Meta Muse Wonder vs Gemini 4 Argon (2026)

The short version

Neither model is better across the board. They come from labs with different goals, and the launch posts show it:

• Muse Wonder is Meta's agent model. Meta's pitch is Active Memory, which keeps context across conversations, and Active Focus, which keeps an agent working when its first plan fails and has it check in before consequential steps. Its strongest published results are on long coding runs (SWE-Marathon) and professional-task suites.
• Gemini 4 Argon is Google's depth-and-precision model. It leads the Vals Index at 68.9% and Vals' Finance Agent benchmark, and Google's launch post focuses on long documents, legal and financial analysis, large-codebase refactors, and finding security vulnerabilities.
• On one shared yardstick they are almost level. The Artificial Analysis Intelligence Index puts Muse Wonder at 52, Argon at 53, and Wonder Full at 56. Claude Opus 5.5 leads that index at 58.
• Meta's stronger model is not in the API. Muse Wonder Full is only available inside Muse Code subscriptions. If you are building a product, you get the standard Wonder.
• Argon is half price for six weeks. $2 / $10 runs until November 30, then the price rises to the same $4 / $20 as Wonder.

Muse Wonder: Meta's first closed frontier model

Muse Wonder is the first frontier model since Meta reorganized its AI work in 2025, and it was trained at Meta's new Prometheus data center in Ohio. It builds on Muse Spark, the small, low-latency model Meta released earlier this year. Wonder is much larger and is aimed at the same agentic coding and knowledge work as the top Claude and GPT models. Meta lists law, finance, health, and coding as its target areas.

The bigger change is how Meta is releasing it. Llama made Meta the main source of open-weight models, but neither Wonder model has downloadable weights. You can't self-host either one, fine-tune it on your own hardware, or run it in your own VPC. You access it through Meta's API like any other closed model.

Meta is selling two features on top of the benchmarks:

• Active Memory keeps relevant personal and professional context between conversations. TechCrunch reports that Meta says Wonder beats Fable, Opus, and Astra on MemoryBench, a continual-learning benchmark. That result comes from Meta's own testing.
• Active Focus is Meta's name for persistence. The model is trained to try other approaches when a plan fails, recover from errors, ask for clarification, and check in before important actions.

On the API, Wonder accepts text, images, audio, and video, returns text, and supports streaming, tool calling, structured output, and adjustable reasoning effort. Meta also offers OpenAI- and Anthropic-compatible endpoints, so trying it in an existing app is mostly a change of base URL and model ID. Limits at launch are 3,000 requests and 4 million tokens per minute. Built-in web search costs an extra $2.50 per 1,000 queries.

Wonder vs Wonder Full

Wonder Full is the stronger model. On Meta's own tables it scores 4–6 points higher on most coding and workflow evals, and 56 vs 52 on the Artificial Analysis index. It has no API and no pay-as-you-go price. You can only use it inside Muse Code, Meta's coding agent, on the Power Usage plan ($50/month) or the Maximum Usage plan ($200/month). Both plans have usage caps. For developers, that means you can try Wonder Full as a coding tool but can't ship it in your product.

Gemini 4 Argon: Google's strongest model

Google calls Argon its most powerful model, and the launch was staged. On September 30 it went only to government, critical-infrastructure, and security partners through Google's Fairwind Program. General access arrived on October 13: the paid Gemini API, Google AI Studio, Google Antigravity, Gemini for business, and paid Gemini app plans. It is not available on the free tier.

Google's launch post focuses on careful long-running work rather than chat:

• Long, mixed documents. It reads files that combine text, tables, charts, and images and produces grounded output. Google's examples include ledger reconciliations, variance analysis, and pulling verified citations from case law. It scores 91.7% on LVBench, a long-video benchmark.
• Large-codebase work. Google says the Chrome team used Argon to finish a refactor across hundreds of files in Chromium's roughly 40-million-line codebase, and that internal teams using it in Antigravity submitted over 35% more agentic code changes with almost a third fewer rollbacks. These are Google's own figures.
• Defensive security. Argon ties for first on CWE-Bench at 68% for fixing vulnerabilities. Google also says it is its most resistant model to indirect prompt injection, with a 0.7% attack success rate on Gray Swan's benchmark across 15 adaptive attempts.
• Honest reporting. Google says Argon is trained to report negative findings, say when it couldn't complete something, and pause for confirmation before irreversible steps. That is the same goal as Meta's Active Focus.

The general version blocks offensive security tasks such as pentesting and exploitation. Defensive work goes through Google's CodeMender or your own security harness, and the version without those guardrails is limited to vetted Fairwind partners. If you build security tooling, test your actual prompts before you commit.

Benchmarks

The two labs published different evaluation suites, so there are few direct comparisons. The fairest one is an independent composite index that scores every model the same way:

Two things stand out. First, the top of the market is now tightly packed: six frontier models fall within six points of each other. Second, the price no longer follows the score. Muse Wonder and Argon score about the same as GPT-6 Astra and Fable 5.1 on this index at well under half their list price. Artificial Analysis also estimates cost per completed index task, which accounts for how many tokens each model uses: $4.52 for Muse Wonder, $5.98 for Opus 5.5, and $3.26 for Astra. A model that uses more tokens per task can cost more overall despite a lower per-token price.

What Meta published

Meta's table compares both Wonder models with GPT-6 Astra and Claude Opus 5.5, all at maximum reasoning effort:

The clearest result is SWE-Marathon, where both Wonder models lead the long-running coding test by more than ten points. Meta did not hide its weaker results: Opus 5.5 and Astra both beat Wonder Full on Terminal-Bench 4.0 and HealthBench Pro. The legal scores are low for every model, which means no current model handles real legal agent work reliably. Meta's table also combines its own runs with vendor-reported and third-party scores, and some coding results used different agent harnesses, so treat small gaps as noise.

What's published for Argon

Google cites Vals AI, which runs its own independent benchmark suite. Selected Argon results from Vals:

Argon is strongest on careful, document-heavy professional work: finance, legal, and code migration. It is weakest at operating a computer. On Vals' computer-use benchmark it placed 8th of 9, so it's a poor fit for browser or desktop automation. Its Terminal-Bench 4.0 score (57.6%) is close to Wonder Full's 57.1% in Meta's table, but the two were measured with different harnesses and are not directly comparable.

Pricing in practice

The list prices match, but three details change the bill:

• Argon's introductory price. Through November 30, Argon costs $2 / $10, half of Wonder's price. That makes it the cheaper model to evaluate now. Don't base a long-term budget on the intro rate.
• Cache pricing ends up the same. Wonder charges $0.20 per million cached input tokens. Argon discounts cached input by 95%, which also works out to $0.20 at full price. For agent loops that re-send a large prompt each turn, neither has an advantage.
• Wonder uses one rate at any context length. Meta charges the same per-token price whether a request uses 10K tokens or close to a million. Check Google's pricing page for Argon's long-context rates before you commit to whole-repository or long-document workloads.

What to check before you build on either

• Data mode on Muse Code. Muse Code defaults to Contributor mode, in which Meta may use your prompts, code, files, and outputs to train future models. Turning it off only applies to future submissions. The Wonder API runs in Standard mode, which is not used for training. Make sure your team knows which one they're using before anyone pastes in a client's code.
• Zero data retention removes Wonder's best features. ZDR is available to approved API organizations, but it disables server-side conversation state, background requests, file uploads, and built-in web search. Active Memory depends on several of those, so under ZDR you get a capable model without the features Meta is promoting.
• Argon refuses offensive security work. Pentesting and exploit prompts are blocked on the general release. Defensive scanning and patching work, but test your real workflow first.
• Neither has a free API tier. Argon is paid-only on the Gemini API. Muse Wonder is only on the paid Meta Model API.
• Long runs need async handling. Both models are designed to keep working on hard tasks, which means long individual calls. Stream responses, set generous timeouts, and don't block a user request on one model call.

Which one fits your workload?

The honest answer

Meta now competes in closed frontier models alongside OpenAI, Anthropic, and Google, at the lower price tier. That price is the most useful part of these launches. Muse Wonder and Argon both score close to GPT-6 Astra and Fable 5.1 on the shared index at $4 / $20 instead of $10 / $50. Opus 5.5 is still at the top of that index for the same $4 / $20.

Choose by the shape of the work. For agents that run for a long time and need to remember context, test Muse Wonder. For work that requires reading carefully, such as financial filings, contracts, large codebases, and security reviews, test Argon, ideally before November 30 while it's half price. Both use standard chat-completion APIs with tool calling, so the most reliable comparison is to run fifty real requests from your product through each one. The labs' benchmark tables mostly measure the tasks each lab chose to highlight.