Meta’s New Muse Wonder AI Model Is Closing The Gap to Anthropic and OpenAI

Meta has released Muse Wonder and Muse Wonder Full, AI models that bring developers stronger coding-test results than its previous Muse Spark model. Independent evaluator Artificial Analysis measured gains in fixing software projects and in terminal tasks that use text commands to control software, with each model powering Meta’s coding assistant. The more accessible Wonder version took longer per tested task.

Released on October 12, Wonder is available in preview through Meta Model API, which lets developers use the model in their own applications, and Muse Code, its ready-made coding assistant. Wonder Full is a separate, higher-capability model restricted to Muse Code’s Power Usage and Maximum Usage subscriptions.

Muse Code can plan a job, edit files and run commands from a developer’s terminal or an automated software workflow. Artificial Analysis tested complete coding systems, pairing each model with the software that lets it act on a project.

Coding Gains Across Different Software Tasks

With reasoning effort set to maximum, Muse Code powered by Wonder scored 63 on Artificial Analysis’s Coding Agent Index, up from 54 with Muse Spark 1.3. Wonder Full reached 68, tying Anthropic’s Claude Code running the Sonnet 5.5 model at maximum effort. Claude Code with Opus 5.5 scored 66 in the same suite.

The coding index gives equal weight to three kinds of work: implementing changes in software repositories, completing terminal tasks and answering repository questions in SWE-Atlas-QnA. Each component averages single-attempt task success across three runs per task. The index summarizes those results; the component percentages show where performance changed.

The largest improvement for ordinary Wonder came on Terminal-Bench 4.0, where Muse Code’s success rate rose from 32% with Spark to 47%. The improvement was smaller on repository implementation, tested by DeepSWE v1.1. Full pulled further ahead chiefly through terminal work.

What Meta’s Muse Models are Good At

The new family extends the coding role of Muse Spark, launched in April. That earlier model could generate websites and mini-games from prompts. Meta positions Wonder for more complex work across multiple files and for agents that keep working on a task over hours.

Meta describes two functions intended to support that persistence. Active Memory preserves important context across conversations and retrieves it when relevant. Active Focus keeps the model working toward a goal, trying another approach when it encounters a blocker and seeking clarification when instructions are ambiguous.

An agent that loses track of earlier information or its intended goal can make mistakes that compound as it keeps working.

Meta’s internal MemoryBench tests recall in conversations containing 64,000 to 100,000 tokens, the units used to process text. It includes distracting information, conflicts and updated facts, with another language model judging the responses. Meta’s model card recommends human confirmation before irreversible actions.

Competitive Progress Varies by Task

At maximum reasoning effort, Artificial Analysis’s separate Intelligence Index put Wonder at 52 and Full at 56, versus Spark 1.3 at 48. That weighted composite, on a scale from 0 to 100, covers agent tasks, coding, general capability and scientific reasoning.

Another independent evaluator, Vals AI, found a particularly large mathematical gain. Wonder Full at maximum effort solved 94 of 100 problems on its private ProofBench mathematical-proof test, compared with 58 for Spark 1.3.

In Vals’s three finance and legal tests, Full’s scores slipped by up to two points from its predecessor while remaining in the top five. Its mathematical-proof gain came alongside largely unchanged results on those professional tasks.

Higher Scores Take More Time and Money

In the coding suite, Wonder averaged $8.05 in API charges per task, compared with Spark’s $3.98. Its active runtime also increased, from 18.4 to 22.4 minutes per task. Full averaged 28.8 minutes; a comparable API cost was unavailable. Artificial Analysis’s runtime measure excludes startup and grading overhead.

The listed rates help explain why using a stronger model can cost more even when it generates less text. Wonder’s Standard API prices are $4 per million input tokens, $0.20 for cached input and $20 for output. Spark’s corresponding rates are $1.25, $0.15 and $4.25. Cached input is previously supplied text that the service reuses. The actual charge depends on the mix of input, reused input and generated output in a job.

Artificial Analysis measured roughly 54,000 output tokens per task for each Wonder variant in its separate Intelligence Index workload, versus Spark’s 60,000. Ordinary Wonder nevertheless cost $4.52 per task in that suite, compared with Spark’s $1.60. Fewer output tokens therefore coexisted with a higher measured bill under the new rates.

Paid access also carries a data choice. Muse Code subscriptions default to Contributor mode, which gives more usage allowance in exchange for permission to train on submitted content. Users can turn it off, reducing their allowance. Wonder has no discounted Contributor variant for API access; separate API keys remain billed on a pay-as-you-go basis. Subscription benefits and availability vary by region.

Wonder’s integration into the broader Muse app is still forthcoming, with subscribers due to receive it first.