Try Muse Ember on repeat tasks, but do not budget from its benchmark bill

Try Muse Ember on repeat tasks, but do not budget from its benchmark bill

Muse Ember deserves a place in the test queue if repeated agent subtasks are getting expensive. In the November logic dataset from llm2014/llm_benchmark, it posts a slightly higher median score than Qwen3.8-Max and Step 5 Preview while recording lower test costs and shorter average runtimes. That makes it a candidate to evaluate, not a ready-made budget for your application.

Access includes the Responses API

In its November 9 announcement, Meta says Ember is available through Muse Code and the Meta Model API under the model ID muse-ember-1.0, with support for the Responses API. For prompts up to 200K tokens, the listed rates per million tokens are $0.10 for input, $0.50 for output and $0.004 for cached input.

Longer prompts cost $0.30 for input, $1.50 for output and $0.012 for cached input, again per million tokens. A single blended token price would hide that distinction. Nor should a benchmark's total cost in yuan be treated as the price of one production request. Check prompt length and then inspect actual billed usage.

A useful shortlist, not a replacement order

The published data file lists Muse Ember at xhigh with a median score of 42.99, a test cost of CNY 6.76 and an average runtime of 131 seconds. Qwen3.8-Max at xhigh records 41.71, CNY 72.87 and 1655 seconds; Step 5 Preview records 41.30, CNY 30.02 and 1307 seconds. These figures come from the current data snapshot cited here, not an archived leaderboard from the November 9 launch.

Our recommendation is to test bounded work with an obvious acceptance check: extraction, routing or a targeted code change. Keep your existing model as the control. Similar logic scores do not establish equal success rates on your tasks.

The benchmark repository describes a private, rotating question bank kept within 28 questions and 270 cases. It prefers intermediary services with official providers selected. Measure latency on your own endpoint and check its supported reasoning parameters rather than copying the table's xhigh label directly.

Hold the task, acceptance criteria and tool permissions constant. Track quality, elapsed time, total charges and retries together. A cheaper attempt that needs repeated repair may not save anything. If your existing model remains better at planning, replace only the execution subtask first; that creates a smaller, more testable migration than switching the whole workflow.

Share this article