Meta's Muse Ember Ties Claude Haiku 5.5 at 36 for About a Third of the Cost per Task. It Does Not Beat It, and Prompts Over 200,000 Tokens Cost Three Times as Much.
Meta Superintelligence Labs released Muse Ember on November 9 on the Meta Model API and Muse Code as muse-ember-1.0: $0.10 per million input tokens, $0.004 cached, and $0.50 output on prompts up to 200,000 tokens, and $0.30, $0.012, and $1.50 above that, with a 1 million token context and up to 512,000 output tokens. Artificial Analysis scores it 36 at max effort, second of 33 in its price class, at $0.14 per Index task, 3.2 minutes per task, and 446 tokens per second. Meta's own table puts Claude Haiku 5.5 at max at the same 36, so the pitch of more intelligence per dollar is a cost win at an equal score; Haiku leads on FrontierCode 2.0. No parameter count, no weights. We read Meta's post and AA's page.
By T.W. Ghost
Meta Superintelligence Labs released Muse Ember today, November 9, 2026, on the Meta Model API and in Muse Code as muse-ember-1.0, with the Responses API supported and availability on Google Cloud and Oracle Cloud. It is the first model Meta names as trained with Muse Wonder Full, the teacher behind last week's specialized-models announcement. The price: $0.10 per million input tokens, $0.004 for cached input, and $0.50 output on prompts up to 200,000 tokens; above that, $0.30, $0.012, and $1.50. Context is 1 million tokens, output up to 512,000. Meta lists text, image, audio, and video input; Artificial Analysis lists text and image.
AA scored it at 36 on the Intelligence Index at max effort, second of 33 in its price class, at $0.14 per Index task, 3.2 minutes per task, and 446 tokens per second, using 115 million output tokens against a median near 90 million. Meta's own chart puts Claude Haiku 5.5 at max at the same 36, at $0.411 per task by Meta's reading of AA's data. So the claim Meta makes, more intelligence per dollar, is precisely true and precisely limited: the same score for about a third of the cost per task. On FrontierCode 2.0, Haiku leads at 34.9 to Ember's 27.2. And the list price carries the same shape of trap we flagged on Haiku, a long-prompt surcharge, here at 200,000 tokens. We read Meta's research post and AA's model page. Here is what held up.
What You Are Calling
| Muse Ember (November 9, 2026) | |
|---|---|
| Where | Meta Model API; Muse Code; Google Cloud; Oracle Cloud; model id muse-ember-1.0 |
| Price, prompts up to 200K | $0.10 in / $0.004 cached / $0.50 out per 1M |
| Price, prompts over 200K | $0.30 / $0.012 / $1.50 per 1M |
| Context and output | 1M context; up to 512K output |
| Input | Text, image, audio, video (Meta); text and image (AA) |
| AA Intelligence Index | 36 at max; second of 33 in price class |
| AA cost and speed | $0.14 per task; 3.2 minutes per task; 446 tokens per second; 4.37 seconds to first answer token; 115M output tokens |
| Haiku 5.5 on Meta's chart | 36 at max; $0.411 per task (Meta's reading of AA data) |
| Size and weights | No parameter count; proprietary; one provider on AA |
| Chart caption | "Muse Ember costs are estimates at proposed pricing" (Meta) |
A tie is not a win, and Meta's own chart says tie. Thirty-six and thirty-six. The per-dollar framing is correct because the denominator moved, not the numerator.
The surcharge is the fine print. Three times the input price and three times the output price past 200,000 tokens of prompt. A long-context agent that lives above that line is not paying $0.10 and $0.50; it is paying $0.30 and $1.50. The same catch sat at 100,000 tokens on Haiku 5.5.
"Proposed pricing" is Meta's phrase. The chart's cost column is an estimate at a price Meta describes as proposed. The rate card above is what the post publishes today.
The Benchmarks, Sorted by Who Reported Them
| Claim | Number | Who |
|---|---|---|
| Intelligence Index | 36 at max; second of 33 in class | Independent (Artificial Analysis) |
| Cost, time, speed per Index task | $0.14; 3.2 min; 446 tok/s | Independent (AA) |
| Haiku 5.5 at 36; Luna at 28 | Meta's relay of AA data | |
| FrontierCode 2.0 | Haiku 34.9; Ember 27.2; Luna 25.2 | Meta's relay of AA data |
| AutomationBench | Ember 37.8 vs Haiku 13.9 | Meta's relay; not confirmed on AA pages here |
| Coding Agent Index | Ember 46 vs Haiku 37 | Meta's relay; not confirmed on AA pages here |
| Agents' Last Exam, Personal Superintelligence Index, OSWorld 2.1, PerceptionBench | Meta internal evals | |
| Terminal-Bench 4.0 rows | Vendor, with different harnesses per model by Meta's own footnote | |
| Ember vs Muse Spark 1.3 | 36 vs 37 at $0.14 vs $2.39 per task | Meta's relay of AA data |
| Partner claims (Box, Cline, Oracle, Cloudflare, You.com p95 2.2 s) | Vendor and partner |
Three honest observations:
- The rows worth trusting are the ones AA published directly: 36, $0.14, 446 tokens a second. The rest of Meta's table mixes AA-derived rows, internal evaluations, and Terminal-Bench runs on different harnesses per model, which Meta footnotes and which makes that column unreadable as a comparison.
- Haiku wins the code-quality row. FrontierCode is a private Cognition benchmark, and on it Haiku leads by nearly eight points in Meta's own relay. Ember's AutomationBench and Coding Agent numbers would be interesting if confirmed on AA's pages; we did not confirm them.
- Speed is the quiet strength. Four hundred forty-six tokens per second and 4.4 seconds to first answer on AA's new method are the numbers a latency-bound agent cares about, and they are independent.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Released Nov 9; surfaces; model id | Verified (Meta post) |
| Price tiers at 200K | Verified (Meta post, footnote) |
| 1M context; 512K output; modalities | Verified (Meta); AA lists text and image only |
| AA 36, second of 33; $0.14; 3.2 min; 446 tok/s; 115M tokens | Verified (AA model page) |
| Haiku 5.5 at 36 and $0.411 per task | Meta's relay of AA |
| FrontierCode, AutomationBench, Coding Agent rows | Meta's relay; AA pages not rechecked here |
| Internal evals and partner quotes | Vendor |
| Parameter count; weights | Not published; proprietary |
Who Should Use It
Use Ember for high-volume agent loops under 200,000 tokens of prompt where Haiku-class intelligence at a third of the per-task cost and twice the speed is the trade you want.
Skip the "beats Haiku" reading; it ties on the independent score and trails on the code-quality row, and skip the low price for long-context work, where the rate triples.
The frontier is elsewhere: Meta's frontier claim this autumn still rests on Muse Wonder, and its open claim on Muse Spark Open, now a generation and a scale behind. Our model comparison has the current standings side by side.
Sources
- Meta AI Research: Muse Ember
- Artificial Analysis: Muse Ember model page
- TechCrunch: Meta releases Muse Ember, a new lightweight model for agent execution
*Not sure whether a tie at a third of the price beats the incumbent for your agent? Take the free 2-minute quiz and get matched. Then read about the model Ember ties and the surcharge it shares: Claude Haiku 5.5.*