New AI Models

Claude Sonnet 5.5

Claude Sonnet 5.5 (claude-sonnet-5-5) is Anthropic's mid-tier model positioned as the best combination of speed and intelligence, offering a 1M token context window, 128K max output tokens (up to 300k on the Message Batches API), and pricing of $2 / $10 per MTok. Adaptive thinking is on by default with an effort parameter controlling thinking depth, and it fits agentic workflows, coding, and high-volume production traffic. Code written for Claude Sonnet 5 needs five updates, including switching to thinking: between_tools to turn off up-front thinking and replacing forced tool_choice calls that now return a 400 error, so Anthropic ships a migration guide plus a dedicated prompting guide. The model is available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

Eleven v4

Eleven v4 is ElevenLabs' new flagship text-to-speech model (model_id eleven_v4), aimed at quality-first scenarios such as content creation, audiobooks, and character voiceovers. Compared with Eleven v3, it improves emotion, delivery, and the nuance of audio tags like [whispering] and [laughing], and its voice cloning captures the source voice's timbre, cadence, and delivery more faithfully. Language coverage expands from v3's 70+ to 90+ languages, the model is served through the Text to Dialogue API for long-form audio, and the same family's Eleven v4 Turbo covers real-time applications.

Eleven v4 Turbo

Eleven v4 Turbo is the realtime tier of ElevenLabs' v4 speech family, served through the Text to Dialogue WebSocket with a median inference latency of about 100 ms, down from roughly 280 ms for eleven_v3_conversational (both excluding application and network latency). It is aimed at support agents, AI assistants and interactive characters that need expressive speech without perceptible delay. Like the flagship Eleven v4, it offers high-fidelity voice cloning, support for over 90 languages and audio tags for fine-grained delivery control.

Kling 4.0

Kling 4.0 is the flagship text-to-video and image-to-video model from Kling AI, unveiled with its full rollout slated for October. It natively generates clips up to 30 seconds, supports 21:9, 16:9, 1:1 and 9:16 aspect ratios, and outputs 720p to 4K, with 10-bit HDR for 1080p and 4K to follow. An omni-reference system combines up to 15 inputs across images, videos, subjects and voice, while up to 10 keyframes give precise control over narrative beats. It also delivers dual-channel stereo audio with multilingual lip-sync, renders multilingual text and emoji, accepts prompts up to 8000 tokens, and supports video editing to add, modify or remove subjects and backgrounds.

Kling 4.0 Flash

Kling 4.0 Flash is the lightweight, cost-efficient tier of the Kling 4.0 video generation family from Kling AI, opened first in a small-scale preview while the full Kling 4.0 rollout is yet to follow. Aimed at high-frequency creation scenarios, it cuts generation time substantially and pairs that speed with low cost. The model covers text-to-video, image-to-video driven by a first frame, and the omni-reference mode, producing clips of 3 to 20 seconds at 720p in 8-bit SDR, a practical setup for fast everyday video production.

LongCat-2.5-Preview

LongCat-2.5-Preview is a preview multimodal model on Meituan's LongCat API platform. It adds image understanding that parses image content and supports cross-modal Q&A, content summarization, and complex visual reasoning. It also excels at code generation, code understanding, and automated programming tasks, and works efficiently in agent coding environments such as Claude Code, OpenCode, and Kilo Code.

Gemini 3.8 Live with Live Avatar

Gemini 3.8 Live with Live Avatar extends the Gemini 3.8 Live speech model with a real-time visual persona, streaming low-latency video alongside speech with precise lip-syncing, natural expressions, and fluid turn-taking. It processes audio and visual inputs simultaneously for multimodal conversation, supports asynchronous tool calling that runs in the background while dialogue continues, and natively syncs speech-to-speech across 97 languages with mid-conversation language switching. Enterprises can pick from a preset avatar library or create custom avatars from a reference image, and all generated audio and video is watermarked with SynthID. The model is offered through Gemini Enterprise and the Google Cloud agent platform.

Gemini 3.8 Flash TTS

An expressive text-to-speech model in the Gemini audio family, built for deep creative direction and character design. Natural language prompts define the role, accent, and personality to create character voices from scratch across 100+ languages and dialects, backed by 2,000+ production-ready voices. It supports line-by-line direction with acting cues, pacing, dialect shifts, and backchanneling via tags like <laughs> and <sigh>, stages native two-speaker scenes from a single script, and generates hours of long-form audio with minimal speaker drift. Google reports a #1 rank on Hume's Voice Design Benchmark (71.4); it is available in Gemini API, AI Studio, and Gemini Notebook, with an Enterprise API to follow, targeting games, audiobooks, and podcasts.

Gemini 3.8 Flash-Lite TTS

Gemini 3.8 Flash-Lite TTS is Google's lightweight text-to-speech variant in the Gemini Audio family, tuned for cost-efficient generation at high volume. It adapts tone and pacing automatically, draws on a production-ready voice library covering over 100 languages, and supports up to 64K output tokens for long-form narration. Aimed at everyday audio workloads such as video dubbing, podcast creation, and real-time voice agents, it ranks second on Hume AI's Overall Quality Index. The model is available in the Gemini API and Google AI Studio, with enterprise API access to follow.

Muse Realtime Avatar

Muse Realtime Avatar is Meta's audio-driven Diffusion Transformer that turns Muse Realtime Voice into expressive, interactive avatars, consuming the same speech-token (VQ) stream to generate synchronized voice-and-video in real time. Conditioned on reference media and a rolling window of recent video latents, it animates portraits, full-body illustrations, animals and everyday objects in causal chunks, keeping appearance and mannerisms coherent for as long as the conversation runs. Self-forcing and distribution matching distillation compress a 120-evaluation guided teacher into a two-step unguided student (a 60x reduction), streaming 448x768 portrait video at 25 fps with roughly 870 ms of end-of-turn latency. In live comparisons raters preferred it over Runway Characters by 78% to 22% and over HeyGen LiveAvatar by 88% to 12%.

Claude Opus 5.5

Claude Opus 5.5 (API model claude-opus-5-5) is the first model in Anthropic's Claude 5.5 family, built for long-running agentic coding and knowledge work such as codebase-wide migrations and audits. It ships a 1M token context window by default, 128k max output tokens, and always-on adaptive thinking, with Fast mode (research preview) on the Claude API for up to 2.5x speed. Anthropic reports it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 at $4/$20 per million tokens, scoring 66.4% on Terminal-Bench 4.0 and 54.4% on FrontierCode v1.1. It is available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

FLUX 3 Action

FLUX 3 Action (F3A) is an open-weight 7B world action model (WAM) from Black Forest Labs, derived from the multimodal FLUX 3 backbone and trained to jointly predict future video frames and robot actions rather than following the VLA paradigm. Pretraining covers image, video and audio with over 95% of training tokens from video, followed by action-aligned midtraining across 14 embodiments and joint video-action fine-tuning on DROID. On the RoboLab-120 benchmark its single-step 7B checkpoint reaches 38.3% success, ahead of Cosmos 3 Nano (36.8%) and Pi0.5 (28.0%), while a guidance-distilled checkpoint raises the state of the art to 42.2%. Weights are downloadable via Hugging Face, and the official team uses gaming navigation as a testbed, anticipating computer-use and other latency-sensitive agents.

Hy Image3.5 preview

Tencent Hunyuan's Hy Image3.5 preview is a cost-efficient image generation model for professional creative work, continuing the Hy Image3.0 line and released as a preview ahead of the official version. It supports text-to-image and image-to-image with up to 5 reference images per request, multi-turn creation and continuous editing based on prior context, multiple aspect ratios, and output up to 2K resolution. In GSB blind tests with over a hundred professional designers, its overall win rate against the previous generation improved by more than 30%, with gains in text rendering and layout, realism and style, and editing consistency. A simplified architecture, efficient distillation, and decoupled text and image deployment keep inference costs low: Tencent Cloud API charges 0.15 CNY per output image at 2K resolution or below (0.024 USD internationally) with free input, and the model is available in products including Yuanbao, ima, WorkRally, Miora, and OnSolo.

GPT-6 Luna

GPT-6 Luna is a low-cost tier in OpenAI's GPT-6 lineup, trained with methods similar to GPT-6 Astra and built to bring that generation's gains in professional work, factuality, coding, and computer use to everyday, high-frequency tasks at scale. API prices drop 50% from GPT-5.6 Luna to $0.10 per million input tokens and $0.50 per million output tokens, with cached input reads discounted 90%. At max effort it scores 66.6% on DeepSWE v1.1, comparable to Claude Opus 5 and Fable 5 at medium effort while costing 93% and 96% less per task respectively, and at high effort it improves on its predecessor by 5.4 percentage points on AutomationBench at 58% lower cost per task. It runs as gpt-6-luna in the API and in ChatGPT Work and Codex for paid plans, while Free and Go users can use it in the desktop app.

GPT-6 Sol

GPT-6 Sol is OpenAI's new mid-tier model in the GPT-6 family, built with the same training methods as GPT-6 Astra and aimed at complex coding and agentic work. OpenAI reports 68.8% on DeepSWE v1.1 at max effort, 33.2% on AutomationBench and 56.4% on Agents' Last Exam, matching or beating stronger-configured competitor models at a fraction of their per-task cost, with roughly half the factual errors of its GPT-5.6 Sol predecessor. API pricing is 50% below GPT-5.6 Sol at $2 per million input tokens and $10 per million output tokens, and cached input reads get a 90% discount. It is available in Codex and ChatGPT Work for Plus, Pro, Business, Enterprise, and Edu users, and in the API as gpt-6-sol.

MiMo-V2.6-Flash

MiMo-V2.6-Flash is Xiaomi's full-modality, high-intelligence, low-cost reasoning model, positioned as the efficiency tier of the MiMo-V2.6 series for high-frequency calls and large-scale tasks in professional workflows. The official model card lists a sparse MoE architecture with 309B total and 15B activated parameters, native text, image, video and audio input, and a 1M token context window. Trained with a single mixed RL run spanning coding, general agents, visual and cybersecurity tasks, it scores 67.9 on DeepSWE v1.1, 87.6 on Terminal Bench 2.1 and 95.1 on CyberGym in official evaluations. It suits cost-sensitive agentic workflows, coding assistants and multimodal automation where throughput matters.

MiMo-V2.6-Pro

MiMo-V2.6-Pro is Xiaomi's flagship omni-modal reasoning model and its most capable model to date, a trillion-parameter open-source model built for complex projects, long-horizon tasks, high-stakes work, cybersecurity, and research. Xiaomi reports a score of 46.32 on the Artificial Analysis Intelligence Index, making it the strongest open-source model to date, along with 71.9 on DeepSWE v1.1 and 94.0 on CyberGym. Combining 3D spatial reasoning, multimodal perception and computer-use agent (CUA) capabilities, it ships with an open technical report, training environments and RL code, and is served through the MiMo API Platform, AI Studio, MiMo Desktop and OpenRouter.

Grok 4.7

Grok 4.7 is the flagship LLM from x.ai, whose pages now carry the SpaceXAI brand, aimed at coding and knowledge work. It moves to a new, larger base model over Grok 4.6 with a longer reinforcement learning run weighted toward tasks that take many hours, and offers 500k context, configurable reasoning, and agentic tool calling. Official numbers include 46.3% on CursorBench 4.0 at the price-performance frontier, plus the company's strongest refusals and jailbreak resistance to date. Pricing stays level with Grok 4.6 at $2 per million input tokens and $6 per million output tokens, with a May 2026 knowledge cutoff, served via the Grok API, Cursor, Grok Build, and third-party coding harnesses.

Qwen3.8-Omni-Flash-Realtime

Qwen3.8-Omni-Flash-Realtime is the realtime interactive counterpart of Qwen3.8-Omni-Flash on Alibaba Cloud Model Studio, handling live audio and video interaction with text and audio output. On top of the base version it adds multichannel audio, video aggregation, and remote MCP tool calling, with WebSocket, WebRTC, and AOQ access. It targets realtime scenarios such as voice assistants and live customer service.

GLM-5.3-FlashX

GLM-5.3-FlashX is the high-speed tier of Zhipu AI's GLM-5 series, offered on Alibaba Cloud Model Studio with direct supply from Zhipu. It runs at up to 200 tokens/s with a 1M-token context window and 128K-token output, and natively accepts image, video, and file input alongside text. It supports deep thinking, function calling, and structured output, fitting low-latency multimodal assistants, long-document and video understanding, and high-throughput agent workloads.