Meta's Muse Ember matches Claude Haiku 5.5 at a third of the cost per task, at the same price per token

Meta Superintelligence Labs released Muse Ember on November 9, its smallest and cheapest model, pitched at the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long-document parsing. It is available as muse-ember-1.0 on the Meta Model API and in Muse Code, and on Google Cloud and Oracle Cloud. It is not an open-weight model. Meta says it is the first model trained using Muse Wonder Full.

The headline claim is "more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting." On Artificial Analysis's Intelligence Index, Ember at max reasoning scores 35.8 for $0.136 per task, against 36.2 for $0.411 for Haiku 5.5 at max. GPT-6 Luna tops out at 27.8 for $0.152.

Same price per token, fewer tokens per task

Ember is not cheaper per token. Its list price is $0.10 per million input tokens and $0.50 per million output tokens, which is exactly what Anthropic charges for Claude Haiku 5.5 and what OpenAI charges for GPT-6 Luna, according to both companies' current pricing pages (Anthropic, OpenAI). The real price difference is in cached input: $0.004 per million for Ember against $0.01 for both rivals. For agents that resend a long system prompt and history on every call, that matters. Artificial Analysis logged about 10 billion input tokens against 115 million output tokens for Ember's run of its index.

The other difference is the long-prompt surcharge. Ember's price triples above 200,000 input tokens ($0.30 / $0.012 cached / $1.50), while Haiku 5.5 switches to $0.50 / $0.05 / $2.50 above 100,000. So the threefold gap per task comes from Ember spending fewer tokens and paying less for the cached ones, not from a lower sticker price. TechCrunch's launch report called Ember "significantly cheaper" than Haiku and Luna at 10 cents per million input tokens; per token, it isn't.

The more striking comparison is inside Meta's own lineup. Meta's table puts Ember at 36 on the index for $0.14 per task and Muse Spark 1.3, which Meta positions for coding and agentic workflows, at 37 for $2.39. That is one index point for about 17 times the cost per task.

Which numbers are independent

Of the ten results Meta shows against Haiku 5.5 and Luna, five come from Artificial Analysis and the rest from Meta's own suite. Ember leads on Artificial Analysis's Coding Agent Index (46 vs 37 for Haiku) and AutomationBench (37.8% vs 13.9%, strict pass@1), and loses on FrontierCode 2.0, which judges whether code is mergeable (27.2% vs 34.9% for Haiku). The Personal Superintelligence Index, where Ember leads 46.8 to 36.1, is Meta's internal composite of eight evaluations.

The Terminal-Bench 4.0 row needs a footnote that Meta does provide: each model ran in its "best published harness," which means mini-swe-agent for Ember, Claude Code for Haiku and Codex for Luna. Ember's 40.9% against Haiku's 39.2% and Luna's 16.4% is therefore a comparison of model-plus-harness pairs, and Luna's low score may say as much about its harness as about the model.

Already tested in the wild

Meta's Alexandr Wang wrote on November 10 that the anonymous test model "crouton-alpha" was an early version of Ember, and that it processed over 5 trillion tokens in testing before launch, with "a bunch of the issues" fixed since.

Partners quoted by Meta, including Cline, Cloudflare and You.com, mostly praise speed and token efficiency for subagent work. That fits the positioning: Meta suggests pairing Ember with Muse Wonder, with the large model planning and Ember "handling the volume." Whether Ember holds up on coding quality is the open question, and FrontierCode, its one clear loss to Haiku, is the benchmark closest to that.