Google is making Gemini 4 Argon more widely available through its app and application programming interface (API). Meta is introducing Muse Wonder as an alternative for long, independently executed tasks, while Liquid AI is focusing on a model designed for local use. For you, the decisive points are less about leaderboard positions and more about three practical questions: where access is available, how quickly costs grow, and which data must be sent to the cloud.
What is available now
Gemini 4 Argon is a new multimodal model that can process text, images, audio, and video while producing text as its output. Multimodal means several types of media can be combined within one task. According to the Gemini 4 Argon release report, Google is positioning it for long, multistep work in areas including coding, finance, law, science, and cybersecurity.
Access runs exclusively through Google’s cloud, including the Gemini API and AI Studio; downloadable open weights are not available. An API is an interface that lets other applications call a model. Argon is also available in Google’s own app, although t3n reports that access is limited to paying customers. Free users reportedly remain on the Flash-Lite base model.
Meta released Muse Wonder and Muse Wonder Full during the same period. Both are described as agent models with a one-million-token context. A token is a small unit of text used to assemble prompts and responses. Muse Wonder is available through the Meta Model API, while Wonder Full is part of Muse Code plans and reportedly leads the DeepSWE and SWE-Marathon coding benchmarks. These provider and benchmark claims have not been independently verified in the supplied material.
Liquid AI is taking a different route. LFM2.5-24B-A2B is an open-weight model intended for local agents. Open weights can be downloaded and run on suitable hardware under your control. According to the provider, the model contains 24 billion parameters but activates only two billion for a task; this is called a Mixture-of-Experts architecture, which uses selected parts of the model rather than all of them at once.
What the limits and prices mean
Argon’s most striking figure is its one-million-token limit. The lead report’s headline calls this a one-million-token output limit, but the text describes a shared window covering both input and output. A large prompt therefore leaves less room for the response. This distinction matters: the million tokens are not an additional allowance for an equally large prompt and answer.
In practical terms, the window can accommodate very long documents, media, or multistep work histories within one context. Google reports a score of 91.7 percent on LVBench, a test of long-video understanding. Argon achieves 55.0 percent on the FrontierSWE v2 coding benchmark, however, placing last among the four models in Google’s listed comparison. Leaderboards depend heavily on the task being measured and do not replace a trial using your own material.
At launch, the Gemini API costs $2 per million input tokens and $10 per million output tokens. Simon Willison also identifies access through a $100-per-month subscription and reports that the API prices are scheduled to double on December 1. His own SVG test involving a pelican illustration cost 17.166 cents. That is a concrete example, but it is not a general measure of what your tasks will cost.
Willison also points out that token rates alone no longer make models directly comparable. The number of attempts and amount of generated text needed to complete a task matter as well. He cites an Artificial Analysis benchmark run costing $2,407 for Argon and places it below the latest models from OpenAI, Anthropic, and Meta on that measure. Meta Muse Wonder starts at $4 per million input tokens and $20 per million output tokens. The supplied information provides no directly comparable price for Wonder Full or Liquid AI’s model.
Who benefits in practice
If you are a beginner, the paid Gemini app is the most straightforward way to reach Argon because it avoids API setup. Start with a clearly limited task and material that can reasonably be sent to a cloud service. You could summarize a long video or combine several documents into a structured overview. Key statements still need to be checked against the originals, especially when the subject is legal, financial, or scientific.
For advanced users, the API provides more control over output length and spending. A small test with a representative task makes more sense than immediately requesting an extremely long response. Willison’s pelican example shows that Argon can also produce scalable vector graphics (SVG), a file format for graphics that remain sharp at different sizes. If you automate recurring work, you should also measure the tokens and correction rounds required to reach a usable result.
Google’s macOS app also includes a coding environment called Antigravity. In Task mode, Argon can reportedly work in the background and access more than 40 connected services, including Shopify, Adobe, and OpenTable. This may reduce waiting in professional workflows, but it also makes access controls more consequential. A system that can act autonomously should not have access to more accounts and data than its particular assignment requires.
A second practical example comes from cybersecurity. Google DeepMind says Wiz used Argon to uncover a critical flaw in hospital software that earlier models had missed, helping protect hundreds of thousands of patient records. Google’s own security teams also reportedly used it to locate and patch flaws in internal and open-source software. These are claims published by Google DeepMind, not an independent assessment documented in the supplied sources.
Liquid AI becomes more relevant when local processing matters more than the simplest cloud access. The provider reports a LiveClaw score of 48.09 and a decoding speed of 184 tokens per second on an Apple M5 Max. Those figures apply to particular hardware and a particular test. They do not show that the model will run quickly, or usefully, on every personal device.
Pros and Cons of the new model choices
Pros:
- Long workflows – Argon and Muse Wonder can reportedly use a context of up to one million tokens for large documents and multistep tasks.
- More access options – Argon is available through a paid app and an API, while Meta offers API access and Liquid AI supplies open weights.
- Different priorities – Google targets complex knowledge work, Meta emphasizes agents and coding, and Liquid AI focuses on local execution.
- Visible cost basis – Published token rates allow at least a rough estimate of API costs for Argon and Muse Wonder.
Cons:
- Cloud dependency – Argon runs only on Google’s infrastructure, so prompts and results cannot remain entirely within your own environment.
- Rising prices – Argon’s introductory API price is reportedly set to double on December 1, potentially making established workflows more expensive.
- Mixed measurements – A strong long-video result sits alongside last place in the coding comparison listed by Google.
- Incomplete comparisons – The sources lack comparable prices for Wonder Full and Liquid AI, while many performance figures come from the providers.
What remains unresolved
The sources provide no Switzerland-specific availability, prices, or confirmed language details. Because Argon is being offered broadly to paying Gemini customers and through the cloud, actual access will likely depend on the relevant account and plan; the supplied material contains no explicit Swiss confirmation. Swiss companies, schools, and other organizations therefore need to determine whether cloud processing of sensitive material is compatible with their internal data protection rules. Liquid AI’s open weights can offer greater control over where processing occurs, but open access alone says nothing about ease of use, hardware needs, or output quality.
The competitive comparison is also only a snapshot. The Verge describes the Google and Meta releases as renewed competition with OpenAI and Anthropic after Google had spent a period delivering mainly incremental improvements. A separate report says OpenAI may release one or two additional models by the end of 2026. Whether either will be its planned major upgrade remains uncertain; safety evaluations and an expected review by the US government could affect the process.
The three releases do not produce a single obvious winner but offer different compromises. Argon combines a very large working context with convenient cloud access, Muse emphasizes agents and coding, and Liquid AI promises greater local control. The total cost of completed workflows, independent confirmation of many performance claims, and responsible handling of sensitive data all remain unresolved.
Sources
- Google Releases Gemini 4 Argon API With 1M Token Output Limit – MarkTechPost, 2026-10-14
- Überraschung: Google veröffentlicht Gemini 4 Argon – t3n, 2026-10-13
- llm-gemini 0.35 – Simon Willison’s Weblog, 2026-10-13
- Developing Gemini 4 Argon responsibly and safely – Google DeepMind, 2026-10-13
- Meta Launches Muse Wonder and Muse Wonder Full – MarkTechPost, 2026-10-13
- Liquid AI Releases LFM2.5-24B-A2B – MarkTechPost, 2026-10-14
- OpenAI soll bis Jahresende mindestens ein bis zwei neue Modelle veröffentlichen – The Decoder, 2026-10-14
- The AI race is back on – The Verge, 2026-10-14
Image: Matheus Bertelli via Pexels