your DAW short, and loop mode waited out the last note instead of returning at the bar line. Both now follow your bar count, whether you set it in the Bars field or wrote "an 8 bar loop".
• ●Changing your project tempo no longer stutters playback. Automating tempo in your DAW made Aether rebuild and restart the clip twice a second while you were listening.
• ●Electronic styles fit the length you asked for. Trance, drum & bass, dubstep and festival each described a full arrangement with fixed section lengths, so a shorter clip squeezed that whole shape into the space available. A 32-bar trance spent half its length in breakdown with the rolling bassline missing.
• ●No more drum parts. Aether does not write drums, but nothing in its instructions said so, and drum & bass kept adding a breakbeat layer you had to delete. Styles still write around the drums you will add, which is why a trance bassline leaves the downbeat clear.
• ●Quantize snaps to the grid your music is on. It offered straight and triplet grids together, so a slightly early sixteenth could get pulled onto a triplet, changing the rhythm instead of tightening it. It now uses the one grid the style works on, and styles with no fixed pulse get the sparsest option rather than the densest.
• ●Dropped MIDI keeps its starting position. A file beginning with a bar of rest had that bar removed. Recordings you make in Aether still snap to your first note.
• ●The BPM box shows your real project tempo. With Host on, a session at 128.5 read as "128". Only the display was rounding.
• ●Trance leads hold their hook. The lead was picking up stray notes a sixteenth off the pulse, which against a rolling bassline reads as a stumble.
• ●Chord preservation works with the section editor on, clip names count bars correctly in 6/8, and changing your audio device's sample rate no longer leaves a loaded clip playing at the wrong speed.

Changed

• ●Time signature defaults to 4/4 instead of "Let AI Choose". Aether can only send its timing reference when it knows the meter, so the old default left generations without one. "Let AI Choose" is still there.
• ●UK garage's shuffle is a little tighter, and holds steady across a clip instead of varying note to note.

1.2.2

Improved

• ●Every genre now writes like itself. Each of the 16 styles carries exact note placements, velocity shapes and layering rules of its own, instead of generic advice. Trap leaves the 808 hitting alone, funk's bass and chords interlock instead of stacking, bossa uses its real two-bar comping pattern, and gospel's vamp escalates each time round.
• ●House / EDM is now Electronic, and it understands what you asked for. One setting used to cover house, techno, drum & bass, dubstep, garage and trance with a single set of festival-oriented instructions and no sense of tempo, so asking for drum & bass gave you house-speed drum & bass. Say what you want and you get it: warehouse techno stays hypnotic and hookless, liquid drum & bass rolls at 174, UK garage swings and skips, dubstep sits in half-time, trance gets its long breakdown, and asking for nothing in particular still gives you house.
• ●Sections sound like sections. A bridge or solo that previously only moved higher and louder while reusing the same rhythm is now required to change where the notes land.
• ●Quiet genres are allowed to be quiet. Ambient, lo-fi and trap were being pushed toward pop density and dynamics by rules written for pop. Those rules are now defaults a style can override, so sparse music stays sparse and a pad layer can breathe.
• ●Mood and tempo now affect the writing itself. Asking for something dark or bright, or working well outside a style's usual tempo, changes the harmony, register and note placement rather than just the description.
• ●Classical understands specific forms. Ask for a waltz, nocturne, fugue, sonata or a particular era and you get guidance for that form instead of a summary of all of them at once.

Fixed

• ●Naming a style in your prompt actually works. Genre detection was reading the message box after it had already been cleared, so free-text requests for house, drum & bass, a jazz ballad or a waltz were ignored and fell back to general songwriting. They now reach the right style.
• ●Velocity colour is visible in the piano roll again. Multi-part clips were rendering with almost no difference between soft and hard notes.
• ●Renaming a clip no longer flattens it. A renamed clip was reloading as a single merged layer and losing its version history.
• ●The playhead now plays from where you put it. After pausing, moving the playhead and pressing play was resuming from the old position.

1.2.1

Improved

• ●Generations are substantially faster. Aether now reuses the unchanging part of its instructions between requests instead of paying to re-read them every time, and lets you tell each model how long to deliberate before writing. The longer you stay in one session, the more of the work is already cached.
• ●Pro Mode is dramatically faster when editing. Asking Pro Mode to change an existing clip made its planning pass quietly write a whole arrangement and throw it away before composing the real one, so every edit did the work twice and the wasted attempt was carried into the second pass. The planning pass now plans. Edit turns finish in a fraction of the time, and the plan itself is unchanged in quality: it still studies every note of your clip.
• ●Reasoning-effort selector now appears for Gemini. The Low / Medium / High control shows for Gemini 2.5 and later. On Gemini, Low switches the model's deliberation off entirely, which is the fastest setting available; Medium stays the default, so nothing changes unless you choose it.

Fixed

• ●Pro Mode no longer stops with a "rate limit" message on fast generations. Now that a generation can finish in a couple of seconds, Pro Mode's composing pass could arrive too soon after its planning pass and be turned away by Aether's own request throttle. The plugin was rate-limiting itself. That throttle now applies only to requests you start.
• ●Clearer message when a provider returns nothing. If a model takes so long that the connection closes before it replies, Aether reported "Invalid JSON response from API", which pointed at a formatting problem that did not exist. It now says the provider returned an empty response, and suggests retrying or switching to a faster model.
• ●Saved sessions no longer carry unnecessary MIDI data. The routine that trims note data out of saved conversations was still looking for an older format and had quietly stopped matching anything, so session files were larger than intended.
• ●Adding a layer no longer sometimes merges your existing layers. When you asked to add a new layer, the existing arrangement could occasionally collapse into a single combined track instead of staying as separate layers. Edits now build on your real, named layers every time.
• ●Reasoning-effort selector now appears for Grok. The Low / Medium / High control that trades speed for depth now shows for the Grok models that support it (such as Grok 4.5), so you can dial in how hard Grok thinks before composing.
• ●Attached MIDI files now land at your project tempo. MIDI exported from Ableton Live carries no tempo information, so Aether read those clips at a fixed 120 BPM while telling the model your session tempo. A 4-bar idea dropped into a 140 BPM project came through as roughly 4.7 bars, and whatever Aether wrote back sat off the grid. Attached files now use your session tempo when the file does not specify one, and their own tempo when it does. Files that happened to be at 120 BPM were always fine and are unchanged.

1.2.0

New

• ●Compose in layers, not just Bass / Harmony / Lead. Aether now builds arrangements from any number of named, instrument-tagged parts. Ask for "a horns section" or "an ostinato under the piano" and the new layer is added alongside what you already have instead of replacing it. Each layer carries the instrument name the model chose, normalized to a General MIDI sound.
• ●Layers view. A toggle in the toolbar opens a dedicated panel listing every layer with its instrument, a mini piano-roll preview, and per-layer Solo / Mute. A header shows the detected key, tempo, and layer count, and you can flip the labels between instrument names and part roles.
• ●Named, GM-mapped MIDI export. Dragging or exporting a multi-layer clip now writes one named track per layer with the matching General MIDI program change, so your DAW loads each part onto the right instrument automatically.
• ●Drag a single layer to your DAW. Alongside the whole-clip drag, each layer has its own colored handle you can drag out on its own.
• ●Per-layer colors in the piano roll. Every layer gets its own color so overlapping parts stay readable at a glance.
• ●Layers work on every provider. Claude, ChatGPT, and Gemini are now asked for layered output through their own structured-output formats, so named, instrument-tagged layers come back whichever model you pick. (Older Gemini models, such as 2.0 Flash and the 1.5 family, still return the classic Bass / Harmony / Lead trio.)
• ●Custom layers get the same polish as the main parts. Note clean-up, pitch clamping, and humanization now run on every layer you create, including instrument-tagged ones, not just Bass / Harmony / Lead.
• ●GPT-5 Pro models now work. OpenAI's "pro" models (such as GPT-5.5 Pro and GPT-5.4 Pro) run on a different endpoint from the standard chat models and previously failed to generate. Aether now talks to that endpoint automatically when you pick a pro model, so they compose like any other.

Improved

• ●Reasoning-effort control now covers OpenAI, and Pro models start fast. The Low / Medium / High reasoning selector, previously Grok-only, now also appears for OpenAI's o-series and GPT-5 models (including the Pro tiers). Each provider keeps its own setting, and OpenAI reasoning models default to Medium, so GPT-5 Pro composes in a couple of minutes instead of running its default High effort for many minutes on a short idea. Turn it up when you want maximum depth.
• ●Solo reflects what you actually hear. Soloing a layer now dims the other layer buttons in the toolbar, so the top bar always shows what is currently audible.
• ●Seamless muting for overlapping layers. Muting or unmuting one layer while parts share pitches no longer briefly cuts or re-triggers the notes held by the other layers.
• ●Onboarding tutorial covers layers. The walkthrough gained dedicated steps for the layer controls and the Layers view.

Fixed

• ●Runs on macOS 11 (Big Sur) and later, on both Apple Silicon and Intel Macs. A build setting had raised Aether's minimum to the very newest macOS, so on earlier systems the standalone app would not open ("cannot use this version of the application with this version of macOS") and the plug-in failed to load in your DAW. Aether now installs and runs across supported macOS versions again.

1.1.5

New

• ●Choose which OpenRouter providers you see. OpenRouter exposes hundreds of models, so the list could be overwhelming. In Settings, a “Providers” picker lets you tick exactly which upstream providers (Anthropic, OpenAI, Google, and so on) appear in the model dropdowns, with Select all / Clear all and Save.
• ●Per-version actions in the Library. When a saved idea has multiple versions (v1, v2, v3...), each version now has its own actions menu: Add to Chat, Show in Finder, and Delete.

Improved

• ●Seamless part muting. Muting or unmuting Bass, Harmony, or Lead during playback no longer briefly doubles the other parts' notes; only the part you toggled changes.
• ●Clearer Library buttons. Action buttons now have tooltips, the redundant “Chat” and “Variation” buttons are merged into a single “Add to Chat”, and deleting a clip now asks you to confirm, since it removes the file from disk.
• ●Better onboarding tutorial. The Raw / Clean / Enhance step is easier to read, the Library and Toolbar steps now explain the action buttons and the Bass / Harmony / Lead controls, and the Settings step covers custom system prompts, temperature, and Pro Mode reasoning.

Fixed

• ●Custom system prompt text now saves. Editing a custom system prompt's name or body and then closing Aether no longer discards what you typed; any in-progress edit is committed to disk on close and restored on the next launch.
• ●Grok's model list no longer shows the code-only and multi-agent models, which are not applicable to music generation.
• ●In System Prompt settings, clicking a prompt in the list now moves the highlight to match the prompt shown in the editor.
• ●Fixed the tutorial's Enhance step, where the navigation buttons could overlap the popover and the step appeared to show twice.

1.1.4

Fixed

• ●Some models (most often Claude routed through OpenRouter) could print a large block of raw JSON into the chat instead of the usual plain-language summary, which could also make the interface feel sluggish. Generation responses are now parsed cleanly no matter how the model formats them, so you only ever see the summary.
• ●The Pro Mode plan view (the “Thought” panel and the reasoning card) now shows the readable composition plan for every model. Previously, models that wrap their output in code fences (most often Claude routed through OpenRouter) could display a block of raw JSON here instead of the formatted plan.

1.1.3

Improved

• ●More reliable licensing. Activation and license checks now retry automatically through brief network hiccups, so an unstable connection is far less likely to cause a failed activation or a false “couldn't reach the server” message.

1.1.2

New

• ●Return to playhead. When you scroll the piano roll away from the playhead, a small “↩ Playhead” button appears so you can jump the view straight back.

Improved

• ●Smoother seek. After double-clicking to set a start point, pressing play now glides the view to the playhead instead of snapping to it instantly.

Fixed

• ●Auditioning a variation while a batch is still generating no longer cuts out when the next variation finishes.
• ●The piano-roll playhead now stays locked to what you're actually hearing after you scroll the view. Previously it could drift out of sync with the audio.

1.1.1

New

• ●Automatic update notifications. Aether now checks for new versions on launch and shows a small, dismissible banner when one is available, with a direct link to download. No more missing an update.

Improved

• ●Better Grok results. Reworked the composition prompting for Grok models so they produce denser, more complete parts instead of sparse or repetitive ones.

Fixed

• ●Fixed an error when using the reasoning-effort setting with Grok models that don't support it.
• ●Stopped occasional false “generation failed” messages that could appear after you cancelled or resent a request.

1.1.0

New

• ●Reasoning effort for Grok. A Low / Medium / High selector in Settings lets you trade speed for depth on Grok models that support it. It only appears when the selected model offers it.

Improved

• ●Cleaner Pro Mode. The plan and result now appear together in a single “Composition plan” card.
• ●Stronger Grok composition prompting for more musical, less generic output.

Fixed

• ●The reasoning card no longer clips its content when expanded.
• ●The reasoning card now appears in the right place: after your message, not before it.

1.0.1

New

• ●See the AI’s reasoning. Aether can now show the plan it’s working from as it composes, for both standard and Pro Mode generations. Turn it on or off with the new “Show AI reasoning” setting.
• ●Per-variation Pro Mode. When generating a batch, each variation now gets its own independent plan and composition, so the options feel more distinct.

Improved

• ●Smarter edits. Mentioning a part in an edit (e.g. “make the bass busier”) now focuses on that part instead of removing the others.
• ●Tidied the interface: fixed the Quick Params clear button icon and removed a setting that no longer did anything.
• ●Hardened privacy in the reasoning view so internal planning context stays internal.

Fixed

• ●Cancelling a generation now stops cleanly, with no leftover “Planning…” spinner.
• ●Pro Mode now saves your actual prompt to history instead of its internal planning text.
• ●Windows: resolved an installer hang on reinstall, made the background helper exit automatically when your DAW closes, fixed a port conflict that could block AI requests, and corrected a packaging issue with the bundled piano instrument.

1.0.0

Highlights

• ●Generate bass, harmony, and lead parts from natural-language descriptions.
• ●Choose your AI provider (ChatGPT, Claude, Grok, Gemini, and more) with your own API key.
• ●Chat-based workflow: ask for changes, regenerate, and refine in context.
• ●Pro Mode for a plan-then-compose approach to more deliberate results.
• ●Built-in piano-roll preview and one-drag export straight into your DAW.
• ●Save and reuse your favorite generations from the library.