Gemini 4 Argon vs Muse Wonder: same $4/$20 list price, different access and strengths

At a glance

The table is as of 13 October 2026. Muse Wonder Full costs $50 a month on Muse Code's Power Usage plan and $200 a month on its Maximum Usage plan, and has no per-token price.

What each is built for

Google pitches Gemini 4 Argon at real-world software engineering, enterprise knowledge work such as legal and finance, and cyber defense. Per Google, it can find, validate and patch software vulnerabilities on its own, and it is already running large internal jobs, including a C/C++ to Rust migration for libgav1 that produced a decoder 2.7 times faster than the earlier Rust port.

Meta pitches Muse Wonder at personal and professional agents. It names two traits. Active Memory carries context about a user across conversations, and Active Focus keeps working through blockers while asking for clarification and checking in before important actions. Meta also lists frontier coding, trained on long-running tasks.

Gemini 4 Argon: our next era of frontier intelligence (Google)Google's announcement. Pricing, staged rollout and internal use cases.blog.google → Muse Wonder: A Frontier Model For Personal & Professional AI Agents (Meta)Meta's announcement. Active Memory, Active Focus and access tiers.research.meta.ai →

Both vendors compare against the same three rivals: GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Overlap is largest in coding and in legal and finance agent work. Argon leans toward enterprise and security buyers. Wonder leans toward assistants that remember a person and act across apps.

Scores: what can be compared

The two vendors publish their own tables, and only some rows line up. DeepSWE v1.1, a test of long software tasks, uses the same rival figures in both charts, so it is the cleanest head-to-head. The independent Artificial Analysis index covers every model on one scale.

On DeepSWE the gap between Argon and Wonder is under one point. The model card says Datacurve, a third party, ran Wonder's DeepSWE results. On the independent index, Argon and Wonder are within a point of each other, and Opus 5.5 is ahead of both. Artificial Analysis flags Argon's price as introductory, so its cost per task may rise once the $4 and $20 rate starts in December.

Most other rows do not match across the two charts. Terminal-Bench 4.0 is one example. Google lists Opus 5.5 at 66.4% and Argon at 57.4%. Meta's chart, which uses Artificial Analysis results, shows Opus 5.5 at 59.6, Wonder at 52.0 and Wonder Full at 57.1. The same rival gets two different scores, so Argon and Wonder cannot be compared there. Legal work has the same problem. Google reports Harvey's Legal Agent Benchmark, while Meta reports Artificial Analysis's version 1.1 with hallucination checks.

Access and safety

Argon rolled out in stages. On 30 September it went only to trusted cyber defenders through Google's Fairwind Program, and the paid Gemini API opened on 13 October. Google says it is taking part in the U.S. government's voluntary pre-release access process, and that trusted defenders and its own teams get a version without cyber guardrails.

Meta split Wonder in two. Muse Wonder is on the API. Muse Wonder Full, the stronger model, is sold only inside Muse Code plans. Meta says Muse Wonder will reach its Muse consumer experience soon, starting with subscribers.

Both labs publish their own risk statements. Meta's model card says it cannot rule out the unmitigated model reaching its high-risk threshold for chemical and biological misuse, so it applies safety training and system-level classifiers, and rates the mitigated model moderate or lower in every domain. Google says Argon is its most resilient model against indirect prompt injection, where hidden instructions hijack an agent, and that it monitors the model's reasoning and actions and stops runs when needed. All of this is self-reported. No independent safety comparison of the two was found.

Through OpenRouter

Both models are listed on OpenRouter, which puts many providers behind one OpenAI-compatible API, so a team can test both with a single key. The slugs are google/gemini-4-argon and meta/muse-wonder. Wonder Full is not listed. These figures are OpenRouter's own, from the first days of each model.

The main difference is routing. Argon has two Google routes, so a failing one can hand traffic to the other, while Wonder has a single provider. Wonder shows higher throughput and Argon shows lower latency, but one sample from launch week will move. Throughput also does not equal time per task. Artificial Analysis reports Wonder used about 130M output tokens to run its index, against a median of 85M.

Which one fits

• Argon fits teams that want the lower price through November, legal, finance or security work, PDF input, or long single outputs of up to 1M tokens. Google's table has it first on the Vals Index, which weights finance, coding, legal and tax work.
• Wonder fits teams that want a model on Meta's API for long coding tasks, where it is within a point of Argon on DeepSWE, or that plan to build on Meta's memory-focused assistant features as they arrive.
• Teams that want Wonder's top version cannot buy it by the token. Wonder Full needs a Muse Code subscription.
• Teams choosing on independent score alone should note that Claude Opus 5.5, the reference model both vendors chart, leads both on the Artificial Analysis index at the same $4 and $20 standard price.

For most buyers the practical test is the same. Run each model on a sample of the team's own tasks and compare cost per finished task, because cost per task depends on how many tokens a model spends. At the same $4 and $20 rate, Artificial Analysis puts Wonder at $4.52 per index task and Opus 5.5 at $5.98.

What to check before buying

• Every score except the Artificial Analysis index is vendor-run or vendor-selected. Meta's model card says its coding rows are not all run on one harness, the program around the model, and that some rival runs are best-effort.
• Meta's benchmark chart is an image, so the Wonder figures here come from outlets that transcribed it. Several outlets agree on 52.0 for Wonder on Terminal-Bench 4.0. One printed 85.6, which looks like an error.
• One outlet reports that Muse Code subscriptions train on user code by default. That was not confirmed in Meta's terms. Check it before pointing Wonder Full at a private repository.
• Meta lists visual and spatial reasoning, especially in 3D, as a known limit of Wonder.
• Argon's $2 and $10 price ends after November. Budget at $4 and $20 for anything that runs into December.

Sources

1. Gemini 4 Argon on Google DeepMind: Google's benchmark table and safety claims.
2. Gemini 4 Argon announcement (Google, 30 Sep): positioning, intro and standard pricing, staged rollout, internal use.
3. Gemini 4 Argon for developers (Google, 13 Oct): API availability.
4. gemini-4-argon model page (Google AI): model ID, context window, output limits, supported inputs.
5. Argon on OpenRouter: providers, latency, throughput, uptime, tier prices and cached input price.
6. VentureBeat on Argon's broad release: access tiers and pricing window.
7. Artificial Analysis: Gemini 4 Argon and NeuralTrust: independent index and cost per task.
8. Muse Wonder (Meta AI Research, 12 Oct): positioning, Active Memory and Active Focus, access.
9. Muse Wonder on OpenRouter: the single provider, latency, throughput, uptime and web search price.
10. Muse Wonder model card: context length, inputs, evaluation methods, risk assessment, limits.
11. VentureBeat on Muse Wonder: API price, limits and Wonder Full access.
12. MLQ, DEV Community, Fello AI and MyClaw: Muse Code plan prices, Meta chart figures, Artificial Analysis index and cost per task.
13. Beri: the unconfirmed report on training on user code.