New AI models this week: Meta’s Muse Wonder and Google’s Gemini 4 Argon

Two frontier launches set the pace between October 11 and October 18, 2026: Meta introduced its agent-focused Muse Wonder models, and Google opened Gemini 4 Argon to developers. Open weights were the other theme, with Meta, Nvidia and DeepSeek each releasing large models that teams can download and run themselves. Prices below are in US dollars per 1M tokens, taken from our price tracker, which reads OpenRouter’s catalogue.

Muse Wonder: Meta’s most capable models, built for agents

Meta released Muse Wonder and Muse Wonder Full on October 12, calling them its most capable models to date. According to Meta, they are built for personal and professional agentic tasks such as insurance claims, lease review and personal finances, as well as coding. Two features stand out: Active Memory, which carries relevant details across conversations, and Active Focus, which keeps an agent working toward a goal, helps it recover from errors and has it ask for confirmation before important actions.

Muse Wonder is available through the Meta Model API and Muse Code, Meta’s coding product, while Muse Wonder Full is available to Muse Code subscribers. The tracker lists Muse Wonder at $4.00 input and $20.00 output with a 1,048,576-token context window. Meta’s announcement gives no benchmark scores, license or pricing, and says a full report with additional safety evaluations is coming. More in Meta’s announcement and in our article on the Muse Wonder models.

Gemini 4 Argon: Google’s new frontier model for long tasks

Google presented Gemini 4 Argon on October 13 as its latest frontier model for multi-step work in coding, finance, legal, cybersecurity and science. The company highlights large refactors, performance optimization, spatial reasoning and multimodal document analysis. By Google’s own reporting, it ranks first on the Vals Index and scores 87% on the ERQA embodied reasoning benchmark; these are vendor-reported results.

The API has a 1M-token unified input and output limit. Google lists an introductory price of $2 per 1M input tokens and $10 per 1M output tokens, with cached input 95% cheaper, matching the tracker. From November 30, 2026, the public rates rise to $4 and $20. The model is also reaching AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search, and vetted cyber defenders can apply to the Fairwind Program for access without cyber guardrails. Read Google’s announcement and our piece on developer pricing.

Muse Spark Open: Meta’s 816B model with downloadable weights

On October 15 Meta published Muse Spark Open on Hugging Face. It is a sparse mixture-of-experts model with about 816B total parameters and about 42B active per token, using 256 experts per layer with 8 routed per token. It takes text, images, audio and video as input, produces text, and supports up to 1M tokens of context.

Weights come in BF16 and NVFP4 checkpoints, and the community has already added llama.cpp and MLX quantizations. Deployment paths include vLLM, SGLang and Hugging Face Inference Endpoints. The tracker prices hosted access at $1.25 input and $4.25 output. The Hugging Face post does not state a license, so check the model repository before building on it. See the Hugging Face post and our coverage.

Nemotron 3.5 Super VL: Nvidia’s vision model with a small active footprint

Nvidia’s Nemotron 3.5 Super VL arrived on October 15 as a vision-language model. Its checkpoint name, NVIDIA-Nemotron-3.5-Super-VL-120B-A12B, points to about 120B total and 12B active parameters. Nvidia’s documentation describes a Nemotron-H hybrid language decoder paired with a RADIO vision encoder, a vision projector and a separate temporal video embedder, plus shared Multi-Token Prediction weights. It handles image and video input.

It is the cheapest model of the week on the tracker, at $0.10 input and $0.40 output, with a 1,048,576-token context window listed there. Nvidia’s documentation page does not state a license. Details are in Nvidia’s Megatron Bridge documentation, and our article covers the Amazon SageMaker JumpStart release.

DeepSeek V4.1 Flash 1017: an MIT-licensed update with agentic gains

DeepSeek published V4.1-Flash-1017 on October 17 and describes it as a minor update to V4.1-Flash. The model card reports gains in reasoning, coding, agentic and visual tasks, along with shorter task execution time and output length. The model has about 763B parameters, a 1M-token context window and an MIT license that covers the weights.

DeepSeek’s self-reported results include Terminal-Bench 4.0 at 42.3, up from 31.2, DeepSWE v1.1 at 77.0 and HLE without tools at 42.6, up from 36.8. On several other tests the card shows Claude Opus 5 ahead. The tracker lists $0.15 input and $0.60 output. The card says no inference provider hosted the weights at the time of review. See the model card and our article.

MOSS-Image-SFT: an open 5B text-to-image model with layout control

Among Hugging Face releases that drew attention, OpenMOSS-Team’s MOSS-Image-SFT is a supervised fine-tuned text-to-image model built on MOSS-Image-Base. It has 5B parameters and is released under Apache 2.0, covering the weights and bundled inference code. It accepts natural-language prompts, tag sequences and JSON layouts that control text placement and object arrangement, and it supports English and Chinese. It runs through Diffusers with a custom pipeline or through a ComfyUI plugin and needs an Nvidia GPU with CUDA. Find it on the Hugging Face page.

The week in numbers

• Models on the price tracker: 5 (Muse Wonder, Gemini 4 Argon, Nemotron 3.5 Super VL, Muse Spark Open, DeepSeek V4.1 Flash 1017), plus one unpriced open image model.
• Cheapest per 1M input tokens: Nemotron 3.5 Super VL, $0.10 (output $0.40).
• Most expensive per 1M input tokens: Muse Wonder, $4.00 (output $20.00).
• Context window: 1,048,576 tokens for all five priced models.

For how these prices compare with other models, see our AI API pricing tracker.