Meta's Muse Ember Undercuts Rivals at $0.10 Per Million Tokens

Meta released Muse Ember on Monday at $0.10 per million input tokens, a price that puts it at roughly a third of Claude Haiku 5.5's cost per task while matching the Anthropic model's score on the Artificial Analysis Intelligence Index. The smallest and cheapest model in the Muse family, built by Meta Superintelligence Labs, targets the unglamorous bulk of agent work: tool calls, coding subtasks, routing, extraction and long-document parsing.

The pricing is aggressive by any measure. Ember costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, with a 1 million-token context window. It can return up to 512,000 tokens in a single response. It's available now through the Meta Model API and Meta's Muse Code coding tool under the ID muse-ember-1.0. Meta has not said whether it will release the weights — a notable omission for a company that shipped an open-weight model in the same family just months ago.

What do the benchmarks actually show?

Meta's headline claim rests on Artificial Analysis's Intelligence Index. At maximum reasoning, Ember scores 35.8 at an estimated $0.136 per task, against 36.2 at $0.411 for Anthropic's Claude Haiku 5.5. OpenAI's GPT-6 Luna peaks at 27.8. Meta's own larger Muse Spark 1.3 scores 37 — but at about $2.39 per task, more than seventeen times Ember's estimated cost. Meta notes the cost figures are its own estimates at list prices.

The full picture is mixed, and Meta discloses who measured what:

• AutomationBench-AA: Ember 37.8% vs. Haiku 5.5's 13.9% and GPT-6 Luna's 20.2% (Artificial Analysis)
• Coding Agent Index-AA: Ember 46 vs. Haiku's 37 and Luna's 31 (Artificial Analysis)
• FrontierCode 2.0-AA: Haiku still leads at 34.9% to Ember's 27.2%, with Luna at 25.2% (Artificial Analysis)
• Terminal-Bench 4.0: Ember 40.9%, Haiku 39.2%, Luna 16.4% (best published harness)
• OSWorld 2.1, partial credit: Ember 69.5%, Haiku 68.9%, Luna 50.5% (Meta internal)
• Agents' Last Exam, Linux subset: Ember 61.6%, Haiku 57.0%, Luna 57.1% (Meta internal)

Several of those tests are Meta's own runs, and the OSWorld figure uses partial credit, so the gaps on those rows carry less weight than the independently measured Artificial Analysis results. On raw coding chops, Haiku 5.5 keeps a clear lead.

Early testers quoted by Meta focused on speed rather than intelligence scores. Cloudflare CTO Dane Knecht said Ember ran up to 30% faster at the tail than a comparable model. You.com reported a 95th-percentile latency of 2.2 seconds on its coding tasks.

Where does Ember fit in Meta's lineup?

Ember is the fourth Muse model this year, after Spark in April, the open-weight Glimmer in August and the frontier model Wonder in October. Meta says it is the first model trained using Muse Wonder Full, and it frames the pairing as a division of labour: "Muse Wonder plans, Muse Ember handles the volume." In Meta's customer-support example, the larger model sets the rules and Ember applies them across tens of thousands of cases.

The model also takes text, images, audio and video as input, so an agent can read screenshots or scanned documents without bolting on a separate vision model.

Why the cheap model matters commercially

A $0.10-per-million-token model gives Meta a high-volume product to sell to businesses. The Information has reported that Meta is building an enterprise AI sales team and lining up consulting firms as resellers (MLQ coverage). On the consumer side, the company keeps adding features to its Muse assistant, including real-time voice chats in the U.S.

If Meta's cost-per-task estimates hold up under third-party testing, Ember puts sustained pricing pressure on the small-model tiers at Anthropic and OpenAI — and gives Meta something concrete to sell into enterprises as its reseller channel takes shape.