Model buying decision · Source-based comparison

Muse Wonder vs Opus 5.5, Gemini 4 Argon and GPT-6.1 Sol

Use GPT-6.1 Sol Max as the low-cost benchmark candidate, Opus 5.5 Max with default fallback as the higher-score candidate, and Muse Wonder Max as an alternative to test on your own agent workload. Argon High belongs on the comparison, but verify access first. Pick by cost per accepted task and your latency limit—not token price alone.

AI Tool Finder Editorial Team · Official documentation reviewed October 13, 2026

This is a documentation and independent-benchmark comparison. We made no model API calls and ran no coding-agent or production task tests. Recommendations below are editorial starting points for a separately authorized evaluation.

Compare the same benchmark and the exact effort setting

The following snapshot uses Artificial Analysis Intelligence Index v4.3.2, a weighted composite of ten evaluations. Scores are index points, not a percentage of your business tasks completed. The displayed Cost per Task is the weighted benchmark cost—not a flat charge for one answer, one API request or one completed application workflow.

Swipe horizontally on a phone to see every column.

AA snapshot checked October 13, 2026 · USD per weighted Intelligence Index task
Evaluated modelEffort / configurationIndex score ↑Cost per Task ↓Decision constraint
Muse WonderMax52$4.52Wonder API; not Wonder Full
Claude Opus 5.5Max, Default Fallback58$5.98Keep the fallback configuration attached to this result
Gemini 4 ArgonHigh53$1.99Phased access; published introductory tariff
GPT-6.1 SolMax52$0.72Do not apply this result to another effort level

Sources: AA Wonder vs Sol, Wonder vs Argon, Opus 5.5 Max with default fallback and leaderboard. Configuration labels are part of the evidence.

What this says—and what it leaves open

Wonder and Sol have the same rounded aggregate score here; that does not make their answers interchangeable. A repository repair, a spreadsheet extraction and a long research run can reward different strengths. Opus has the highest aggregate score of these four, while Sol has the lowest displayed task cost. Neither observation establishes a winner for every task.

AA weights each evaluation's input, cache, reasoning and answer-token costs by its contribution to the index. The result changes with workload composition and effort. Do not divide the headline score into the price to invent a “cost per successful task”: an index point is not a success probability.

Meta's launch report emphasizes long-running agents, coding and computer use, and includes vendor-presented evaluations. Those claims are useful hypotheses, separate from the independently measured AA table and from any results your team obtains. This page does not mix Harvey LAB, Harvey LAB-AA or legal-specialist task costs into the general-index ranking.

API token prices are a different table

USD per one million text tokens, checked October 13. These are Standard or stated introductory prices; they are not the benchmark's per-task charges.

Published token rates and purchasing conditions
ModelUncached inputCache hitOutputExtra condition
Muse Wonder
muse-wonder-1.0
$4$0.20$20Standard only on Model API; search queries billed separately
Claude Opus 5.5$4$0.20$20Cache writes: $5 (5 min), $8 (1 hour). Fast mode is a separate premium.
Gemini 4 Argon$2 introductory$0.10 introductory$10 introductoryAnnounced launch rates; access and end of introductory period need confirmation
GPT-6.1 Sol$2$0.10$10Published cache-write rate $2.50; tool charges may be additional

First-party sources: Meta pricing, Anthropic pricing, Google's Argon launch and price footnote, OpenAI's Sol model page.

Google's announcement says Argon will move to $4 input / $20 output after the introductory period. We did not verify a universal self-service entitlement or a fixed end date. A published price and an AA evaluation do not grant access; confirm your account's model availability and applicable rate before budgeting.

Why equal token prices do not produce equal task costs

Wonder and Opus share the listed $4/$20 input/output rates, yet their AA task costs differ. Token usage, cache behavior and evaluation configuration matter. A model that explores more branches can spend more before emitting a short final answer. The visible answer length is therefore a poor proxy for the bill. Conversely, a more expensive attempt can be cheaper to operate if it materially reduces retries and human correction in your actual workload.

Use our existing Claude API pricing guide for Claude-specific billing choices and Opus Fast vs Sol Ultrafast comparison for speed tiers. Their buying questions are separate from this same-index task-cost comparison.

Choose the evaluation by the work you need to finish

1. Short responses

For classification, extraction or a brief answer, start with the cheapest accessible candidate that can satisfy your schema and accuracy threshold. Sol is the low-cost candidate in this snapshot, but Max may be unnecessary. Evaluate the exact lower effort you intend to deploy; do not carry its Max score into that decision.

Measure: valid outputs, correctness, time to first usable answer and cost per accepted item.

2. Long-running agents

Compare Wonder and your current accessible baseline on the same tool permissions, timeout and task history. Include repeated context, tool results, cache misses and recovery attempts. An agent that runs for longer without asking for help has not necessarily completed the job.

Measure: verified completion, failed or duplicate tool actions, total wall-clock time and total spend.

3. Complex code work

Include Opus Max with its documented fallback and Wonder Max when task value justifies testing a more expensive route. Use held-out repository changes with existing tests and review criteria. Add Argon only after access is confirmed. The aggregate AA lead is a reason to test, not proof that a patch is correct.

Measure: accepted patches, regressions, review time and cost including failed attempts.

Token speed is not task latency

Output tokens per second describes a portion of generation. Time to first token, time to first answer token, tool waits, network overhead and retries are different measurements. AA's weighted time-per-task estimate excludes first-token and other overhead; it should not become your end-to-end SLA. For an interactive feature, measure from user action to usable result. For an overnight agent, measure to verified completion and record timeout failures.

A reusable cost-per-accepted-task estimate

Keep token buckets mutually exclusive. Let U be uncached input, H cache-hit input, W separately billed cache writes and O billable output, including reasoning where the provider bills it as output. Rates below are per million tokens.

Attempt cost = (U × input rate + H × cache-hit rate + W × cache-write rate + O × output rate) / 1,000,000 + tool/service charges

Cost per accepted task = (all attempt costs + runtime costs + review/rework cost) / number of accepted tasks

Follow each provider's usage fields: do not charge a cache-write token again as ordinary input, or count reasoning twice when it is already included in output. If no tasks are accepted, cost per accepted task is undefined; the spend still occurred. Monthly spend also depends on frequency: a cheap attempt repeated across thousands of jobs can dominate the budget.

Illustrative Wonder budget—not a measured run

Assume one attempt uses 20,000 uncached input tokens, 40,000 cache-hit tokens and 8,000 total billable output tokens. The token subtotal is $0.248: $0.08 + $0.008 + $0.16. If that attempt also uses ten Meta web-search queries at the documented $2.50 per 1,000 queries, it becomes $0.273, before other applicable costs.

For 100 requested tasks plus 20 additional full-size retry attempts, that hypothetical spend is $32.76. If only 80 deliverables pass review, the model-and-search cost is about $0.41 per accepted task. Human review, hosting and taxes remain outside this example. Real retry sizes and cache hits vary; these assumptions are not forecasts.

Log returned usage instead of assuming a repeated prompt hit the cache. Count every billed partial or failed attempt, verify how rejected requests are treated, cap retries and stop when another attempt is unlikely to improve the outcome. Provider errors and application rejection are different events; do not assume every unsuccessful attempt is free.

Buy the right access—and check what happens to your data

Meta Model API versus Muse Code

Meta's model catalog separates muse-wonder-1.0 from muse-wonder-full-1.0. Wonder Full is a different model, available through the Power Usage and Maximum Usage Muse Code subscriptions. It is not the “Max” effort setting used for Wonder in the AA table, and it does not offer the same public per-token Model API route.

The Muse Code subscription documentation ties the allowance to the credential connected through CLI onboarding. Additional Model API keys are pay-as-you-go. Plans have usage limits and resets; a subscription is not unlimited general API access. Power and Maximum include Wonder Full, subject to eligibility and the limits shown during onboarding.

Standard and Contributor modes

Wonder on Model API uses Standard mode; Contributor is not available there. Meta says Standard-mode content is not used to train its models. Muse Code subscriptions, however, default to Contributor mode: prompts, code, files, context and responses may be used for training in exchange for more usage. The subscription setting can be turned off for future content. Check the actual setting before sending a repository or business document.

“Not used for training” is not a promise of zero retention or approval to send confidential data. Your team still needs permission to transmit the material, acceptable retention and processing terms, regional availability and appropriate access controls. Redact or use synthetic material until those conditions are clear. Do not transfer a privacy promise between an API, a consumer product and a third-party gateway.

Make a decision your team can reproduce

  1. Confirm access and constraints. Record provider, model ID, effort, fallback, data mode, price tier and a spend cap.
  2. Define acceptance before comparing. Use the same representative tasks, tool access and success criteria. Include hard failures and timeouts.
  3. Count all attempts. Record cache categories, reasoning/output usage, paid tools, retries and review time alongside the final artifact.
  4. Choose the lowest-cost option that clears your quality and latency thresholds. Route exceptions to a more capable or slower candidate only when the observed improvement justifies the extra cost.

This is an evaluation plan, not a test we have run. For packaged developer workflows, continue to our coding assistants guide; for tool orchestration and deployment choices, use the agent tools guide.

Evidence date and remaining limits

Sources were read October 13, 2026. The AA table records v4.3.2 and its displayed configurations; live scores, tariffs and availability can change. The figures are not AI Tool Finder scores. No task-level production success rate, confidential-data approval, account entitlement or paid-model result was verified.

Primary technical sources are linked beside the relevant claims. Google launch pricing is conditional; Wonder Full is excluded from the four-model cost ranking. No cross-suite legal benchmark costs have been combined with this table.