Changelog

Text models32

• GPT-6.1 Sol

  GPT-6.1 coding and agent text model from OpenAI delivering near-Astra performance at lower cost

  All usersAnonymous

• MiMo-V2.6-Flash

  Omnimodal sparse Mixture-of-Experts text model from Xiaomi with a 1M-token context window

  All usersPrivate

• Claude Opus 5.5 Fast

  Faster-output variant of Claude Opus 5.5 from Anthropic with a 1M token context window

  All usersAnonymous

• Claude Sonnet 5.5

  Everyday agentic text model from Anthropic, a step-change improvement over Sonnet 5

  All usersAnonymous

• Aion 3.5

  Multi-model roleplaying and storytelling text system from AionLabs, built on the GLM family

  All usersAnonymous

• Aion 3.5 Mini

  Compact multi-model roleplaying and storytelling text system from AionLabs, built on the GLM family

  All usersAnonymous
• Show 26 more

Image models3

• GPT Image 2.5 Flare

  Image generation variant of GPT Image 2.5 from OpenAI

  All usersAnonymous

• GPT Image 2.5 Sunburst

  Image generation variant of GPT Image 2.5 from OpenAI

  All usersAnonymous

• Muse Image

  Image generation model

  All usersAnonymous

Video models9

• Seedance 2.5 US

  Text-to-video and image-to-video generation model from ByteDance

  All usersPrivate

• Seedance 2.5 US R2V

  Reference-to-video variant of Seedance 2.5 US, enabling image-guided video generation

  All usersPrivate

• MiniMax H3 Max Multi-Angle

  Multi-angle video generation model from MiniMax

  All usersPrivate

• Wan 3.0 Prime Pro

  Image-to-video variant of Wan 3.0 Prime Pro from Alibaba Cloud

  All usersAnonymous

• Wan 3.0 Prime Pro Reference

  Reference-to-video variant of Wan 3.0 Prime Pro from Alibaba Cloud

  All usersAnonymous

• Wan 3.0 Pro

  Image-to-video variant of Wan 3.0 Pro from Alibaba Cloud

  All usersAnonymous
• Show 3 more

Audio models4

• ElevenLabs TTS v4

  Expressive text-to-speech model from ElevenLabs with audio tag support

  All usersAnonymous

• ElevenLabs TTS v4 Turbo

  Low-latency text-to-speech model from ElevenLabs with audio tag support

  All usersAnonymous

• ElevenLabs Music v2

  Realistic music generation model from ElevenLabs with optional vocals and configurable duration

  All usersAnonymous

• ElevenLabs Music v2.5

  Latest realistic music generation model from ElevenLabs with optional vocals and configurable duration

  All usersAnonymous

Deprecated9

• Gemma 3 27B→replaced byGemma 4 26B A4B Uncensored
• GLM 5.1→replaced byGLM 5.2
• GPT OSS 20B→replaced byGPT OSS 120B
• GPT-5.2 Codex→replaced byGPT-5.3 Codex
• Mistral Small 4→replaced byMistral Small 3.2 24B Instruct
• Qwen3.6 35B A3B Uncensored→replaced byGemma 4 26B A4B Uncensored
• Sora 2 Pro→replaced byMiniMax H3 Max
• Wan 2.1 Pro→replaced byWan 3.0 Pro
• Wan 2.2 A14B→replaced byWan 3.0

Chat40

• Free With Pro Model Tag

  Eligible users in the free-models experiment now see a 'Free with Pro' tag on models made free for paid tiers in the model switcher.

• Auto Mode Default Model Test

  Auto mode can now start chat and route images on a different default model for users in the default-model experiment.

• Mind Assistant Model Selector

  Mind Assistant now has a model selector, and the chosen model carries over between Agentic Chat and Mind Assistant.

• Open Shared Mobile Chats on Web

  Chats shared from mobile via an encrypted link can now be opened in the web app at a dedicated shared-chat route.

• More Models For Everyone

  DeepSeek V4.1 Flash, GLM 5.3 Flash, GLM 5.3, Kimi K3, the end-to-end encrypted Kimi K3, Claude Sonnet 5, Claude Opus 5, Grok 4.6, and Fable 5.1 are now available to all users in Minds instead of alpha testers only.

  Beta

• More Voice Agent Voices

  Realtime voice conversations now accept the full set of voices from the XAI TTS model instead of a hardcoded list of five.

• Show 34 more

Image, video and audio28

• Smarter Edit Template Refinement

  Edit template refinement now analyzes uploaded images to pick a subject-appropriate setting, translates casual customizations into precise visual language while dropping conflicting template instructions, and prioritizes user fields over the literal template.

• Wan 3.0 Pro Promo

  An announcement card promoting 30% off Wan 3.0 Pro appears in the app and links to Video Studio with the model preselected, and the discount on Wan 3.0 Pro and Wan 3.0 Prime Pro is extended through the end of October 7 PT.

• Lyria 3.5 For Everyone

  Lyria 3.5 music generation is now available in Studio for anon, free, and pro users.

• Music Uncensored Tag Removed

  ElevenLabs Music and ACE-Step 1.5 no longer carry the uncensored badge and become selectable in store builds and by the autorouter.

• Trend Aspect Ratio Picker

  Trend runs now support an aspect ratio picker in the composer and an optional aspect ratio parameter, letting users reframe outputs as vertical or square on engines that support multiple shapes.

• Trend Photos Saved To Library

  Photos uploaded or taken for trends are now stored locally and saved to the Studio inputs library for reuse across trends.

• Show 22 more

Library and media3

• Minds Monetization Disabled

  The Minds creator monetization UI, including the pricing tab, earnings page, and owner earnings strip, is now hidden for all users.

• Persisted Minds Rating Filters

  Selected Minds maturity rating filters are now saved to user preferences and restored on each visit instead of resetting.

• Translated Mind Rejection Reasons

  The mind publish popover now shows translated rejection reasons when a mind submission is rejected.

Account and settings5

• Thumbnails Included In Backups

  Local backups now include cached image thumbnails so restored conversations keep their preview images.

• Anonymous Email Relay Sender

  The anonymous email relay now sends through an additional SendGrid sender for better delivery.

• Auto-Refreshing API Usage

  The API usage dashboard now polls for fresh usage data every two minutes.

  Pro users

• Telemetry Privacy Boundary

  Users with telemetry disabled now still load PostHog for tier-based feature flags, though event capture stays off.

• Edit Images Onboarding Removed

  New and upgrading users no longer see the Edit Images step in the onboarding modal.

Wallet and payments7

• Annual Pricing Default

  The upgrade modal now defaults to yearly billing and shows annual tiers as a monthly equivalent price with billed-annually wording.

• Free Models For Paid Tiers

  Listed credit-billed chat and image models are now free for paid-tier users as part of a running experiment.

  Pro users

• Downgrade Credit Loss Warning

  A confirmation now warns users with banked bundled credits that downgrading their subscription will forfeit unused credits before the plan change proceeds.

• Crypto 3 and 6 Month Plans

  Pro plans can now be purchased with crypto on 3- and 6-month terms, and all payment methods share a unified checkout screen with a single confirm step.

• Sales Tax At Checkout

  Subscription and credit checkout now display Stripe-calculated sales tax and authoritative totals, including in upgrade and scheduled-change breakdowns.

• Payment Method On Subscription Badge

  The Stripe subscription badge now shows whether the cycle was paid with Credit Card or USDC, and hides when the cycle is unpaid.

  Pro users
• Show 1 more

Mobile app11

• Model Explorer Feature Filters

  Model Explorer filters now include a Pay Per Use pricing option and a Features section for Reasoning, Vision, Speed, and Uncensored.

• Image Editing Workspace V2

  A new mobile editing workspace handles Edit, Combine, Crop, Doodle, Remove Background, and Upscale with persistent drafts and improved controls.

• Agentic Video Particle Loader

  Agentic video generation now shows an animated particle loader while work is in progress.

• iOS Background Media Generation

  Image and video generation can now continue in the background on iOS.

• Large Backup And Restore

  Backups now support up to 500 MiB of text and 2 GiB of media with crash-safe, memory-bounded restore.

• Copyable Usage History IDs

  Usage history now shows the task/request reference ID beneath the model name, with a tap to copy the full ID.

• Show 5 more

• Higher Partner Tier 1 Rate Limits

  Partner Tier 1 rate limits doubled for DeepSeek V4.1 Flash, Gemma 4 31B, and Qwen 3.5 9B.

  Pro users

• Astra Reasoning Replay Fix

  Follow-up turns on the Responses API for GPT-6 Astra no longer fail with a 400 when replaying stored reasoning items.

• Fable Tool Choice Fix

  Forced tool_choice requests to Claude Fable 5.1 are coerced to auto instead of failing with a 400 error.

  Pro users

• Featured Model API Fields

  Agent model configuration endpoints now return isFeatured and capabilityTags so clients can highlight models and show capability subtitles.

  Pro users

• Reasoning Disable Switch

  The reasoning.enabled parameter is now honored on /chat/completions and /responses, so setting it to false actually disables reasoning.

• Responses API Accepts Image URL Strings

  The Responses endpoint now accepts standard OpenAI image_url string inputs alongside object form, with image detail preserved through translation.

• Show 39 more

• Draft Submission Fix

  Submitting a Mind message right after the app loads no longer silently drops the draft while local storage is still initializing.

  Chat

• Unified Database Worker

  Chat data now loads through a single shared web worker, fixing duplicate worker instances and Safari out-of-memory crashes.

  Chat

• Pre-Relay Balance Check

  Text inference requests from accounts below the minimum credit balance now fail upfront with a 402 balance_exhausted error instead of breaking mid-stream or relaying without charge.

  Pro usersAPI

• Share Link Key Leak Fixed

  Share link decryption keys are no longer sent to analytics in flags requests or cookies.

  Chat

• Claude Reasoning Round-Trip Fixes

  The chat completions API now preserves Claude thinking blocks and signatures across tool-use loops, and streamed replay no longer lets readable summaries contaminate encrypted reasoning items while reasoning-only chunks are stripped correctly when thinking responses are disabled.

  API

• Accurate Video Duration Quotes

  Chat now quotes upscale and video-input models using the attached video's actual duration instead of falling back to the model default, which could show roughly double the correct credit cost.

  Chat
• Show 65 more