Meta unveils Muse Glimmer for single-GPU use

Meta Muse Code, its first AI coding agent, was released on August 5. Priced at $1.25 per million input tokens and $4.25 per million output, it is touted as more affordable than both Claude and Codex. Anthropic's Fable 5 model, used in Claude Code, is priced higher at $10 per million input tokens and $50 per million output. A "contributor" tier at Meta is even more budget-friendly, at $0.10 and $0.20 per million input and output tokens respectively, though this means users' content could be used for product improvement.

To test the affordability versus quality claim, Muse Code was compared against Claude Code in three coding tasks: a debugging task, a multi-file refactor, and an open-ended reasoning task. The tests used the same prompts and a fresh session for each test to ensure the agents were the only variable.

In the bug hunt task, both agents successfully found and fixed the planted bugs without altering any test files or affecting public behavior. However, when the same test was run with timezone assumptions baked into two unrelated tests, Claude Code managed to make the suite pass without touching the test files, while Muse Code identified the issues but did not correct them. The Muse Code test ran for 2 minutes and 58 seconds, costing 2 cents, while Claude Code took 5 minutes and 31 seconds, costing $3.02.

The refactor task required both agents to improve the readability and structure of a 467-line JavaScript file without changing the library's behavior or public API. Muse Code made minor changes, extracting one new module and breaking the large methods into smaller helpers within the original file. Claude Code, on the other hand, extracted three modules and significantly reduced the main file's length from 467 lines to 353.

Both agents successfully passed all 773 tests and had no lint errors. The Muse Code refactor was completed in 1 minute and 44 seconds for a penny, while Claude Code took 6 minutes and 9 seconds for $2.40.

In the reasoning task, both agents had to investigate why the library's test suite failed in certain time zones and decide on the best fix. While both agents successfully fixed the issue, Claude Code managed to make the full suite pass in any timezone without weakening test coverage and maintaining the library's public behavior.

The Muse Code solution left behind some failed tests, while the Claude Code solution provided a more comprehensive solution. Both tests were completed in under 10 minutes, with Muse Code taking 1 minute 44 seconds and Claude Code taking 6 minutes and 9 seconds.

In conclusion, Meta Muse Code is indeed cheaper than Anthropic's Fable 5 model used in Claude Code. However, the quality of the results produced by each AI agent varies. While both successfully completed the testing tasks, Claude Code demonstrated a more thorough approach, resulting in a more robust solution. The decision between the two ultimately rests on the balance between cost and quality that each user is willing to accept.