The same headline rate, two different rate cards
The launch claim is a cost claim, so start with the rate cards — all three checked against the providers' own pricing pages on November 9, 2026. Ember lists at $0.10 per million input tokens and $0.50 per million output tokens on the Meta Model API. Anthropic lists the same headline rates for Claude Haiku 5.5, and OpenAI lists the same for GPT-6 Luna at Standard tier. Meta has not undercut anyone's headline price per token.
The parity has edges, and they are on the rate cards too. Anthropic lists higher input and output rates — $0.50 and $2.50 per million tokens — for Haiku requests whose prompts exceed 100,000 tokens. Meta lists a 1M-token context window for Ember and states no prompt-length pricing tier in its launch post. Cached input runs the other way: Ember charges $0.004 per million cached input tokens, where Anthropic and OpenAI both list $0.01 for Haiku and Luna cache reads. Within Ember's own card, the cached rate is one twenty-fifth of its standard input price — 96% lower.
What those cards do not establish is why Meta's reported cost per task differs. Meta reports about $0.14 per Artificial Analysis Index task for Ember, describes it as about a third of Haiku's cost, and coverage of the Index figures (Mad Robot, November 9) puts Haiku's figure at $0.41. A per-task cost depends on token counts, input-output mix, cache treatment and reasoning settings, none of which the launch material itemises. The desk reports the estimates as estimates: equal headline rates, a lower reported cost per Index task, and no published breakdown that would let a reader attribute the gap to any single cause.
What Meta's table shows, and what its footnotes qualify
At maximum reasoning, Meta's selected results are strong where the work is structured. Ember scores 37.8% on AutomationBench-AA end-to-end business workflows, against 13.9% for Haiku and 20.2% for Luna. It leads the Coding Agent Index-AA 46 to 37 over Haiku, and edges Terminal-Bench 4.0, 40.9% to 39.2%.
Two entries in the same table deserve a slower read. On FrontierCode 2.0-AA — the test Meta's table labels "real engineering tasks judged on mergeability" — Haiku scores 34.9 to Ember's 27.2. The computer-use result needs its scoring kept attached: Meta reports 69.5 partial-credit points and 33.8 binary-score points for Ember on OSWorld 2.1, against 68.9 and 27.8 for Haiku. These are partial-credit and binary benchmark scores, not demonstrated completion rates on customers' workloads. And two of the table's largest Ember leads carry Meta's own qualifications in its footnotes: Agents' Last Exam is a Meta internal evaluation of the Linux subset, and the Personal Superintelligence Index is an internal composite index. The Intelligence Index and the cost-per-task figures are the parts of the table Meta attributes to Artificial Analysis, an external evaluator; the desk has not independently audited that methodology.
Where it sits
Meta's own family framing assigns the hardest problems to Muse Wonder, coding and agentic workflows to Muse Spark, and repeatable tasks at scale to Ember. The comparison inside the family is steep: against Spark 1.3, Ember gives up one Intelligence Index point, 36 to 37, at a reported $0.14 per Index task instead of $2.39 — about 17 times as much per Index task, on those two estimates. What that ratio means for any real workload depends on the workload; it is a comparison of two Index estimates, not a price list for arbitrary traffic.
What Ember is not, on the launch post's evidence, is an open release: the post offers no downloadable weights, and the model runs through Meta's API and resellers such as OpenRouter. On safety, Meta says its evaluations under the Superintelligence Scaling Framework found Ember below the framework's high-risk thresholds in the categories assessed — Meta's assessment, not an independent certification. A Box partner quote appears in the launch post itself; the speed and token-efficiency quotes attributed to Cline, Oracle, Cloudflare and You.com appear in launch coverage, and are partner claims either way, not independent tests.
Sources
- [1] Meta, "Introducing Our Smallest, Most Efficient Model: Muse Ember", research.meta.ai, November 9, 2026Read source
- [2] Anthropic, Claude Platform pricing documentation (Claude Haiku 5.5 rates, long-context pricing, prompt caching), read November 9, 2026Read source
- [3] OpenAI, API pricing (GPT-6 Luna, Standard tier), read November 9, 2026Read source
- [4] Mad Robot, "Meta Muse Ember: Haiku-level AI at a third of the cost per task", November 9, 2026Read source