Hot AI News

New AI models this week: Claude Fable 5.2 and Gemini 4 Hydrogen

Claude Fable 5.2 leads a week with Gemini 4 Hydrogen and Carbon, Meta's low-cost Muse Ember, and open-weight models from Thinking Machines, MiniMax, Zyphra and Cactus.

New AI models this week: Claude Fable 5.2 and Gemini 4 Hydrogen
Graphic: Aiholics

Anthropic’s Claude Fable 5.2 is the biggest release of the week, arriving alongside a restricted sibling called Mythos 5.2. Google added two Gemini 4 models, Carbon and Hydrogen, and Meta listed a very cheap Muse Ember. On the open side, Thinking Machines, MiniMax, Zyphra and Cactus Compute all published weights on Hugging Face.

Claude Fable 5.2: Anthropic’s top model, now cheaper than Fable 5.1

Anthropic logo
Graphic: Aiholics

Anthropic released Claude Fable 5.2 on November 9, together with Claude Mythos 5.2. Anthropic says the two are the same underlying model: Fable is open to all users and carries strong cyber and biology safeguards, while Mythos is limited to vetted cybersecurity and biomedical research organizations. The API price is $8.00 per 1M input tokens and $40.00 per 1M output tokens, and an optional fast mode costs $16 and $80 per 1M.

Anthropic says output is up to 30% faster than Fable 5.1 and reports 49.1% on FrontierCode 2.0 (against 45.5% for Claude Opus 5.5) and 89.3% on OSWorld 2.1 (against 87.2%). These are the company’s own test results, and the post was not reviewed by outside organizations. Anthropic also suggests starting with Opus 5.5 for most coding and knowledge work and moving to Fable only for the hardest tasks. The tracker lists a 1,000,000-token context window. Details are on Anthropic’s announcement.

Advertisements

Gemini 4 Hydrogen: Google’s fast, lower-priced model for agentic work

Benchmark chart comparing Gemini 4 Hydrogen to other models
Image: Google

Gemini 4 Hydrogen is Google’s new model for agentic workflows and everyday tasks, announced on November 12. It costs $1.50 per 1M input tokens and $7.50 per 1M output tokens, with a 1,048,576-token context window. Google reports a score of 51 on the Artificial Analysis Intelligence Index at its default setting, one point above the larger Gemini 4 Argon, and 54 at the new xhigh thinking level. Those figures come from Google and were not independently verified here.

It is available in the Gemini app, Google Search, AI Studio, the Gemini API and Gemini for business. Paid AI Pro and Ultra subscribers get it in Antigravity. See Google’s announcement and our reports on pricing and benchmarks, the GitHub Copilot rollout and the Devin results.

Gemini 4 Carbon: Google’s heavier model for long-horizon tasks

Google logo
Graphic: Aiholics

Gemini 4 Carbon is built on Gemini 4 Argon and, according to its model card, uses more thinking tokens to test hypotheses and run verification loops. Google positions it for long-running agentic work in knowledge work, engineering, cybersecurity and scientific reasoning. The tracker lists it at $2.00 per 1M input tokens and $10.00 per 1M output tokens, and the model card gives a context of up to 1M tokens across text, images, audio and video input.

Google’s card reports 65.2% on Terminal-Bench 4.0 and 77.4% on Agents’ Last Exam v1.1. These are the maker’s numbers. A Fairwind Program gives verified cyber defenders specialized capabilities. Read the model card, and see how Antigravity’s Orchestrator lets Carbon and Hydrogen share one coding session.

Muse Ember: Meta’s cheapest model on the tracker

Muse Ember
Image: Meta

Muse Ember is a Meta model that appeared on OpenRouter on November 9. The tracker lists it at $0.10 per 1M input tokens and $0.50 per 1M output tokens, with a 1,048,576-token context window. That makes it the cheapest model of the week by a wide margin. We could not open a Meta page for it, so capability claims are left out until Meta publishes its own details. The OpenRouter listing has the current provider terms.

Advertisements

Inkling 2 Small: Thinking Machines’ open-weight mixture-of-experts model

Thinking Machines Lab published Inkling 2 Small on Hugging Face under the Apache 2.0 license. According to the model card, it has 594B total parameters with 17B active per token, routing each token to 8 of 384 experts. It takes text and image input, produces text, and supports up to 1M tokens of context.

The main NVFP4 checkpoint needs roughly 500 GB of combined GPU memory, for example two B300 or four H200 cards, so it targets servers rather than desktops. The card lists agentic and tool-use systems, coding assistants and retrieval applications as intended uses. It also warns of hallucinations and weaker performance in long multi-turn conversations. Find it on Hugging Face.

MiniMax-H3-Image: An open-weight image generator with a custom license

MiniMax H3 image, Open Weights
Image: MiniMax

MiniMax-H3-Image from MiniMax handles both text-to-image generation and editing with up to 16 reference images in one checkpoint. The card describes a 33.1B-parameter diffusion transformer, distilled to 6 sampling steps, that outputs images of about 16 megapixels without an upscaler. It runs in SGLang, vLLM-Omni, diffusers, LightX2V and ComfyUI.

The license is the MiniMax-H3-Image Community License Agreement, not Apache or MIT, and the card says separate written authorization is requested through an application form. The main model file is about 68.65 GB, and the text encoder is a separate download of about 66.7 GB. See the model card.

Zonos3-Mini: Zyphra’s open text-to-speech model for 43 languages

Zyphra’s Zonos3-Mini is a text-to-speech model released under Apache 2.0. It uses about 100M active parameters per token, covers 43 languages and can clone a voice from a reference clip. Users can also design a voice by setting age, accent and vocal texture, and control emotion and pace. Zyphra says it runs in real time on phones and laptops, including in the browser through WebGPU. Weights are on Hugging Face, and we covered the release in our own report.

Advertisements

Mirage: Cactus Compute’s small vision model for devices

Mirage from Cactus Compute is an Apache 2.0 vision model for image understanding and structured extraction, tagged for tasks such as document question answering, object detection and visual question answering. The model card gives no parameter count. It belongs to Cactus Compute’s collection of foundation models for tiny devices and installs with the cactus-compute Python package. Our article covers its size and phone speed, and the weights are on Hugging Face.

Arete-27B: Nous Research’s fine-tune of a Qwen model

Arete-27B
Image: Qwen

Arete-27B is a 27B-parameter model from Nous Research, released under Apache 2.0. It is a supervised fine-tune of Alibaba’s Qwen3.8-27B, so it counts as a derivative rather than a new base model. Nous says it aims to reinforce Western values on a Chinese-trained model. The developers report that answers containing propaganda fell from 22% to 0.9%, judged by an LLM, and that general benchmark scores stayed roughly flat. These results have not been independently verified. See the model card.

The week in numbers

  • Models in this roundup: 9 entries, including 4 with API prices on our price tracker.
  • Cheapest per 1M input tokens among the priced models: Muse Ember at $0.10.
  • Most expensive per 1M input tokens: Claude Fable 5.2 at $8.00.
  • Open-weight releases with permissive Apache 2.0 licenses: Inkling 2 Small, Zonos3-Mini, Mirage and Arete-27B.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Alex CarterAI News & Big Tech Correspondent

Alex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like