Muse Wonder Full Joins AI’s Frontier While Muse Wonder Gets Charged

🔍 Read the full analysis: Muse Wonder Full Joins AI’s Frontier While Muse Wonder Gets Charged on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Meta has released Muse Wonder and Muse Wonder Full, two proprietary models that mark a substantial benchmark improvement over Muse Spark 1.3, according to an Artificial Analysis evaluation published October 12, 2026. Wonder Full scored 56 on the Intelligence Index, while the model available through Meta’s API, Muse Wonder, scored 52; the evaluation estimated a Wonder task at $4.52, raising questions about its cost relative to similarly scoring alternatives.

Artificial Analysis said Muse Wonder Full improved on all 10 evaluations in its Intelligence Index v4.3. Its score of 56 was eight points above Muse Spark 1.3’s 48. Among the 27 configurations compared in the source report, only Claude Opus 5.5 at 58 scored higher, while Claude Sonnet 5.5 also scored 56. Muse Wonder Full ranked in the top three on six of the ten evaluations.

The results included a 57% Terminal-Bench 4.0 score, up from 33% for Muse Spark 1.3, and a 71% score on AutomationBench-AA. The report also listed GDPval-AA at 1,857 Elo, close to Claude Opus 5.5’s 1,866 and Claude Sonnet 5.5’s 1,838; it said those results were within the models’ confidence intervals. On the Coding Agents Index, Muse Code paired with Wonder Full scored 68, tying the reported Claude Code and Claude Sonnet 5.5 configuration.

Muse Wonder, the model customers can access through Meta’s API, scored 52 on the Intelligence Index. It reached 52% on Terminal-Bench and 68% on AutomationBench-AA. The evaluation estimated a task at $4.52 for Muse Wonder, compared with $0.72 for GPT-6.1 Sol, which also scored 52, and $1.99 for Gemini 4 Argon, which scored 53. These are benchmark task estimates, not a universal price for every use.

Strong Scores, Uneven Access and Cost

The release puts Meta’s models closer to the leading systems in the cited evaluations, but access to the strongest result is limited. Muse Wonder Full is available only inside Muse Code, Meta’s coding agent, and has no per-token price listed. Customers cannot compare a published token rate or calculate its cost per benchmark task from the source information.

By contrast, the API model has published rates of $4 per million input tokens and $20 per million output tokens, the same list prices reported for Claude Opus 5.5. Artificial Analysis’s task estimate highlights a practical purchasing issue: a model’s benchmark score is only one part of its value if running it costs more than alternatives with similar results. The estimate reflects the evaluation’s workload and should not be treated as a forecast for every customer’s usage.

Reliability also remains relevant for organizations considering the models for work where mistakes carry a cost. The report puts Muse Wonder Full’s hallucination rate at 23% and Muse Wonder’s at 27%, compared with 33% for Muse Spark 1.3. Those figures indicate improvement in the reported test, but do not establish how either model will perform in a particular deployment.

From Spark to Two Muse Models

The two models arrive six weeks after Muse Spark 1.3. Both are proprietary, accept text, images, speech or audio, and video as input, return text, and have a reported one-million-token context window. Their access and pricing differ: Wonder is offered through Meta’s API and in Muse Code, while Wonder Full is limited to Muse Code.

The distinction matters when interpreting the benchmark results. The Coding Agents Index measures a model together with its software harness, so the 68 score for Muse Code plus Wonder Full is not a direct comparison of Wonder Full alone against a different model in the same setup. The source report explicitly cautions that the result does not show how Wonder Full would perform inside Claude Code.

Artificial Analysis also reported a large input-token footprint for Muse Wonder in its Intelligence Index tasks: 7.69 million input tokens per task, compared with 6.03 million for Spark 1.3. Wonder’s output averaged about 54,000 tokens per task. The source says input accounted for 76% of the estimated $4.52 cost. These measurements help explain the evaluation bill, but task-specific token use may differ from ordinary customer workloads.

“Meta closed the capability gap. It did not close the price gap.”

— Thorsten Meyer, ThorstenMeyerAI.com

Pricing and Real-World Results

The source material does not give a per-token price for Muse Wonder Full, which is available only through Muse Code. It also does not provide enough detail to estimate what the model costs per coding task in that product. Meta’s pricing rationale for Muse Wonder, including whether rates may change, is not established in the material.

Benchmark scores are measurements under specific test conditions, not guarantees of performance across customer applications. The provided account does not include the full evaluation methodology, uncertainty ranges for every reported score, or independent results from everyday deployments. The task-cost comparison also depends on the benchmark’s token consumption and pricing assumptions. How those figures translate to other workloads remains unclear.

Watch API Access and Pricing

The immediate next questions are whether Meta will make Muse Wonder Full available beyond Muse Code and whether it will publish pricing that lets customers compare its operating costs. For Muse Wonder, buyers can review the listed token rates alongside workload-specific testing, rather than assuming the benchmark task estimate will match their own bills.

Further independent evaluations and deployment reports may show whether the reported gains hold across tasks beyond the current benchmark suite. Until then, the confirmed picture is a higher-scoring model with restricted access, alongside a broadly accessible API model whose benchmark score is lower and whose measured evaluation cost exceeded several cited rivals.

Key Questions

What are Muse Wonder and Muse Wonder Full?

They are two proprietary Meta models released six weeks after Muse Spark 1.3. Both accept text, images, audio or speech, and video as input, and have a reported one-million-token context window.

How did the models score?

Artificial Analysis scored Muse Wonder Full at 56 and Muse Wonder at 52 on its Intelligence Index v4.3. Muse Spark 1.3 scored 48 in the same reported comparison.

Can customers use Muse Wonder Full through Meta’s API?

Not according to the source report. Muse Wonder Full is available only in Muse Code; Muse Wonder is the model listed as available through Meta’s API and Muse Code.

How much does a Muse Wonder benchmark task cost?

Artificial Analysis estimated $4.52 per Intelligence Index task for Muse Wonder. That is a benchmark-specific estimate, not a fixed charge for every task or customer workload.

What remains unknown about the release?

The source does not state a per-token price for Muse Wonder Full or establish how the benchmark results will translate to real-world deployments. Meta’s future pricing and any broader access plans are also not specified.

Source: ThorstenMeyerAI.com