This page is the authoritative changelog. A standalone version also lives on the
marketing site. Every surface is green when you check status.
July 2026
🔐 Vision grounding proxies locked down. The grounding proxies now require a shared secret or a first-party origin and enforce per-IP caps. Third parties can no longer ride your vision quota through an open proxy.
🏢 App management and self-serve onboarding. Update an app’s name, default mode, and embedding origins with PATCH /v1/apps/, and see per-app quota overrides as effective caps in GET /v1/usage. One-command client onboarding and secret-key rotation round out the developer dashboard.
💳 Runtime-only billing. Developer billing now meters runtime only: creating and editing worlds doesn’t burn paid credit. Pricing and quota docs were rewritten to say exactly what counts and what’s free.
📚 New guides across the docs. Fresh guides for vision and interactive objects, world-generation examples, the Studio, and version rename/prune, with the OpenAPI spec synced to every new route.
👁️ Vision grounding lands in /v1. Ground objects, describe scenes, and run perception checks against live world frames straight from the API. Vision calls meter under their own quota kinds so you can budget them separately from generations.
🌿 Rename and prune world versions. The version tree gets full CRUD: PATCH retitles a snapshot, DELETE prunes a node and re-roots its children onto its parent so the tree never fragments, and listing returns newest-first. HEAD is protected: you can’t prune the version you’re on.
🎨 Creator Studio in the browser. A standalone Creator Studio SPA ships at /studio: author, edit, and publish playable worlds on top of the same /v1 API you integrate with, no local tooling required.
🛡️ /v1 hardened after adversarial audit. A red-team pass against the public API produced a round of fixes (tenant scoping, input validation, and abuse-path closures), now live, with the OpenAPI spec and hardening results published alongside.
June 2026
💬 Read character transcripts. GET /v1/characters//transcript returns a player’s full chat history with a character (secret-key auth). Cross-session memory got a reliability pass too. Characters now recognize returning players and welcome them back.
🎭 The Characters API. Create a persistent character, talk to it, give it a voice and a live video presence, and embed it in your product, with memory that survives across sessions.
🧩 Graph editing and the embed runtime. Program a world’s logic over REST: state and event CRUD, batched ops, lint/validate, natural-language edits via /edit, and git-style version snapshots. The browser SDK lands as @alakazamworld/embed.
🚀 The programmable-worlds API launches. POST a premise or a seed frame to /v1/worlds and get back a versioned, playable world, then mint short-lived session tokens so your users play it embedded in your product, branded as your own.
📖 Book to Game. Paste a story you wrote and get back a branching, Bandersnatch-style playable campaign that follows your text’s real spine, surfaced directly in Creator Studio.
⏱️ Metered chat time and Plus. Character chat is now metered: a 3-minute free taste for everyone, 2 hours per month on Plus, and credit overage beyond that. The account page gains a credits wallet with top-up packs.
🗣️ Live-video conversation characters. Talk to a character rendered on a live world-model video stream: voiced replies, stance-aware body language, and drift-refresh so the stream never degrades mid-conversation.
🤖 Chat-to-edit worlds and version trees. The Stagecraft editor gains a kernel chat agent that turns natural language into validated graph edits, plus a branching version tree with per-field diffs.
May 2026
🚨 Platform lockdown pass. Server-trusted positions for anti-cheat, request rate limits, a tightened CORS allowlist, and row-level-security lockdown across the database.
🌍 Multi-region GPU streaming. World streaming deploys across multiple GPU regions with a timezone-routed region picker, cutting latency for far-away players. Wake-on-demand spins up idle capacity only when someone plays.
April 2026
☁️ Server-side GPU inference for every game. World-model inference moves off the player’s device onto multiplexed cloud GPUs, so low-end hardware plays the same worlds. A new analytics pipeline starts tracking sessions and per-game events.
🧙 Narrator-programmable worlds. The narrator can mutate a running world: prose becomes worldOps that live-edit geometry, items, and story beats mid-session, the pattern that grew into the /v1 API.