What's new

Every door we've opened lately

New models, sharper workflows, and product improvements across magicdoor.ai.

Latest update May 12, 202624 public updatesRSS feed
Performance

Installed Chat opens faster

We reduced the work that blocks the first magicdoor.ai/chat shell, so installed and bookmarked launches can show the chat surface sooner on mobile.

Permalink
  • Keeps auth and subscription guardrails in place before Chat opens
  • Loads preferences, balance, renewal details, and sidebar history after the shell starts rendering
  • Improves standalone launch polish with matching theme colors and safe-area padding
Model update

Homepage model labels refreshed

We refreshed the main magicdoor.ai homepage marquee so it shows the current GPT-5.5 and Claude Opus 4.7 labels instead of older model names.

Permalink
  • Updates the homepage marquee from GPT-5.1 to GPT-5.5
  • Updates the Claude Opus label from 4.6 to 4.7
  • Keeps the public landing page aligned with the current model registry
Product

Find the right AI model before you start

A new public /which-ai-model page helps visitors choose a practical model for writing, coding, research, images, or everyday chat before signing up.

Permalink
  • Adds a four-question model recommender with one primary recommendation and alternatives
  • Uses current magicdoor.ai model labels, capabilities, and usage pricing for cost posture
  • Keeps the flow public and separate from future first-login onboarding
Model update

Kimi upgraded to K2.6

We updated the Moonshot AI chat model in magicdoor.ai to Kimi K2.6, adding the new Together AI endpoint and refreshed multimodal coding metadata.

Permalink
  • Routes new Kimi chats to moonshotai/Kimi-K2.6
  • Keeps older saved Kimi K2.5 chats resolving to the new K2.6 endpoint
  • Refreshes model labels, pricing, and marketing copy to show Kimi K2.6
Product

Chats now know their active model

Magicdoor now includes the active model context in each chat generation prompt, so assistants can answer model-identity questions based on the model currently handling the message.

Permalink
  • Adds the active model ID to chat system context for each generation
  • Includes model label and provider details when available
  • Keeps assistant instructions, memory, and project instructions intact
Product

Install magicdoor.ai from your phone

A new public /install page explains how to open magicdoor.ai/chat first, then add it to your home screen for faster access to Chat.

Permalink
  • Shows a dismissible mobile banner on Chat and Images that links to the install guide
  • Directs mobile users to start from magicdoor.ai/chat before using Add to Home Screen
  • Includes iPhone Safari and Android Chrome install steps
  • Keeps the install guide publicly reachable at magicdoor.ai/install
Model update

Grok upgraded to Grok 4.3

We updated the xAI chat model in magicdoor.ai to Grok 4.3 while keeping the existing pricing and capability metadata in place.

Permalink
  • Routes new Grok chats to the grok-4.3 model ID
  • Keeps existing saved Grok chats resolving to the new Grok 4.3 endpoint
  • Refreshes model pages, resources, and product copy to show Grok 4.3
Model update

GPT-5.5 now leads the chat picker

We reordered the chat model picker so GPT-5.5 is the first default option for new chats, followed by Claude Sonnet 4.6 and GLM-5.1.

Permalink
  • Makes GPT-5.5 the first model new and trial users encounter
  • Keeps Claude Sonnet 4.6 as the broadly available Claude option near the top
  • Keeps GLM-5.1 available directly below Claude Sonnet 4.6
Trust

Chat internals now use direct model modules

We cleaned up an old chat compatibility layer so model metadata, pricing, markdown rendering, and chat creation now import from their source modules directly.

Permalink
  • Removes a stale chat utility barrel file
  • Keeps chat cost and model capability helpers wired to their source modules
  • Aligns usage cost typing with image usage records
Performance

Chat compaction is now server-owned

Manual chat compaction now uses persisted server history, versioned snapshots, and structured artifact metadata so long chats keep safer context.

Permalink
  • Stores v2 compaction snapshots with summary previews, token estimates, quality metadata, and artifact manifests
  • Computes manual compaction ranges on the server instead of trusting client split points
  • Adds the shared stream event contract needed for future auto-compaction updates
Product

GPT-5.5 chat tools now finish reliably

GPT-5.5 chats now keep running after OpenAI tool results so Magicdoor tools and code interpreter sessions can complete with a final answer.

Permalink
  • Restores multi-step GPT-5.5 tool calls for search, memory, canvas, and other chat tools
  • Keeps code interpreter progress and generated-file capture deterministic
  • Persists code interpreter containers even when no generated file is cited
Product

Chat onboarding now shows models and images together

New users now get a clearer chat onboarding flow that highlights model switching first, then includes one complimentary Seedream 4.5 image after they try a second model.

Permalink
  • Guides new users to compare models in the same chat
  • Adds one complimentary Seedream 4.5 onboarding image
  • Keeps future image generation clearly billed per image
Model update

GLM-5.1 now leads the chat picker

We reordered the chat model picker so GLM-5.1 is the first default option for new chats, with Gemini 3.1 Pro and Claude Haiku moved lower.

Permalink
  • Makes GLM-5.1 the first model new and trial users encounter
  • Keeps Gemini 3.1 Pro and Claude Haiku available lower in the picker
  • Preserves MiniMax M2.7 and legacy M2.5 routing
Model update

MiniMax upgraded to M2.7

We replaced the deprecated MiniMax M2.5 endpoint with MiniMax M2.7 across magicdoor.ai after the provider sunset notice.

Permalink
  • Routes new MiniMax chats to MiniMaxAI/MiniMax-M2.7
  • Keeps older saved M2.5 chats resolving to the new M2.7 endpoint
  • Refreshes model labels and resource copy to show MiniMax M2.7
Performance

Long chats now reuse cached prompt context

Prompt caching is now enabled for Claude and GPT chats, lowering cost and latency when longer conversations reuse the same context.

Permalink
  • Claude system context is marked cacheable for supported long chats
  • GPT chats use account-scoped prompt cache routing
  • Cached token reads and writes are tracked separately for accurate billing
Workflow

Compare image models from one picker

The /images flow now lets you choose one model for a normal generation or select multiple compatible models to compare outputs side by side.

Permalink
  • Model rows generate with a single model, while checkboxes add models for comparison
  • Aspect ratios stay compatible with the selected model set
  • Compare results show model-labeled image tiles with per-model loading and error states
Model update

ChatGPT Image upgraded to gpt-image-2

We upgraded the ChatGPT Image path in magicdoor.ai to Replicate gpt-image-2 and aligned the images flow around the newer model.

Permalink
  • Updated the main image route and library plumbing for the upgrade
  • Refreshed image model labels across the product
  • Added route and image utility test coverage for the change
Workflow

File context is now visible above the composer

Chats now show which files and images are currently in model context, with detach controls when you want to narrow the prompt.

Permalink
  • Historical attachments stay visible in the context bar
  • Historical image thumbnails now render cleanly
  • Non-vision TogetherAI models safely strip file parts before requests
Workflow

Image editing mode landed in /images

The /images page gained a dedicated editing mode for uploaded images, plus signed uploads and library updates to support the flow cleanly.

Permalink
  • Edit existing images from the main images workflow
  • Signed URL support added for image uploads
  • Image form and library UI updated for the new path
Trust

Privacy and security got a public home

We launched a public trust page covering data handling, account-scoped conversations, and product privacy details in plain language.

Permalink
  • Public /trust landing page is now live
  • Canonical metadata and structured data added
  • Middleware updated so the page stays publicly reachable
Performance

PWA startup got faster on repeat visits

We shipped cold-start quick wins for the PWA, including better app-shell caching, deferred scripts, and lighter initial chat loading.

Permalink
  • Improved service worker caching for /chat
  • Deferred the markdown fix script until after interactive
  • Dynamic imports reduced the upfront chat payload

Ready to try what's new?

The latest models and workflows are already inside magicdoor.ai.

Open chat, compare outputs, or jump into images without stacking more subscriptions.

Get started for free