Meta's cheapest AI model matches Anthropic's on an independent ranking at about a third of the cost per task

Meta released Muse Ember on Monday, its smallest and cheapest AI model. An independent testing firm says it matches Anthropic's comparable model for about a third of the cost per task.

That matters to any company running AI agents, software that carries out multi-step jobs on its own. Meta pitches Ember at the repetitive steps in that work, where small savings per task add up quickly.

The savings do not come from a lower price tag. Ember lists at $0.10 per million input tokens and $0.50 per million output tokens. Tokens are the word fragments AI companies bill by. Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna list at exactly the same rates.

Same price tag, smaller bill

Artificial Analysis, which runs standardized tests across AI models, scored Ember 36 on its Intelligence Index at maximum effort. That ties Haiku 5.5. Each test task cost about $0.14 with Ember and $0.41 with Haiku. GPT-6 Luna cost about the same as Ember, $0.15 per task, but scored only 28.

Most of the gap comes from how much each model writes. Artificial Analysis says Ember used roughly a third as many output tokens as Haiku. OfficeChai, citing its data, puts that at about 84,000 tokens per task versus about 226,000.

The second lever is repeated input. Agents resend the same instructions and documents again and again, and providers charge less for this "cached" input. Ember charges $0.004 per million cached tokens, versus $0.01 for Haiku and Luna. Artificial Analysis says cached input makes up most of the cost in its agent tests.

Running Haiku at a lower effort setting is cheaper still, about $0.09 per task. But its score then drops to roughly 25, according to a review by eesel, which sells AI support software.

Meta's own lineup shows how steep the trade-off is. Its larger Muse Spark 1.3 scored 37, one point above Ember, at $2.39 per task. Its top model, Muse Wonder, scored 44 at about $6.95 per task, roughly 50 times Ember's cost. Meta's pitch is to pair them, with Muse Wonder making the key decisions and Ember doing the volume work.

Where the claims get thinner

Several of the most striking numbers are Meta's own. Its launch table shows Ember finishing 37.8% of tasks on a business-workflow test, against 13.9% for Haiku. But Meta's evaluation report calls its runs of rival models "best-effort" and says they may not be tuned optimally.

According to eesel, four of the table's ten rows come from Meta's own testing, including a composite score Meta built itself. Meta also chose the partner endorsements in its launch post, from companies including Box, Cloudflare and Oracle.

Independent results are more mixed. On FrontierCode 2.0, which checks whether AI-written code would be accepted by a real software project, Ember scored 27.2%. Haiku scored 34.9%.

General knowledge is Ember's clearest weakness. On Artificial Analysis's knowledge test, it answered only 21.1% of questions correctly, the lowest among similarly ranked models. Haiku scored 34.7%. Ember declined to answer about 64% of questions, and its rate of made-up answers was 17.4%, about half Haiku's.

Meta says Ember stayed below its internal danger thresholds for weapons, cyberattacks and loss of control. No outside group has confirmed that, and Meta's published evaluation report describes no safety testing.

The real cost depends on the job

Long jobs carry a pricing catch. Ember's rate triples to $0.30 and $1.50 once a prompt passes 200,000 tokens, and the higher rate applies to the whole request.

Rivals have their own thresholds. Haiku's surcharge starts sooner, at 100,000 tokens, and is steeper at five times the base rate. Luna's starts later, at 272,000 tokens, and is milder.

So the cheapest model depends on the workload. Nadir, which sells software that routes requests between models, estimated the cost of one 250,000-token document. It came to about 2.7 cents with Luna, 8 cents with Ember and 13 cents with Haiku. Nadir stresses these are illustrative assumptions, not measured customer bills.

Other details cut against Ember for some buyers. Haiku offers a 50% discount on jobs that can wait, which Ember's published rate card lacks. And Meta has not released the model for download, so companies cannot run it on their own servers.

Meta sells Ember through its own API and its Muse Code coding tool. OpenRouter, a marketplace that resells access to AI models, lists Meta as the only company serving it.

The launch puts Meta in direct price competition with Anthropic and OpenAI for the high-volume, low-margin work that agents generate. For buyers, the lesson is narrower than the marketing. Ember looks like a strong, cheap option for repetitive, well-defined tasks. It is not a drop-in replacement for every job, and the bill depends on how each company's work is shaped.