Meta's Muse Ember Matches Haiku 5.5 at a Third of the Cost
By @sharedot · · 7 pages
- Programming
- AI Models
- Developer Tools
Meta's new agent model Muse Ember ties Claude Haiku 5.5 on the Artificial Analysis index while costing roughly a third as much per task.
Meta ships its smallest model yet
Meta AI Research released Muse Ember on November 9, 2026, describing it as the company's smallest model and its most efficient option for repeated agent and coding work.
It is available the same day through Muse Code and the Meta Model API as `muse-ember-1.0`, supporting the Responses API, priced at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens.
Meta says agents spend most of their tokens on small, structured steps — tool calls, coding subtasks, routing, extraction, and long-document parsing — and Ember is built to run that layer end to end.
It is also the first model trained using Muse Wonder Full, and it natively processes text, images, audio, and video.
Why the pricing reversal is surprising
According to OfficeChai, Ember scores 35.8 at max reasoning on the Artificial Analysis Intelligence Index v5, effectively tied with Claude Haiku 5.5's 36.2, while Artificial Analysis puts Ember's cost per task at $0.136 against $0.411 for Haiku at max — roughly a third of the price for the same score.
Even Ember's lower xhigh setting (33.9 at $0.116 per task) beats Luna's best result of 27.8 at $0.152. unite.ai notes the same trade-off against Meta's own Muse Spark 1.3: Ember scores 36 versus 37, but costs $0.14 per task against $2.39.
The per-token rate card matches Luna's, so the savings come from fewer tokens burned, not a lower sticker price.
The benchmark evidence — and its limits
Per unite.ai, Meta reports Ember at maximum reasoning scoring 37.8% on AutomationBench-AA business workflows versus 13.9% for Claude Haiku 5.5 and 20.2% for GPT-6 Luna, plus 61.6% on Agents' Last Exam (Linux subset), 40.9% on Terminal-Bench 4.0, a Coding Agent Index-AA of 46, and 69.5% partial credit on OSWorld 2.1.
OfficeChai reports Ember leads or ties on nine of ten published evaluations. The exception matters for developers: Haiku leads FrontierCode 2.0-AA, which scores real engineering tasks on mergeability, at 34.9% against Ember's 27.2%.
Several evaluations, including Agents' Last Exam and OSWorld, are Meta internal runs, and Terminal-Bench placed each model in its best published harness, so the comparison is not fully like for like.
Stakes: a cheaper execution layer under a planner
The intended architecture is a planner-and-worker split: Muse Wonder, at $4 per million input and $20 per million output tokens, plans the hardest problems, Muse Spark handles coding and agentic workflows, and Ember executes repeatable versions of those tasks at scale — the same pattern OfficeChai notes Anthropic pitches for Haiku 5.5 alongside Opus 5.5 and Sonnet 5.5.
Early partner results, as reported by unite.ai, back the latency story: Cloudflare CTO Dane Knecht said Ember ran up to 30% faster at the tail and handled multi-repository investigations well, while you.com inference lead Chak Pothina reported 2.2-second 95th-percentile latency and 36–48% fewer output tokens.
Meta evaluated Ember under its Superintelligence Scaling Framework and reports it stayed below high-risk thresholds, including tests on resisting instructions hidden in web pages and tool outputs.
What comes next for agent economics
For working developers, Ember creates a low-cost execution layer that can sit beneath a planner and cut the number of repetitive calls sent to a more expensive model — Meta's numbers support the division of labor, while the FrontierCode result marks its limit, since efficiency does not guarantee superiority on every engineering task.
Meta says even more efficient models are in training, and OfficeChai links the release to Meta's consumer agent push with partners like Shopify and Expedia, where a model cheap enough to run in every product makes the economics work.
The release also lands in a wider no-code tooling wave: The Herald Business reports that South Korea's ETRI released TANGO 2 as open source on GitHub, a no-code tool that automates training, optimization, and deployment of AI models for practitioners without coding expertise.