Muse Ember released: priced exactly like the rivals Meta named — and unranked, for now.
Muse Ember, Meta’s smallest and most efficient model, was released on November 9, 2026 and enters our database today at $0.10/$0.50 per million tokens — headline rates identical to Claude Haiku 5.5 and GPT-6 Luna, the two rivals Meta’s announcement names. This post prices the tie from our own database, works the one rate that does differ — cached input — through an agent workload, sets Ember on the Muse family ladder, and states plainly why it carries no League Score yet.
By AI Pro League · Published

Muse Ember is the newest entry in our model database — the 64th model we track. Our November 11, 2026 changelog records the release from Meta’s official announcement (research.meta.ai, read today): released Nov 9, 2026 at $0.10 input / $0.50 output per million tokens, cached input $0.004, as muse-ember-1.0 on the Meta Model API and in Muse Code, and the first model Meta trained using Muse Wonder Full. It is built for the work agents repeat: tool calls, coding subtasks, routing, extraction and long-document parsing. What follows is computed from the database at build time.
Muse Ember: the release record, priced and dated
| Measure | Muse Ember |
|---|---|
| Release date | Nov 9, 2026 |
| Input / output per 1M tokens | $0.10 / $0.50 |
| Cached input per 1M tokens | $0.004 |
| Blended rate (3:1 input:output) | $0.20 |
| Context window / max output | 1,000,000 / 512,000 tokens |
| Inputs | Text, image, audio, video |
| League Score | Unranked — no tracked benchmark yet |
The record’s shape is a small-model shape: a 1M-token window with an output ceiling of 512,000 tokens — four times the 128,000-token cap we record for both named rivals — closed weights under a proprietary licence, and availability through the Meta Model API, Google Cloud and Oracle Cloud. The full rate card sits in our AI model pricing index.
The price tie: three models, one rate card
| Model | Input / output per 1M | Cached per 1M | Blended | Workload cost | Max output | League Score |
|---|---|---|---|---|---|---|
| Muse Ember | $0.10 / $0.50 | $0.004 | $0.20 | $2.00 | 512,000 | Unranked |
| Claude Haiku 5.5 | $0.10 / $0.50 | $0.010 | $0.20 | $2.00 | 128,000 | 69.5 (40% coverage) |
| GPT-6 Luna | $0.10 / $0.50 | $0.010 | $0.20 | $2.00 | 128,000 | 53.3 (40% coverage) |
Meta’s announcement frames Ember against Claude Haiku 5.5 and GPT-6 Luna on intelligence per dollar. Priced from one database, the per-token contest is a dead heat: all three charge $0.10 / $0.50, all three blend to $0.20 — the joint-cheapest blended rate among the 0 models tied at that floor in our database — and all three cost exactly $2.00 on the standard workload. The only rate that separates them is cached input: Ember’s $0.004 is 60% below the rivals’ $0.010, a 96% discount off its own input rate against their 90%.
What the cached gap is worth on an agent workload
Agent loops re-send the same context constantly, so price the pattern they actually run: 8M cached input + 2M fresh input + 2M output. Ember costs $1.23 — 8 × $0.004 + 2 × $0.10 + 2 × $0.50. Either rival costs $1.28 on the same mix. The cached advantage is real, and at this mix it is worth $0.048 a run — it grows only as the cached share grows, which is the field-wide pattern priced in our prompt caching pricing analysis. Meta’s headline figure — about $0.14 per Artificial Analysis task, against Muse Spark 1.3’s $2.39 in the same announcement table — is a per-task cost at Meta’s chosen reasoning settings, not a per-token rate; on per-token rates the gap to Spark is the family ladder below.
Where Ember sits in the Muse family
Meta’s own pairing is explicit: Muse Wonder plans, Muse Ember handles the volume. The ladder, priced on the standard workload from our records: Ember at $2.00; Muse Spark 1.3 at $21.00 (League Score 69.0, 40% coverage); Muse Wonder at $80.00 (League Score 73.0, 40% coverage). Ember is 10.5× cheaper than Spark and 40× cheaper than Wonder on the same tokens — the price of the executor tier in one line. In Meta’s announcement table the two smaller models sit a point apart on the Artificial Analysis Index (Ember 36, Spark 37); what that index does and does not say about League standing is the next section.
Why Muse Ember has no League Score yet
The launch tables report no result on any of the five benchmarks the League scores — no LiveBench, SWE-bench Verified, GPQA Diamond, MMLU-Pro or LMArena figure from a tracked leaderboard source. The headline 36 is the Artificial Analysis Intelligence Index, which the League does not score, so it cannot be converted into a League Score without inventing a mapping — and we do not invent mappings. The model therefore enters unranked at 0% coverage, exactly as our score coverage explainer describes for launch-day models. It is not alone: 13 of the 64 models we track currently carry no League Score, against 51 scored models in the AI model rankings. Inkling 2 Small and Gemini 4 Carbon entered the same way last week and are still waiting.
What the unranked entry does not yet prove
- At the same price, the rivals are known quantities. Claude Haiku 5.5 scores 69.5 and GPT-6 Luna 53.3, each on a single recorded LiveBench component at 40% coverage. Ember’s counter-offer today is the cheaper cached rate and the 4× output ceiling — not a measured score.
- Every Ember benchmark so far is vendor-published. Meta’s own table is not a clean sweep: on FrontierCode 2.0-AA, Ember’s 27.2 trails Haiku 5.5’s 34.9. Independent re-measurement on a tracked leaderboard is what turns the launch claims into a record.
- The per-task headline is setting-dependent. The $0.14 figure prices Meta’s chosen reasoning effort; per-token, Ember ties its rivals exactly. Buyers comparing rate cards should compare the table above, not the task headline.
None of that dims the release itself: a 1M-context executor at the database’s cheapest blended rate, with the deepest cached discount we record, is a serious price position. The score to watch is the first one — whichever tracked leaderboard publishes on Ember first will decide whether the tie at $0.20 is also a tie on merit. The model’s full record lives on its Muse Ember page.