New AI Models

Claude Fable 5.1

Claude Fable 5.1 (claude-fable-5-1) is Anthropic's successor to Claude Fable 5 for long-running agentic coding, knowledge work, and research. It pairs a 1M token context window and 128k max output tokens with always-on adaptive thinking at $10/$50 per MTok, with cache reads cut to $0.25 per MTok. Improvements over Fable 5 center on multi-hour agentic coding sessions, multistep research, vision, long-context reasoning, and computer use, widening at higher effort. It suits demanding reasoning and long-horizon agentic work on the Claude API, AWS, Google Cloud, and Microsoft Foundry.

Claude Mythos 5.1

Claude Mythos 5.1 (claude-mythos-5-1) is Anthropic's gated model for Project Glasswing participants, launched on September 1, 2026 alongside Claude Fable 5.1 and continuing the defensive-cybersecurity line that began with Mythos Preview. It defaults to a 1M token context window, 128k max output tokens, and always-on adaptive thinking, priced at $10/$50 per MTok like Claude Fable 5, with prompt cache reads cut to $0.25. Access stays invitation-only, and its output text carries Anthropic's text watermark.

Muse Voice Transcribe

Muse Voice Transcribe is Meta Superintelligence Labs' first real-time audio perception model, delivering streaming ASR, speaker diarization for 20+ speakers, and endpointing. Built on the Muse Spark family with 80ms audio chunks and RL-trained adaptive delay, it supports 70+ languages (25 validated at launch), native code-switching, keyword and context biasing, and long audio beyond one hour. It ranks first on Artificial Analysis streaming speech-to-text at 3.1% WER and leads public diarization benchmarks at 17.5% DER. Available via Meta Model API, Meta AI for Mac, and Muse Code.

Qwen-Audio-3.1-ASR-Flash-Message

Qwen-Audio-3.1-ASR-Flash-Message is the voice-message transcription member of Qwen's 3.1-generation ASR family (September 2026), served over a WebSocket realtime endpoint. It applies voice activity detection and, once the speaker pauses, returns each utterance as one complete punctuated transcript without streaming partial results, a mode suited to chat voice notes and asynchronous voice input. Recognition is multilingual with automatic language detection, leaving live captions to its streaming sibling.

Qwen-Audio-3.1-TTS-Flash

Qwen-Audio-3.1-TTS-Flash is a high-performance speech synthesis model built for real-time interaction, supporting multiple languages and Chinese dialects. It provides free-style instruction following and fine-grained label control over emotion, tone, role, speaking rate and volume, and its voice cloning stays robust under complex acoustic conditions such as noise and reverberation. As the Flash tier it targets low-latency scenarios like voice assistants, live conversation and intelligent customer service, priced from CNY 1.5 per million input tokens and CNY 12 per million output tokens.

Qwen-Audio-3.1-TTS-Next

Qwen-Audio-3.1-TTS-Next is an Audiogen model from the Qwen team built for unified audio generation, going beyond conventional TTS that handles speech synthesis alone. From inputs such as text, timestamps, and reference audio (up to 3 clips), it produces complete audio in a single pass, combining speech, sound effects, and ambient sounds on one timeline with flexible timing control. It supports multilingual TTS, single-speaker speech, multi-speaker dialogue, podcasts, cinematic soundscapes, ambient audio, and sound effects, outputting WAV, MP3, or PCM files of up to 240 seconds for podcasts and 120 seconds for other scenes. This makes it a fit for content creation, short-form video, podcast production, film and television work, and game audio.

Qwen3.7-Text-Embedding-Flash

A lightweight multilingual text embedding model trained on Qwen3.7 and available on Alibaba Cloud Model Studio for cost and throughput sensitive workloads. Compared with text-embedding-v4, it expands language coverage from about 100 to 201 languages, extends long text input to 128K (131,072) tokens, and adds Sparse Embedding, with output vector dimensions from 256 to 1024 (1024 by default). Alibaba states that its retrieval quality stays on par with v4 across semantic, cross-lingual and code retrieval benchmarks, with about 3% gain on AIRBench-zh, making it suited to large-scale text search and RAG.

Qwen3.7-Text-Rerank

Qwen3.7-Text-Rerank is a multilingual general-purpose text reranking model trained on the Qwen3.7 base model and served through Alibaba Cloud Model Studio. Compared with qwen3-rerank, it delivers large gains across general and web search, code retrieval, instruction following, and agentic and memory retrieval, with official benchmarks showing roughly 35% relative improvement on websearch and better results on MTEB. It accepts up to 32768 input tokens, making it suitable for relevance reranking of candidate documents in multilingual retrieval and RAG pipelines.

Solaris

Runway Solaris is a research release and the first model in Runway's Interface World Models family, a real-time system that generates the interface itself, frame by frame, replacing pre-coded apps. Built on Gen-4.5, it treats user input as conditioning for the next frame, like text or images, pairing autoregressive generation with distilled few-step denoising to sustain interactive 720p sessions. In a 250-person study it was preferred over Claude-coded interfaces 61% to 24% on instruction following and 71% to 21% on natural behavior, for uses like virtual try-on and agent training.

hy4-preview

Hy4 preview is Tencent Hunyuan's new flagship open-source LLM built on a Mixture-of-Experts (MoE) architecture, with 770B total parameters, 49B activated per token, and a 1M-token context window. A native MTP layer enables speculative decoding for faster inference. Tuned for real productivity work in software engineering, office and data analysis, game development, and scientific research, it edged out GLM 5.3 and Kimi K3 in an internal blind evaluation across 203 engineering tasks. Weights ship under Apache 2.0, including an FP8 variant, and serve via vLLM or SGLang.

Qwen-MT-Image-2.0

Qwen-MT-Image-2.0 is an image translation model service on Alibaba Cloud Model Studio that translates pictures in Chinese, English, Japanese and 55 languages in total into a target language while preserving the original layout and content. It supports custom options such as term definitions, sensitive word filtering and product subject detection, providing flexible and accurate image localization for multilingual visual content.

Gemini Omni 1.1 Flash

Gemini Omni 1.1 Flash is Google DeepMind's generative video model, available in production through the Gemini API in Google AI Studio. Scene extension draws on up to 10 seconds of prior context, extends clips in 10-second increments to 40 seconds total, and accepts up to three seconds of video as reference input for visual consistency. Its 360p draft mode renders up to 60% faster at about a third of the cost of standard 720p, with 1080p and 4K upscaling available.

MiniMax H3 Max

MiniMax H3 Max is a speed-optimized video generation model post-trained by fal Research from MiniMax H3 and launched on MiniMax official channels, supporting text-to-video and image-to-video with 480P/768P output from 5 to 15 seconds. fal reports it renders a 5-second clip in under 3 seconds, at roughly 35x the throughput of the official MiniMax H3 endpoint, and ranks #1 in overall quality, prompt understanding and aesthetics in its human preference evaluations among leading video models. It is suited for rapid iteration on short-form creative and marketing videos.

V8.2 Edit Model

Released on August 27, 2026 for open community testing, this is Midjourney's first V8.2 image edit model. It edits images from text instructions, generates new images from up to 4 image references at once (replacing omni-reference), and supports inpainting and outpainting for local changes and canvas expansion. Personalization, moodboards, and srefs are also supported. It is available on midjourney.com and alpha.midjourney.com, or via the --edit command on Discord.

V8.2 image edit model

Midjourney's first V8.2 image edit model is now open for everyone to test. It edits images from text instructions, generates new images from up to four references (replacing omni-reference), and supports inpainting and outpainting, along with personalization, moodboards, and srefs. You can use it by attaching an image to the prompt bar, clicking edit in the lightbox, or typing --edit on Discord.

Qwen-Drive-1.0-4B

Qwen-Drive-1.0-4B is an open source vision-language foundation model for autonomous driving, released by Qwen under Apache 2.0. It retains the pretrained Qwen3.5-4B VLM architecture unchanged and adds two external modules: a BEV perception head jointly performing 3D object detection, semantic occupancy prediction and BEV map segmentation, plus a flow-matching Planning Expert (planner-sft and planner-rl variants) that generates future ego trajectories. The shared VLM still answers free-form questions about driving scenes, unifying 3D perception, driving VQA and motion planning in one research-ready framework.

GLM-5.3-Flash

GLM-5.3-Flash is the first natively multimodal model in Zhipu AI's GLM-5 series, with 320B total and 18B activated parameters, a hybrid sparse-plus-linear attention architecture, and a context window of up to 1M tokens. It outperforms GLM-5.2 across coding and agentic benchmarks at one-tenth the price and approaches Claude Opus 4.8, scoring 57 on the Artificial Analysis Intelligence Index at $0.045 per task. Suited to high-volume coding, agent workflows, and visual document understanding, its weights are available on Hugging Face.

Gemini 3.5 Transcribe

Gemini 3.5 Transcribe is Google's most precise speech-to-text model, built for intelligent voice interactions. It turns raw audio directly into polished, formatted text, with two API modes: real-time streaming with sub-second latency via the Live API, and pre-recorded transcription with speaker attribution and word-level timestamps. It auto-detects over 85 languages and averages 4.0% WER in streaming and 2.6% in non-streaming use cases. It is available in public preview through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform.

Qwen3.8-Flash-Next

Qwen3.8-Flash-Next is an experimental, open-weight preview of the architecture that will underpin Qwen4. It is a vision-capable causal LM with 125B total parameters and 6B activated, plus 51B n-gram embeddings and 4B MTP, pairing Gated DeltaNet with Qwen Sparse Attention (QSA) alongside Gated Residual and MoE, with native 262K context extensible to 1M tokens. As the first open-weight release of this architecture, it lets the community adapt tooling early and gather feedback for Qwen4.

Qwen3.8-Flash

Qwen3.8-Flash is a multimodal LLM from Alibaba's Qwen team that pairs strong understanding and generation capability with fast response times. It natively supports a million-token context window, handling long documents, code repositories, and complex conversations in a single pass. The model targets coding assistance, agent workflows, and vision-language tasks, works with OpenAI and Anthropic API protocols in tools such as Claude Code and Codex, and offers competitive inference costs for high-concurrency applications.