Changelog

What shipped, in order

Every release, what it actually does, no adjectives doing heavy lifting.

  1. Update The Studio plugin is on the Creator Store

    2026-09-22

    Apple Studio, the plugin Apple builds through, can be installed from the Roblox Creator Store again, and it is free. Install the plugin walks through getting it, enabling it, pairing it and allowing edits.

    The public site

    • The site and the app now share one look, so moving between them feels like one product.
    • Two modes, Plan and Agent, with Autonomous as a toggle on Agent. Apple and Apple MAX are models, not plans.

  2. v0.2 One model, measured

    2026-08-30

    Every mode moved onto a single model, chosen by the eval suite rather than by taste — and the migration made Apple both better and cheaper at once.

    The engine

    • One engine now runs planning, building, memory and vision — 1M context, native tool calling, multimodal. Which engine that is, is an implementation detail Apple chooses per step; you never pick it and it can change without you noticing.
    • Scored 98.9 on the Roblox eval suite against 96.8 for the model it replaced.
    • A full Agent build fell from $0.0139 to $0.0056 — 2.5× cheaper for the same work, on a model that scores higher.
    • Multi-model routing was removed. It existed to move cheap steps onto a smaller model; measurement showed that routing the post-inspection step changed the outcome for the worse — the small model read the project tree and declared the task finished instead of building it.

    Spend control

    • The gateway now bills from the engine's own reported figure per call rather than from an estimate.
    • Credits recalibrated: 1 Credit = 30 neurons, so a Credit stays a meaningful unit now that the same work costs less.
    • Four independent gates, each proven against production: a per-request ceiling, a daily ceiling, a monthly billable cap and an instant kill switch — all adjustable at runtime with no redeploy.
    • Automatic retries removed entirely. A retry is a second bill for the same work.

    Leaner context

    • Tool results capped at 3,000 characters, agent transcript at 24,000.
    • Tool definitions filtered per mode, so Plan does not pay to be told about playtesting tools.
    • Run budgeting and metering were tightened while the agent loop was being stabilized.

  3. v0.1 First public release

    2026

    Apple opens its doors. The full describe-to-build loop: a web workspace, a real Studio plugin, and an agent that changes your actual place — carefully.

    The workspace

    • Chat-driven building with live streaming responses and a visible tool trace — you see every get_project_tree, edit_script and run_and_check as it happens. Since narrowed: the independent Studio plugin never executes code it is sent, so run_and_check and the other tools that would need it are not offered while it is the plugin you have paired. The trace shows what actually ran. See the plugin docs for the current refusals.
    • The early work-mode controls were simplified as the product matured. Today there is nothing to choose: describe what you want, and Apple shows a short plan and builds it. See Getting started.
    • Unlimited projects per account, each with its own chat history and memory of what you built.
    • Image attachments — paste a screenshot straight into the chat.
    • A stop button that actually stops the agent mid-run.

    The Studio plugin

    • Pairing by short-lived code (GLM-XXXXXX, 10-minute expiry) — no tokens to copy around. Since changed: codes are six characters with no prefix, and the expiry is shown counting down in the pairing dialog.
    • Reads and searches your scripts, edits them through Studio's own script editor APIs.
    • Creates, moves, configures and deletes instances with typed, validated operations.
    • Wraps every operation in its own ChangeHistory waypoint — native Ctrl+Z undo, one step per change rather than one per run.
    • Starts and stops run mode, captures logs, and reports results back for verification. Current implementation uses bounded Run-mode controls for Agent verification and can capture periodic viewport snapshots when Studio grants screenshot permission.

    Safety net

    • Automatic checkpoint before every agent run; manual checkpoints on demand.
    • One-click restore of any checkpoint — scripts and instances — even across sessions.

    The deal

    • Free tier: 60 Credits per day, hard quota, resets daily, no card ever collected. Since repriced: the free tier is 231 Credits a day. A reader comparing this page with Pricing was seeing two different numbers for the same thing.
    • Pro (400 Credits/day, priority queue, extended checkpoints) opens as a waitlist. Since withdrawn: there is no Pro tier. The paid tiers are Builder and Studio — see Pricing.
    • Private projects are never used to train models.

Apple is in active development — expect the next entry sooner than you think.