Meta · OpenRouter
Muse Wonder
meta/muse-wonder
AA Intelligence Index
52
Muse Wonder (Max)
AA cost per task
$4.52
Index-weighted average · upstream cost
← All models
MULTIMODAL / MODEL COMPARISON
Muse Wonder vs Gemini 4 Argon
Compare two multimodal models with million-token contexts through independent benchmarks, lab-reported results and three distinct pricing bases.
Sources checked 2026-10-14 · public evaluation snapshot
Google · OpenRouter
Gemini 4 Argon
google/gemini-4-argon
AA Intelligence Index
53
Gemini 4 Argon (High)
AA cost per task
$1.99
Index-weighted average · upstream cost
Benchmarks from the same evaluator
Artificial Analysis Intelligence Index v4.3.2. All published values below come from the same comparison page, and higher is better. The evaluated configurations are Max and High respectively.
| Benchmark | What it measures | Muse Wonder (Max) | Gemini 4 Argon (High) |
|---|---|---|---|
| Intelligence IndexIndex points | Weighted composite of 10 evaluations | 52 | 53 |
| AA-Briefcase v1.1Elo | Complex knowledge work and deliverables | 1659 | 1487 |
| GDPval-AA v2.1Elo | Real professional tasks and deliverables | 1770 | 1624 |
| AutomationBench-AAPercentage | SaaS workflows and tool use | 68% | 78% |
| Terminal-Bench 4.0Percentage | Agentic coding in a terminal | 52% | 57% |
| SciCodePercentage · Marked under review by the evaluator | Scientific code generation | 60% | 62% |
| Humanity's Last ExamPercentage | Advanced knowledge and reasoning | 50% | 57% |
| GDP.pdfPercentage | Professional work grounded in long PDFs | 31% | 22% |
| CritPtPercentage · Marked under review by the evaluator | Physics reasoning and computation | 28% | 27% |
| AA-OmniscienceIndex points | Knowledge reliability with hallucination penalties | 25 | 42 |
| AA-LCR v1.1Percentage | Long-context reasoning | 81% | 80% |
The composite overlaps with the component evaluations, so these are not 11 independent votes. Elo, percentages and index points should not be added together. AA currently flags SciCode and CritPt as under review; this page preserves the checked snapshot.
Use the results to choose tasks to evaluate
Muse: investigate deliverables and document work
Muse has higher published values on AA-Briefcase, GDPval-AA and GDP.pdf. Include it in evaluations of complex source synthesis, document analysis and deliverable creation, then check whether the advantage holds on your material.
Gemini: investigate automation, knowledge and terminal tasks
Gemini has higher published values on AutomationBench-AA, AA-Omniscience, HLE and Terminal-Bench. The overall index differs by one point; task-level quality, reliability and total cost are more useful for a production decision.
Additional lab results, with separate provenance
These results come from Meta’s launch page and Google’s model card. The two DeepSWE values were published separately; the chart and video evaluations have a verified number only on Google’s side.
| Benchmark | Muse Wonder | Gemini 4 Argon | Scope |
|---|---|---|---|
| DeepSWE v1.1 | 78.8% | 77.9% | Long-horizon software engineering; settings may differ across reports |
| Chartography | No matching score found | 71.6% | Chart understanding; reported by Google |
| LVBench | No matching score found | 91.7% | Long-video understanding; reported by Google |
Meta attributes ordinary Muse’s DeepSWE result to Datacurve. These are reported figures rather than a controlled run in one shared agent harness. Missing scores refer only to the launch pages and model cards reviewed here, and do not imply zero capability.
Pricing: distinguish upstream rates from platform billing
All rates are per million tokens. Current upstream prices refer to the checked Standard synchronous routes. Credits use list prices and the existing platform margin rule.
1. Current OpenRouter prices
Gemini currently has a 50% promotion; Muse uses its regular published rates. Cached input uses the cache-read price. Promotions and provider routes can change.
| Usage | Muse Wonder | Gemini 4 Argon |
|---|---|---|
| Uncached input | $4.00 | $2.00 |
| Cache read | $0.20 | $0.10 |
| Output | $20.00 | $10.00 |
2. List-price basis for platform credits
The two models currently share the same input, cache-read and output list prices. Gemini credit rates use those list prices, so the upstream promotion does not change its platform rate.
| Usage | Muse Wonder | Gemini 4 Argon |
|---|---|---|
| Uncached input | $4.00 | $4.00 |
| Cache read | $0.20 | $0.20 |
| Output | $20.00 | $20.00 |
3. Published platform USD and credit rates
These are the rates displayed on model cards, model pages and the playground, using the existing 30% target margin rule. Successful requests cost at least one credit.
| Usage | Muse Wonder | Gemini 4 Argon |
|---|---|---|
| Uncached input | $5.72 · 57,200 credits | $5.72 · 57,200 credits |
| Cache read | $0.29 · 2,900 credits | $0.29 · 2,900 credits |
| Output | $28.58 · 285,800 credits | $28.58 · 285,800 credits |
An upstream example with fixed token counts
Assume 10,000 uncached input tokens and 2,000 output tokens: current published upstream prices give $0.08 for Muse and $0.04 for Gemini. This arithmetic example fixes token usage; the models may consume different amounts or need retries on the same real task.
Matching published rates do not guarantee identical bills: token usage, caching, reasoning output and rounding can differ. Gemini settles at the published token rates; Muse retains settlement based on OpenRouter’s reported actual cost plus the platform margin. AA cost per task is an independently weighted upstream metric with a different pricing basis.
Available through this service
Both use the shared multimodal playground and chat API, so you can try each with the same source material.
| Capability | Muse Wonder | Gemini 4 Argon |
|---|---|---|
| Provider context | 1,048,576 tokens | 1,048,576 tokens |
| Accepted inputs | Text, images, video, audio, PDF | Text, images, video, audio, PDF |
| Output | Text | Text |
| Platform response limit | 8,192 tokens | 8,192 tokens |
| Chat API | /v1/chat/completions | /v1/chat/completions |
File-count, upload-size and response limits belong to this service and should be read separately from provider specifications.
Conditions to keep with the scores
Identify the version and reasoning setting
The main table compares ordinary Muse Wonder (Max) and Gemini 4 Argon (High). Muse Wonder Full is a separate version. Max and High are model-specific evaluation settings and do not imply equal compute budgets.
Understand what the index covers
The AA Intelligence Index primarily evaluates English text tasks. Image, audio, video and multilingual evaluations are separate; the index does not directly measure all those capabilities.
Interpret gaps at the task level
Published point estimates of 52 and 53 do not establish statistical significance. Original units and sources are retained; different metrics are not summed, and small gaps are not treated as reliable wins.
Match the execution environment
The playground provides non-streaming text and media chat with an 8,192-token response cap. It does not expose agent tool execution or adjustable reasoning effort. Reproducing agent benchmarks requires the appropriate harness, reasoning settings and output budget. This page reviews public results and does not claim a benchmark run by this service.
Primary sources and verification
Sources checked 2026-10-14 · public evaluation snapshot
Composite index, ten component results, reasoning settings and weighted task costs.
Index version, weights, coverage, scoring units and execution conditions.
Reported DeepSWE, Chartography and LVBench results and model capabilities.
Ordinary Muse versus Full, the reported DeepSWE value and list prices.
Settings and evaluation methodology, including attribution of DeepSWE results to Datacurve.
The model ID, supported inputs and published prices for this integration.
The model ID, supported inputs and promotional Standard-route prices.