Infrastructure · NOVEMBER 12, 2026
Muse Ember cuts benchmark task costs; customer-acquisition gains remain untested
Meta’s lightweight model approaches Haiku 5.5’s benchmark score at about one-third the cost per task, but prospecting economics still need production testing.
Meta released Muse Ember on November 9 for repeated agent work, with Artificial Analysis measuring about $0.14 per Intelligence Index task against $0.41 for Claude Haiku 5.5 at maximum reasoning effort.
Both models round to 36 points, though Meta’s underlying chart puts Ember at 35.8 and Haiku at 36.2. That’s a cost comparison at similar aggregate scores, not evidence of equivalent performance on every task.
Token consumption changes the cost comparison
Meta’s announcement, “Introducing Our Smallest, Most Efficient Model: Muse Ember,” describes an execution engine for tool calls, extraction, routing and document parsing. TechCrunch frames it as the lighter model beneath a larger planner.
Ember has a one-million-token context window. Meta lists these rates per million tokens:
| Prompt length | Input | Cached input | Output | | --- | ---: | ---: | ---: | | Up to 200,000 tokens | $0.10 | $0.004 | $0.50 | | Above 200,000 tokens | $0.30 | $0.012 | $1.50 |
Artificial Analysis says Ember shares Haiku’s starting input and output prices, but cached input is 60% cheaper and Ember uses roughly one-third as many output tokens at maximum effort. TechCrunch instead describes Ember’s starting input rate as cheaper than Haiku’s; those pricing characterizations conflict.
Separately, Vals AI reports 60.48% on the Vals Index at $0.15 per test, the lowest test cost among models on that index.
Completed objectives aren’t completed workflows
Meta reports 37.8% strict first-attempt completion on AutomationBench-AA. Artificial Analysis reports 68.1% on AutomationBench-AA. These figures shouldn’t be treated as interchangeable: VentureBeat explains that its headline measure credits individual objectives completed within the benchmark’s rules, rather than necessarily finished workflows.
Knowledge is another limitation. Artificial Analysis identifies it as Ember’s weakness, reporting 21.1% accuracy on AA-Omniscience, with 64% of questions declined.
VentureBeat recommends measuring accepted results with retries and escalations included. Its article puts the operator takeaway this way, without identifying a speaker:
"Those designs let teams measure cost-per-accepted result, including retries and escalations, alongside latency and how often context pushes requests into the higher price tier."
Prospecting needs a production denominator
The customer-work market runs from model platforms such as Meta’s Muse Ember to small-business services such as LemonLime, which finds prospective customers and prepares personalized outreach. That category connection doesn’t establish that the service uses Ember.
For an owner or founder assessing repeated prospect research and outreach preparation, the $0.14 benchmark result suggests a reason to test, not a production budget. A bounded, human-reviewed trial could measure verified prospect fits and usable drafts against research, retries, escalation and review costs. A person should check the work before sending.
If comparable economics carry into those workflows, small teams could have more room to run them regularly. But the evaluations don’t establish customer-acquisition gains, customer wins or lower finished-software prices. The practical comparison remains cost per accepted result, including the work needed to verify it.