Models

Google released Gemini 4 Hydrogen on November 12, a model it pitches as the fast execution engine for coding and agent work beside the larger Gemini 4 Carbon and Argon. It costs $1.50 per million input tokens and $7.50 per million output tokens across a 1 million-token context and is rolling out in the Gemini API, AI Studio, the Gemini app, Search and Antigravity. Artificial Analysis scores it 51 on its Intelligence Index at the default high setting and 54 at xhigh, where it ranks 16th of 78 models; most other benchmark figures are Google's own.

MiniMax released H3-Image on November 11, an open-weight text-to-image and image-editing model built on its H3 family, with a 33.1 billion-parameter diffusion transformer distilled to six sampling steps. MiniMax says it generates a 4096x4096 image in under 10 seconds and accepts up to 16 reference images; those are its own claims. The weights are on Hugging Face under a community licence whose territory excludes the European Union, the United Kingdom, South Korea and the United States, and which requires written authorization above $1 million in annual revenue or for hosted services.

Runway released Praxis-1 on November 11 with open weights under Apache 2.0, including inference and fine-tuning code. The 27-billion-parameter model predicts video and robot actions together and scores 46.2% success on Nvidia's RoboLab-120 simulation benchmark, first among world action models and ahead of FLUX 3 Action (42.9%). On the leaderboard's overall table it ranks fifth, behind systems including an agentic language-model policy at 86.0%. Runway says it adapts to new robots with light fine-tuning.

Google launched the Gemini Decisions API on November 11, powered by a new model, Gemini Decision 1, that scores closed-set questions (yes/no, 2 to 255 choices, or rubrics) against one text, image, audio, video or PDF input and returns probabilities and confidence scores instead of generated text. It handles up to 1,000 questions per request at $0.10 per million input tokens with no output charge. Google's own benchmarks put it ahead on five of seven shared tests; the independent Artificial Analysis Decision Index ranks it ninth of 18, per VentureBeat.

Alibaba opened a preview of Qwen4-Max on November 11, the flagship of its next-generation Qwen family, with text, image and video input, reasoning on by default and a 1M-token context window. API access is through its Token Plan subscription only, with no pay-as-you-go option, and Alibaba says the preview model will be retired after November 25. Alibaba has not published benchmark figures in its model documentation, and Artificial Analysis did not list the model when we checked.

Meta released Muse Ember on November 9, its smallest model, aimed at the repetitive steps agents run all day: tool calls, coding subtasks, routing, extraction and long-document parsing. It lists at $0.10 input and $0.50 output per million tokens, the same starting rates as Claude Haiku 5.5 and GPT-6 Luna. Artificial Analysis scored Ember 36 on its Intelligence Index, level with Haiku, at about $0.14 per task against $0.41, partly because Ember used fewer output tokens. Weights and size are not public.

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, the same underlying model behind two gates: Fable for all users with cyber and biology classifiers, Mythos for vetted cybersecurity and life-sciences organisations. Both cost $8/$40 per million tokens, which CNBC puts at 20% below 5.1. Anthropic says it limited the model's kernel-optimisation and architecture-research skills to about Opus 5.5's level. Its benchmark wins are its own; Cisco and IBM said they will build Mythos into security products.

Updated Nov. 12: Baidu on November 11 released ERNIE 5.5 Flash Preview, the first post-trained model built on ERNIE 5.5 Flash Base, with 280 billion total and 14 billion active parameters and a 1 million-token context. The API is available on Baidu's Qianfan platform, and Baidu says the weights will be open-sourced under Apache 2.0 soon; TokenPost reports before November 25. Baidu calls the Preview an intermediate post-training checkpoint. Earlier: on November 5 Baidu released the Base checkpoint itself, pretrained on 33.5 trillion tokens, under its own licence.

Updated Nov. 13: Thinking Machines Lab published the weights of Inkling 2 Small on Hugging Face on November 12 under an Apache 2.0 licence, in NVFP4 and MXFP8 checkpoints, meeting the release window it set at launch. The main NVFP4 checkpoint needs at least 500GB of aggregated GPU memory, for example two Nvidia B300s or four H200s. The model is also served by inference providers including Nvidia, Fireworks and Together. Original report: the lab introduced the 594 billion-parameter, 17 billion-active multimodal mixture-of-experts model on November 5 on Tinker with closed weights, promising open weights after external safety testing; its benchmark figures are lab-reported.

Google has put Gemini Omni 1.2 Flash into general availability on the Gemini API and Google Cloud, raising native clip length to 15 seconds from 10 on Omni 1.1 Flash and cutting the 720p price 30% to $0.07 per second. The update adds multi-turn conversational editing and stronger reference adherence; outputs carry SynthID, and Provisioned Throughput is available on Cloud. Independent ranking boards were not checked for this release.

Ricursive Intelligence, a lab building AI for chip design, released Ripple-65B Preview on November 4, a 65-billion-parameter dense model under the Apache 2.0 licence. It reports 80.52 on TerminalBench 2.1, ahead of several far larger US open-weight models, and says the whole project cost $32 million. All benchmark figures are the company's own.

Tencent released Hy4 on November 4 with image understanding and computer use for the first time, including training on Blender and CAD. It is live in Tencent's apps, on Tencent Cloud and through OpenRouter at the preview's prices. Tencent says the weights and system cards will follow "in the coming days". Its finance and hallucination figures come from evaluations Tencent ran itself.

Meituan released LongCat-2.5 on November 2, a natively multimodal model with 1.6 trillion parameters, about 48 billion active per token, built for long agentic coding tasks. Meituan says it trained and runs the model on domestic AI accelerators with under 80 GB of memory each. Its own benchmark table puts it close to Claude Opus 5 on Terminal Bench v2.1 but well behind on the newer v4.0. It is available by chat and API; weights are planned but not yet out.

NVIDIA released Nemotron 3.5 Transcribe on October 28, an open-weight 30B-parameter model with about 4B active that transcribes 40 languages, translates 31 and attributes speakers in 13, for recordings up to an hour. It also released the Granary v2 speech dataset. NVIDIA's leaderboard claims, including first among open models on Open ASR English, are its own; the model does not support streaming.

Updated Nov. 11: Google made Gemini 4 Carbon available on November 10 through the Gemini API for all paying developers and enterprises, and began rolling it out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Antigravity. The introductory price of $2 per million input tokens and $10 per million output tokens ends on November 30, when $4 and $20 apply. Artificial Analysis now lists Carbon (high) at 60 on its Intelligence Index, fourth behind three settings of Anthropic's Claude Fable 5.2. Google's new benchmark, cost-per-task and safety claims are its own. Earlier: Carbon went first to Fairwind cyber defenders.

Liquid AI said on October 27 that the first release of its LFM3 family covers six models in the 2B-parameter range, all trained from LFM3-1.7B-Base: a text model, a vision-language model, an audio model, an omni model and two decision models. Base and post-trained weights are on Hugging Face under Liquid's own LFM Open License v1.0, which limits commercial use to organizations with less than $10 million in annual revenue. The benchmark comparisons are Liquid's own.

Google DeepMind announced Genie 4 on October 27, a world model that generates interactive simulations together with music, ambient sound and dialogue, lets users change a running world with text prompts, and uses Gemini 3.8 Flash as an agent that monitors and steers each scene. It is available as a research preview to trusted testers, with Google AI Ultra subscribers and early-access Gemini API partners to follow. The Verge reports experiences can last about two minutes, up from one for Genie 3.

Ant Group's inclusionAI lab published the weights of Ling-3.1-flash-1019 on October 27 under the MIT license, along with the Ling-3.1-flash-base model. The model card describes a mixture-of-experts model with 560 billion total parameters, 25 billion active per token and a 1-million-token context window, post-trained for reasoning and agent tasks. Artificial Analysis now scores it 29 on its Intelligence Index v5.0, a different scale from the 44 on v4.3.2 that the model card cites.

OpenAI introduced GPT-6.1 Terra on October 26 at $1 per million input tokens and $5 per million output tokens, which OpenAI says is 50% below GPT-5.6 Terra and GPT-6.1 Sol. Its system card says that in a test run without safeguards, Terra saw a barrier and finished the task another way that "violated the spirit of the restriction" in 29.6% of maximum-effort rollouts, and that the test mostly covers low-stakes situations. The benchmark gains in OpenAI's launch post are OpenAI's own figures; Artificial Analysis independently scores Terra at max effort 48 on its Intelligence Index.

Updated November 5: Moonshot has published the Kimi K3.1 Pro weights on Hugging Face under its own Kimi K3.1 Pro License, meeting its 'by Nov 4' promise. The model card lists 2.8 trillion total parameters with 104 billion active, a 1-million-token context and MXFP4 weights, and repeats the 81.0% xdev-bench-hard score for the cyber-trimmed release. Artificial Analysis had ranked the API model fourth on its Intelligence Index; benchmark figures are Moonshot's own.

Block says Dessa, built by a small lab that reports to Jack Dorsey, is a full-duplex speech-to-speech model trained from scratch on human conversation, with no text and no pretrained models, and that it runs locally on a MacBook Pro M4 Max or better. Block has shared only its own recordings. It has published no benchmarks, model size, weights, API, pricing or release date, and early access is going to a small group of AI safety researchers.

Recursive Superintelligence released Flint, a 40B-total, 2.5B-active mixture-of-experts coding model, with weights downloadable from Hugging Face under Apache 2.0. Recursive says Flint leads MoE models under 50B parameters on three agentic coding benchmarks and runs at more than 125 tokens per second on an M4 Pro MacBook Pro, and that its Eureka Machine contributed to nearly every part of the build. All benchmarks are Recursive's own.

Mistral released Voxtral 3 on Oct. 20: speech-to-text models covering 50+ languages, Voxtral Speech 3 (Mini TTS 3) for 35+ languages and 50+ dialects, and a LiveKit-based voice-agent SDK. Listed prices include $0.0035/min for Mini Transcribe batch and $0.04/min for the agent API with Mistral Small 4. Benchmark wins are Mistral's own.

MiniMax published M3.1 Flash's weights on Hugging Face on Oct. 23 under a MiniMax community licence. Commercial users must display attribution, and products above $20M in yearly revenue need MiniMax's written authorization, so the release is open weights, not open source. Artificial Analysis had not yet updated its listing and scores the model about 42. Released Oct. 19 at $0.15/$0.60 per million tokens.