Meta Releases Muse Ember, the Cheapest Agent Model at $0.14 a Task
Meta's new efficient-tier model matches Claude Haiku 5.5 on intelligence at a third of the cost, with 1M context and native multimodal input.
- Meta released Muse Ember, its smallest reasoning model, matching Claude Haiku 5.5 intelligence at one-third the cost.
- Pricing: $0.10 input / $0.004 cached / $0.50 output per 1M tokens, available as muse-ember-1.0.
- 1M-token context window with native text, image, audio, and video input.
- First Meta model distilled from Muse Wonder Full, designed as a subagent paired with Wonder.
- Beats Haiku 5.5 on Terminal-Bench (40.9% vs 32.8%) and AutomationBench (68.1%), trails on FrontierCode.
- Weakness: 21.1% on AA-Omniscience, declining 64% of knowledge questions to avoid hallucinations.
Meta Superintelligence Labs has released Muse Ember, its lightest and least expensive Muse model. On the Artificial Analysis Intelligence Index, Ember scores 36 at its maximum reasoning setting, matching Claude Haiku 5.5. Artificial Analysis estimates a cost of $0.14 per index task for Ember and $0.41 for Haiku, a gap driven by cheaper cached input and shorter outputs.
Developers can access the model through the Meta Model API and Muse Code under the identifier muse-ember-1.0. It supports the Responses API, a 1 million-token context window, and native text, image, audio, and video input. Those capabilities allow agents to process meeting recordings, scanned pages, screenshots, and long document collections without a separate input model.
Cache and output length cut the bill
Meta lists standard input at $0.10 per million tokens, cached input at $0.004 per million tokens, and output at $0.50 per million tokens.
| Token type | Price per million tokens |
|---|---|
| Standard input | $0.10 |
| Cached input | $0.004 |
| Output | $0.50 |
The standard input and output rates match Claude Haiku 5.5 and GPT-6 Luna. Ember's cached-input rate is roughly 60% lower, according to Artificial Analysis. That discount can materially reduce costs for agents that repeatedly send the same system instructions, tool schemas, code, or conversation history.
Ember also generated about one-third as many output tokens per benchmark task as Haiku 5.5. Combined with the cache discount, that output efficiency produced the $0.14 estimated task cost. Models with similar Intelligence Index scores, including DeepSeek V4.1 Flash and GLM-5.3-Flash, cost about four times as much per task. Muse Wonder scores eight points higher at 44, while its estimated task cost is roughly 50 times Ember's.
Benchmarks favor automation and terminal work
Meta designed Ember for repetitive, bounded subtasks inside multi-model agents. Its benchmark results are strongest in automation, terminal execution, and coding-agent workflows, while long-horizon coding and factual recall remain weaker.
| Benchmark | Ember | Comparison |
|---|---|---|
| AutomationBench-AA | 68.1% | Level with Muse Wonder and GPT-6 Astra |
| Terminal-Bench 4.0 | 40.9% | Claude Haiku 5.5 scores 32.8% |
| FrontierCode 2.0 | 27.2% | Claude Haiku 5.5 scores 34.9% |
| Coding Agent Index | 46 in Muse Code | Haiku 5.5 scores 37 in Claude Code; GPT-6 Luna scores 31 in Codex |
| GDP.pdf | 22.8% | Claude Haiku 5.5 scores 20.8% |
The Coding Agent Index compares model-and-harness combinations, so the result reflects Muse Code as well as Ember. Teams considering a migration should reproduce the test with their own tools, prompts, approval rules, and repository structure.
Ember's 21.1% accuracy on AA-Omniscience is the lowest among models with similar overall Intelligence Index scores. It declined to answer 64% of questions, producing a reported hallucination rate of 17.4%, about half Haiku 5.5's rate. Workflows that can escalate refusals to a larger model may benefit from that conservative behavior, particularly when an incorrect tool call could modify data or trigger an external action.
Wonder teaches the worker tier
Ember is Meta's first model distilled from Muse Wonder Full. Distillation trains a smaller model to reproduce selected behavior from a larger teacher, trading some capability for lower inference cost and faster high-volume execution. Meta built that process around a two-tier agent architecture in which Wonder plans, sets rules, and decomposes work before Ember executes the resulting subtasks.
A customer-support system illustrates the intended pattern: Wonder can determine how an issue should be handled, then Ember can apply those rules across tens of thousands of cases. The same structure applies to document pipelines, test suites, code maintenance, and other workloads with a costly planning stage followed by many similar operations.
Route work by task shape
A practical deployment assigns Ember bounded tasks with clear success criteria and routes uncertain or knowledge-intensive work to a stronger model.
| Workload | Suggested role | Reason |
|---|---|---|
| Document parsing and extraction | Primary worker | High volume, repeated context, and multimodal input |
| Routing and classification | Primary worker with confidence thresholds | Bounded outputs and low per-task cost |
| Terminal automation and test execution | Primary worker with tool permissions | Strong Terminal-Bench result |
| Targeted code fixes | Worker under a larger model's plan | Competitive agent score, weaker long-horizon coding |
| Knowledge-heavy Q&A | Router or formatter | Low factual-recall accuracy and frequent refusals |
| Open-ended research | Supporting subagent | Broader synthesis requires stronger recall and reasoning |
Teams evaluating Ember as a replacement for Claude Haiku 5.5, GPT-6 Luna, or DeepSeek V4.1 Flash should measure task completion, tool-call validity, refusal handling, latency, cache hits, and total output tokens on production traces. Benchmark task costs depend on each test's prompts, token use, and caching pattern.
Vals AI reports a similar cost profile in its independent evaluation. Ember ranks 12th among 35 models on the Vals Index with a score of 60.48% and a cost of $0.15 per test, the lowest test cost on that index.
Comments
