Cursor Adds Meta Muse Wonder with 46.1% Record
Cursor integrates Meta's Muse Wonder reasoning model, hitting 46.1% on CursorBench with a 1M token window and $20 per million output tokens.
Meta’s reasoning model lands in Cursor with a 1-million-token window, direct API pricing, and a new benchmark high for the Muse family.
Code editor Cursor rolled out native integration for Muse Wonder on Thursday (Oct. 15), bringing Meta Superintelligence Labs’ frontier reasoning model into its developer environment. Running at standard reasoning effort on CursorBench 4.0, Muse Wonder scored 46.1%. That marks the highest benchmark result recorded by any model in Meta’s Muse lineup inside the editor.
The rollout arrived just three days after Meta initially unveiled the architecture. For Anysphere, the startup behind Cursor, the rapid addition maintains parity against developer tools built around competing engines from OpenAI, Anthropic, and xAI.
CursorBench performance
The 46.1% score moves Muse Wonder past the prior family record in the editor, previously held by Muse Spark 1.3 at 41.6% under its maximum effort setting. CursorBench evaluates coding agents against real developer workloads, testing multi-file edits, complex codebase refactoring, and project-wide debugging.
External benchmarks show additional headroom. Independent evaluations released on X by LuminaBench indicate that Muse Wonder climbs to 50.4% on CursorBench when pushed to its maximum reasoning effort setting (Max effort). The numbers place Meta’s newest release in direct contention with premier agentic coding models.
Pricing and context limits
Cursor mirrors Meta’s official API pricing with zero platform markup. Calls run at $4.00 per million input tokens, $20.00 per million output tokens, and $0.20 per million cached tokens. Anysphere does not charge surcharges for running queries across extended contexts.
In day-to-day use, the model supports contexts ranging from 300,000 tokens up to 1 million tokens. Operating inside Cursor’s agent harness, Muse Wonder can issue terminal commands, retrieve web pages, run automated browser tests, and inspect uploaded screenshots.
The Full version and BYOK limits
Meta launched the Muse Wonder family on October 12 under Alexandr Wang at the newly established Meta Superintelligence Labs. However, the release structure splits capabilities between commercial tiers.
Only the baseline Muse Wonder model is accessible through the public API and third-party tools like Cursor. The more capable Muse Wonder Full variant remains locked behind paid subscriptions to Muse Code, Meta’s proprietary coding environment.
Community requests for bring-your-own-key (BYOK) support for Meta accounts in Cursor also remain unaddressed. For now, that capability is labeled as RUMOR: Anysphere has provided no official confirmation or engineering timeline for third-party Meta keys.
Running Muse Wonder in Cursor
At $20 per million output tokens, Muse Wonder matches the going market rate for heavy reasoning models. That price point demands intentional usage. Firing off the model for basic boilerplate or quick syntax edits makes little financial sense.
The economics work best on stubborn engineering tasks. Tracing elusive bugs across distributed services, restructuring legacy modules, and parsing massive failure logs make good targets. When the model untangles a complex architectural refactor in a single prompt run, the saved engineering hours easily justify the token invoice.