Meta’s Muse Wonder arrives in Cursor

Cursor has added Meta’s Muse Wonder to its model picker, giving developers direct access to the agentic coding model inside the editor. According to Cursor’s announcement, Wonder scores 46.1% on CursorBench at the default reasoning setting, the best result yet for a Muse model.

Built by Meta Superintelligence Labs, the research organization led by former Scale AI CEO Alexandr Wang, Wonder is the API-accessible model in a two-model release. The larger Muse Wonder Full remains exclusive to Meta’s Muse Code product, while Cursor offers the version sold through token-based API access.

A premium model with a million-token window

Cursor’s pricing matches Meta’s direct API rates and sits in roughly the same tier as Claude Opus 5.5 and GPT-6 Sol. Output tokens cost five times as much as input tokens, so long patches, explanations, and repeated agent loops can drive the bill. The price favors deliberate agent runs over routine autocomplete.

A one-million-token context window defines how much source code, documentation, conversation history, and tool output the model can process in one session. That capacity can accommodate broad repository snapshots and long-running tasks, though file selection and retrieval quality still affect accuracy in large monorepos. The underlying model accepts multimodal input; the attachment types available during a Cursor workflow depend on the editor’s integration.

Muse’s best score still trails the leader

Within Meta’s Muse lineup, Wonder’s 46.1% CursorBench result sets the high mark. Across the full leaderboard, the current leader sits more than 11 percentage points ahead. The default reasoning designation means Cursor measured the model using its standard reasoning budget instead of a more compute-intensive setting.

Released on September 10, 2026, CursorBench 4.0 evaluates long-horizon coding work drawn from realistic software-development scenarios. Those tasks require models to navigate repositories, edit multiple files, use tools, run tests, and recover from failures over several steps.

Meta reports stronger relative performance on its own agentic coding evaluations:

Different task sets, scaffolding, reasoning budgets, and scoring methods can produce different benchmark rankings. Meta’s figures are vendor-published results, while CursorBench measures performance in Cursor’s own agent environment. Repository-specific evaluations remain the clearest basis for comparing success rate, latency, retries, and total token use.

Meta trains for memory and retries

Meta says Wonder was trained around two behaviors it calls active memory and active focus. Active memory aims to retain and synthesize information supplied by the user throughout a long session. Active focus encourages the model to pursue alternative approaches after a failed command, broken test, missing dependency, or other blocker.

During an agentic coding session, those behaviors target failures that often emerge several steps into a task. A model may produce a plausible first patch, then lose track of earlier edits or stop after a test failure. Wonder’s value in Cursor depends on whether it can preserve relevant context and continue debugging without repeated user intervention.

Tasks that justify the token bill

Muse Wonder’s rates favor work where its large context window and retry behavior can offset the higher cost, including:

• Long-running agent tasks across large repositories, where broad context reduces manual file selection.
• Debugging sessions that require repeated tool calls and recovery after failed tests.
• Cross-file migrations where the agent must track earlier edits and dependencies.
• Multimodal tasks that combine code with screenshots, design mockups, or diagrams.

Routine completions and small edits remain better matched to lower-cost options such as Muse Spark 1.3. Cursor also offers models with higher overall CursorBench scores. Teams choosing a default agent can compare task completion rate and total cost instead of relying on a single benchmark result.

Wonder moves behind hosted access

Wonder extends Meta’s shift from downloadable Llama releases to hosted frontier models. Earlier Llama models gave organizations public weights they could run on their own infrastructure. Both Wonder models use closed, hosted distribution, with Muse Wonder available through paid API channels and Muse Wonder Full confined to Muse Code. Organizations that require on-premises or offline inference therefore cannot deploy Wonder locally.

Following Llama 4’s launch and benchmark controversy in April 2025, Meta reorganized its AI work around Wang’s Meta Superintelligence Labs. Cursor now provides metered access to the resulting model without requiring a separate Meta subscription. The integration gives developers a practical way to measure Wonder against their own repositories, where completion rate, persistence, latency, and token cost will determine whether it earns a place in the default agent configuration.