Evaluating Muse Wonder for Coding Across Software Workflows

On October 12, 2026, Meta introduced the Muse Wonder series, presenting Muse Wonder and Muse Wonder Full as frontier models designed for complex agentic workflows, multi-file software engineering, and multi-step execution. Meta built both models on larger foundations than its previous Muse Spark 1.3 releases and trained them at Meta's Prometheus data center. Evaluating Muse Wonder for coding reveals how Meta positions these 1,000,000-token context models for developers managing large repositories and automated tool environments.

Meta contrasts Muse Wonder with its lighter Muse Spark line and standard code completion assistants that operate strictly within single-file buffers. Unlike tools that halt when a compiler error occurs, Muse Wonder uses an architectural routine named Active Focus to retry failed commands, test alternative syntax paths, and execute parallel edits across multiple files in a repository. In Meta's published benchmarks, Muse Wonder Full achieved a score of 79.0 on DeepSWE 1.1 and 57.5 on the extended SWE-Marathon evaluation, exceeding the reported results of several general-purpose foundation models on long-horizon software engineering benchmarks.

For engineering teams and technical leads deciding on model infrastructure, this release introduces a concrete split between Application Programming Interface (API) integration and hosted developer tooling. Muse Wonder is accessible via the Meta Model API and OpenRouter for custom pipelines, while the higher-capacity Muse Wonder Full model is restricted to proprietary Muse Code subscriptions starting at $50.00 per month. Understanding where these models succeed and where they fail helps engineering organizations decide whether to alter existing developer setups.

Benchmark Results for Muse Wonder for Coding Tasks

Meta's published benchmark results position Muse Wonder and Muse Wonder Full as competitive choices for long-horizon engineering, though rival models retain leads on specific programming tests.

On the DeepSWE 1.1 benchmark, which tests long-horizon software engineering, Muse Wonder Full scored 79.0 and Muse Wonder scored 78.8. These scores sit above Anthropic's Claude Opus 5.5 at 74.2, Claude Fable 5.1 at 67.4, and OpenAI's GPT-6 Astra at 74.1. On the ultra-long-horizon SWE-Marathon benchmark, evaluated at extra-high reasoning effort, Muse Wonder Full reached 57.5 and Muse Wonder achieved 53.1. Claude Opus 5.5 scored 46.3, and GPT-6 Astra scored 42.5 on that same test.

Standard coding benchmarks show a different distribution of capabilities. On ProgramBench, which evaluates from-scratch programming and standalone script generation, Claude Opus 5.5 scored 70.5, while Muse Wonder Full scored 59.0 and Muse Wonder recorded 50.0. On Terminal-Bench 4.0-AA, measuring command-line agent operations, Claude Opus 5.5 scored 59.6 and GPT-6 Astra scored 59.1, outpacing Muse Wonder Full at 57.1 and Muse Wonder at 52.0. On the composite Artificial Analysis Index, Claude Opus 5.5 leads at 58, compared to 56 for Muse Wonder Full and 52 for Muse Wonder. Meta noted that its internal numbers reflect runs at maximum reasoning effort, while external teams have yet to publish independent verification across these suites.

• DeepSWE 1.1: Muse Wonder Full scored 79.0, Muse Wonder scored 78.8, Claude Opus 5.5 scored 74.2, and GPT-6 Astra scored 74.1.
• SWE-Marathon: Muse Wonder Full reached 57.5 at maximum reasoning, Muse Wonder reached 53.1, Claude Opus 5.5 reached 46.3, and GPT-6 Astra reached 42.5.
• ProgramBench: Claude Opus 5.5 leads at 70.5, Muse Wonder Full reached 59.0, and Muse Wonder reached 50.0.
• Terminal-Bench 4.0-AA: Claude Opus 5.5 leads at 59.6, GPT-6 Astra reached 59.1, Muse Wonder Full reached 57.1, and Muse Wonder reached 52.0.

Using Muse Wonder for Coding Across Large Repositories

Large-repository reasoning in Muse Wonder relies on a 1,000,000-token context window combined with Active Focus routines that handle multi-file refactoring.

When developers point Muse Wonder at extensive codebases, the model ingests broad dependency trees, configuration manifests, and documentation files simultaneously. In typical agentic coding setups, context limits force tools to prune files aggressively, causing models to miss interface definitions located in separate modules. The 1,000,000-token window allows Muse Wonder to maintain references to internal libraries, database schema definitions, and shared utilities across lengthy troubleshooting sessions.

Active Focus functions as an iterative execution loop rather than a single generation step. If an automated test fails after a file modification, Muse Wonder inspects the terminal output, revises the modified code, and re-executes the test suite without terminating the task. Meta also designed the model to ask for clarification before completing irreversible operations, such as dropping database tables or overwriting remote branches. For engineering teams maintaining legacy systems, this persistent execution model reduces the need for manual prompt restarts during repetitive refactoring jobs.

Developer Tooling Integration and Pricing Structure

Access to Muse Wonder differs between direct API consumers and subscribers to Meta's hosted developer environments.

The standard Muse Wonder model, designated as muse-wonder-1.0, is available through the Meta Model API and OpenRouter. The Meta Model API maintains compatibility with the OpenAI Software Development Kit (SDK), allowing teams to substitute existing client libraries by updating endpoint URLs and authorization headers. Pricing for Muse Wonder sits at $4.00 per million input tokens, $0.20 per million cached input tokens, and $20.00 per million output tokens. Context caching discounts lower the operating cost for teams that pass static codebase indexes repeatedly.

Muse Wonder Full, designated as muse-wonder-full-1.0, is excluded from the pay-per-token API. Meta reserves Muse Wonder Full exclusively for subscribers to its Muse Code developer environment. Teams accessing Muse Wonder Full must purchase either the Power Usage plan at $50.00 per month or the Maximum Usage plan at $200.00 per month. The Power Usage plan provides 20 times the usage volume of Meta's Everyday tier along with higher file upload limits, while the Maximum Usage plan provides 80 times the base allocation. Engineering managers cannot integrate Muse Wonder Full into external Integrated Development Environments (IDEs) through generic API keys.

• Muse Wonder API: Costs $4.00 per million input tokens, $0.20 per million cached tokens, and $20.00 per million output tokens.
• Muse Wonder Full Availability: Available only through Muse Code subscriptions, not via public API endpoints.
• Muse Code Power Usage: Costs $50.00 per month for 20 times baseline usage allocations and expanded file handling.
• Muse Code Maximum Usage: Costs $200.00 per month for 80 times baseline usage allocations.
• SDK Compatibility: The Meta Model API conforms to OpenAI client standards for simplified migration.

When Muse Wonder for Coding Fails and Alternatives to Pick

Muse Wonder exhibits distinct operational constraints in standalone script synthesis, proprietary tooling lock-in, and independent verification.

The model's lower ProgramBench score indicates that Muse Wonder struggles more than Claude Opus 5.5 when generating functional code from scratch without surrounding repository context. Teams requiring quick, isolated scripts or algorithm generation often receive cleaner initial drafts from rival models. Furthermore, withholding Muse Wonder Full from the public API creates vendor lock-in, forcing engineering organizations to adopt Meta's proprietary Muse Code interface rather than using their established IDE extensions or internal developer portals.

Engineering organizations requiring standalone code generation or flexible API pipelines should not select Muse Wonder Full, choosing instead Claude Opus 5.5 or GPT-6 Astra for custom development infrastructure. Independent third-party validation confirming Meta's SWE-Marathon numbers across broader open benchmarks, or Meta releasing Muse Wonder Full directly onto the pay-per-token API, would shift this assessment in Meta's favor.

Security Guidance and Review Standards for Generated Code

Shipping model-written code into production environments requires structured static analysis, automated linting, and human verification before deployment.

Meta trained Muse Wonder under the Meta Superintelligence Scaling Framework, implementing safeguards against Chemical, Biological, Radiological, and Nuclear (CBRN) risks and offensive cyber misuse while permitting defensive security tasks. However, these safety boundaries do not prevent common coding defects, such as introducing insecure dependencies, mishandling Cross-Site Scripting (XSS) sanitation, or misconfiguring Continuous Integration and Continuous Deployment (CI/CD) pipelines. Teams deploying model-generated pull requests must run automated vulnerability scanners and dependency checkers before merge approval.

Regulated organizations in financial and healthcare sectors must also establish strict data boundaries. While Meta states that Muse Wonder avoids sharing private information without permission, sending proprietary code to hosted endpoints carries operational risk. In automated codebase migrations and routine technical audits, unreviewed agentic diffs frequently introduce subtle logic drift or misconfigured environment variables. Audit your internal repository workflows against our verification checklist before deploying Muse Wonder for coding in production branches.