New AI Models

Muse Wonder

Muse Wonder is Meta's frontier model for personal and professional AI agents, released alongside Muse Wonder Full and trained at Meta's gigawatt-scale Prometheus data center. Its Active Memory carries important context across conversations and actively retrieves task-relevant details, while Active Focus sustains progress on long-running goals by trying alternative approaches, working through blockers, and checking in before important actions. Meta reports state-of-the-art results on agent, memory, and coding benchmarks including the Artificial Analysis Intelligence Index, AutomationBench-AA, MemoryBench, SWE-Marathon, and DeepSWE, with domain evaluations spanning law, finance, and health. The model is available through the Meta Model API and Muse Code, and will power Muse consumer experiences to follow, first for subscribers.

Muse Wonder Full

Muse Wonder Full is the top tier of Meta's Muse Wonder series, frontier LLMs for personal and professional agents that Meta reports as state-of-the-art on agentic, memory, and coding benchmarks such as MemoryBench, SWE-Marathon, and DeepSWE. The model pairs Active Memory, which carries key context about the user across conversations and retrieves task-relevant details, with Active Focus, which sustains long-horizon work by trying alternative approaches, working through blockers, and adapting to new information. Built on a larger and stronger base model than Muse Spark 1.3 and trained at Meta's gigawatt-scale Prometheus data center, it brings domain expertise in law, finance, and health, frontier coding performance, and better-than-predicted results on out-of-distribution tasks as test-time compute scales. Muse Wonder Full is available exclusively to Muse Code subscribers and ships with Meta's strongest safeguards to date against CBRN and cyber misuse, with careful handling of private information.

GLM-Legal

GLM-Legal is a legal-domain model from Zhipu AI built on the lightweight GLM-5.3-Flash foundation, with expanded legal corpora in mid-training and reinforcement-learning post-training across tens of thousands of simulated legal environments. It scores 19.2% on the Harvey Legal Agent Benchmark, a 12.5 percentage point gain over its base model that ranks first among open-weight models in Vals AI's independent evaluation, and it approaches larger closed-source frontier models on LegalBench and PLawBench despite a leaner parameter footprint. A Harness runtime orchestration framework coordinates multi-step agentic legal work such as consultation, case analysis, and document drafting, with these capabilities also arriving in the AutoClaw desktop platform.

Qwen-Image-2.1-Turbo

Qwen-Image-2.1-Turbo is the accelerated checkpoint of Qwen-Image-2.1, a 7B unified text-to-image and image editing model from Qwen. It runs in 8 denoising steps with a pre-configured sampling schedule and prefix KV caching, cutting inference cost while keeping the base model's capabilities: native RGBA transparency, multi-reference composition with up to 10 reference images, local edits via circles, painted annotations or masks, and reliable text rendering at native 2K resolution. It fits workflows that need fast iteration on generation and editing, and is also served through the Alibaba Cloud Model Studio API.

Claude Haiku 5.5

Claude Haiku 5.5 (API id claude-haiku-5-5) is Anthropic's newest Haiku-class model tuned for high-volume, latency-sensitive work, available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry. It ships a 1M token context window, 128k max output tokens, and adaptive thinking with an effort parameter for trading cost against intelligence. Anthropic calls it its most capable Haiku generation to date, with large benchmark gains over Haiku 4.5 in areas like computer use (OSWorld) and Terminal-Bench, and it works as a fast subagent alongside Opus 5.5 and Sonnet 5.5. Migrating from Haiku 4.5 involves breaking changes such as manual extended thinking with budget_tokens, changed sampling parameter behavior and computer tool version updates, so Anthropic published a dedicated whats-new page and migration guide.

GPT-6 Instant

GPT-6 Instant is OpenAI's everyday conversational model for the ChatGPT Chat experience, taking over the slot held by GPT-5.6 Instant. It rolls out to Plus, Pro, Business and Enterprise first (backed by GPT-6 Sol) and then to Free and Go tiers (backed by GPT-6 Luna), both tuned for everyday conversation. It debuts Intelligent UI, composing replies from text plus interactive components like buttons, forms and charts, and interleaves thinking with answering: OpenAI says it starts answering web-search questions 44% sooner on average than GPT-5.6 Instant. The update covers only ChatGPT's Chat experience, leaving Work and Codex models unchanged.

Vidu Q4 Preview

Vidu Q4 Preview is the public preview of Vidu's next-generation flagship video generation model, exposed in the API under the model id viduq4-preview. It supports joint audio-video generation with generated dialogue and sound effects, automatic shot switching, and 3 to 16 second clips at up to 4K resolution. The reference-to-video endpoint accepts up to 15 reference images and 3 reference audio tracks with voice cloning and lip sync, while image-to-video extends a single still into smooth high-motion footage with effects like explosions and particles. It targets AI short dramas, advertising, social media content, and film production, with API pricing starting at 9 credits per second for 540p output.

EmbeddingGemma 2

EmbeddingGemma 2 is Google's open, natively multimodal embedding model, built on the Gemma 4 architecture and released under the Apache 2.0 license with 740M parameters. It maps combinations of text, code, images, audio, and video into a unified embedding space, handling an 8K token context (up to 5.5 minutes of audio, 29 images, or 58 video frames) with Matryoshka truncation from 768 down to 128 dimensions, and achieves leading scores among sub-1B multimodal embedders on MTEB Code and MAEB. Quantized, it runs on a Pixel 11 Pro with as little as ~191MB active RAM for text-only workloads (270M parameters, with optional 170M vision and 300M audio encoders for the full model), making it suited to on-device search, semantic code search, and privacy-first RAG.

Mistral Large 4

Mistral Large 4 (unofficially ML4, codenamed le Chonk) is Mistral AI's flagship natively multimodal LLM, built on a granular MoE architecture with 1.05T total parameters, 49B active parameters, and a 1.6B vision encoder, available now in public preview on the Mistral Studio API with a 1M token context. Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, it targets coding, agentic workflows, and multimodal understanding, ranking state-of-the-art among open models on cybersecurity, finance, and legal workloads, with visual grounding that surpasses some frontier closed models. Open weights are slated to follow, and Mistral positions it as the foundation for a new generation of specialized models. It suits enterprise scenarios such as security operations, document and chart understanding, financial and legal knowledge work, and scientific computing.

Qwen-Image-2.1-Pro

Qwen-Image-2.1-Pro is a unified text-to-image generation and image editing model in the Qwen series, built around a compact visual generation component of only 7B parameters in a 32-layer single-stream DiT with hybrid-granularity attention and prefix KV cache reuse. It natively generates transparent RGBA images, edits transparent layers, and extracts subjects from photos in a single model; editing accepts up to 10 reference images and marks local regions with circles, hand-drawn annotations, or separate masks while preserving the identity of people and products. It also improves typography, portrait lighting, and detail rendering, and is served through Alibaba Cloud Model Studio.

FLUX 3 Image

FLUX 3 Image is Black Forest Labs' flagship image model for controllable generation and editing, served through a single endpoint where the prompt itself decides whether to generate or edit, with no mode field. It supports pixel-exact local edits (recolor, replace, move, resize, or remove several elements in one request), composition of up to ten reference images, and layout control via JSON bounding boxes written directly in the prompt. Resolution spans 768sq fast previews to 4K (about 16 megapixels) across 15 aspect ratios, priced from $0.041 to $0.607 per image. It fits poster and cover design, typography-heavy layouts, and multi-reference compositing.

MAI-Transcribe-2-Streaming

MAI-Transcribe-2-Streaming is Microsoft MAI's first streaming speech recognition model, built for real-time conversational voice agents. It produces low-latency transcripts in 60 languages with automatic continuous language detection, emitting first partial hypotheses just over 100 ms after receiving audio and revising them into a stable transcript as more context arrives. It ranks no. 1 on Artificial Analysis for the accuracy of both final and partial transcripts, and Microsoft's internal evaluations show words appear 2x faster than its closest competitor in live dictation and subtitling. It is offered at an introductory price of $0.54 per hour of audio and pairs naturally with MAI-Voice-2.1-Flash for low-latency voice agent loops.

MAI-Voice-2.1

MAI-Voice-2.1 is the highest-fidelity, most expressive text-to-speech model in Microsoft's MAI-Voice family, generating natural speech in 23 languages from text or a short reference audio clip. It supports gated instant voice cloning, SSML-based emotion and style control, and long-form generation with stable speaker consistency, making it suited to audiobooks, podcasts, lectures, educational content, and voice overs. The model is offered in public preview via Azure Speech, alongside the faster MAI-Voice-2.1-Flash variant that targets latency-critical scenarios such as real-time voice agents.

MAI-Voice-2.1-Flash

MAI-Voice-2.1-Flash is the blazing-fast text-to-speech variant in Microsoft's MAI-Voice 2.1 family, built for fast, low-latency generation. It produces high-fidelity, natural, and expressive speech across 23 languages, supports gated instant voice cloning from a short reference clip, and offers fine-grained emotion and style control via SSML. Optimized for real-time responsiveness, it targets voice agents, assistants, IVR, and call-center flows where latency and cost are critical. The model is available through the Azure Speech SDK, Voice Live, and Microsoft Foundry.

PixelUMM

PixelUMM is an NVIDIA-developed encoder-free unified multimodal model that handles understanding and generation across text, images, and video directly in pixel space. It uses a Qwen3-8B language backbone in a 15.2B-parameter decoder-only Transformer, represents RGB images as 16x16 pixel patches, and generates images and video with an iterative denoising head. Supporting text-to-image generation, image-conditioned text generation, video-conditioned understanding, and video generation, it targets research on unified multimodal modeling and pixel-space representation learning, with the checkpoint released under a non-commercial research license.

Praxis-1

Praxis-1 is Runway's first open-weight world action model, a generalist robotics policy that turns the large-scale web video pretraining behind Runway's world models into control for real robots. Because it learns physical plausibility and object behavior from video, the same policy generalizes across embodiments and environments, from bimanual arms to mobile bases. In official evaluations, web video pretraining matched teleoperation data on placement tasks (16.1 cm vs 16.0 cm final placement error), and simulated results tracked real-world performance with a 0.95 correlation. It is being tested with early partners Noble Machines, Standard Bots and Ultra, with the open-weight public release to follow.

Suno Speech

Suno Speech is Suno's new spoken-audio model, which the company describes as the first audio model that generates voice and music together as one cohesive track. Users type an idea, a poem or something they have written, then describe the voice and musical style they have in mind, and Speech returns spoken audio set to original background music. The beta is built into Suno and now open to everyone, fitting uses such as dramatic readings, meditations, pep talks and bedtime stories.

Gemini 4 Argon

Gemini 4 Argon is Google's new frontier model and the first release of the Gemini 4 series, built to sustain deep reasoning across complex, long-horizon workflows in real-world software engineering, enterprise knowledge work in legal, finance, and tax, and defensive cybersecurity. Google reports state-of-the-art results including 77.9% on DeepSWE v1.1, 91.7% on LVBench for long video understanding, and a tie for first at 68% on CWE-bench v1, alongside a 1M token output limit. The model rolls out in phases, starting with trusted network defenders in the Fairwind Program before reaching developers, enterprises, and consumers, with introductory pricing of $2 per million input tokens, $10 per million output tokens, and cached input at 5% of the input price.

Ideogram 4.5

Ideogram 4.5 is Ideogram's precision image edit model, built to curb the pixel shifts, color drift, and texture artifacts that image models accumulate over repeated edits. Its precise-edit pipeline copies pixels the edit does not touch exactly from the source image, supports an optional mask plus up to 4 reference images (3 when a mask is used), and returns the result at the input's exact width and height. The model edits high-resolution images without downsizing, with the official demo showing local edits stitched seamlessly back into a 24.2 MP photo, while generation supports 1K/2K presets up to 2048x2048 and aspect ratios from 1:6 to 6:1. Use cases include product colorways, photo restoration, sign and text edits, and large-format print retouching, and it is available in the Ideogram app and via API.

GPT-6.1 Sol

GPT-6.1 Sol is OpenAI's upgrade to GPT-6 Sol, with major gains in agentic coding, computer use, and professional work. It delivers near GPT-6 Astra performance at about one fifth the standard token price, with API pricing of $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. OpenAI reports that on GDP.pdf it scores above Opus 5.5 at less than half the per-task cost, and on OSWorld 2.0 it narrows the gap to Astra to within 2.1 percentage points. It is available in ChatGPT Work and Codex and callable via the API as gpt-6.1-sol.