Meta Muse Ember: Haiku-level reasoning at one-third the cost

Meta Superintelligence Labs has released Muse Ember, its smallest reasoning model, positioned to reshape agentic AI economics by matching the performance of Claude Haiku 5.5 while cutting per-task costs to approximately one-third.

What changed

Meta introduced Muse Ember as the fourth model in the Muse family, joining Muse Wonder, Muse Spark, and Muse Glimmer. Pricing begins at $0.10 per million input tokens for standard inputs and $0.50 per million for outputs, with cached input tokens priced at $0.004 per million. The model launched on the Meta Model API and Muse Code, Meta’s hosted coding agent platform.

Muse Ember is the first model trained using Muse Wonder Full, Meta’s largest reasoning model. The model arrives with a 1 million-token context window and can produce up to 512K output tokens, four times the 128K output limit on comparable offerings from Anthropic and OpenAI. Meta designed it for the repetitive steps inside agents: tool calls, coding subtasks, routing, extraction, and long document parsing.

Why it matters

The real story centers on cost per completed task rather than raw token pricing. Meta reports Muse Ember achieves a score of 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5 at maximum reasoning effort. Yet Muse Ember accomplishes this at approximately $0.136 per indexed task versus $0.411 for Haiku, representing a 66% cost reduction.

This matters most for agent workflows that spawn hundreds of small, structured subtasks per session. Teams building agentic systems spend most tokens on numerous small steps rather than single large requests. Cost per task governs whether an agent can operate at scale. Muse Ember removes a pricing ceiling that previously made high-volume agents economically marginal.

The model outperforms GPT-6 Luna across most benchmarks while surpassing Haiku on Terminal-Bench 4.0 with a score of 40.9 versus 39.2 and on the Artificial Analysis Coding Agent Index at 46 versus 37. Early partners in testing included Box, Cline, Oracle, Cloudflare, and You.com, indicating adoption is already underway in enterprise automation and search applications.

On the AutomationBench-AA benchmark measuring business workflow completion across 657 Zapier workflows, Muse Ember reaches 37.8%, compared to 20.2% for GPT-6 Luna and 13.9% for Claude Haiku 5.5. This performance advantage directly translates to the agent execution focus Meta claims for Muse Ember.

What to test

Before adopting Muse Ember for production agent workflows, several vendor claims warrant scrutiny.

Context window limits in practice. Meta claims a 1 million-token context window, but the source does not specify whether this limit applies uniformly across all use cases or under certain reasoning settings. Test whether the full context window remains stable when reasoning effort is set to maximum.

Output token quality at 512K. The 512K output cap substantially exceeds Haiku’s 128K limit. Verify that coherence and accuracy hold across the full output range, especially for long document parsing or multi-step code generation. No quality metrics for outputs near the ceiling are available in the source material.

Per-task cost assumptions. Meta’s $0.136 per task figure assumes specific token usage patterns across workflows. Measure actual token consumption on your own workflows before assuming the three-times cost advantage holds consistent. Token usage varies based on routing decisions, tool complexity, and output length.

Weak performance on code quality. Muse Ember scores only 27.2 on FrontierCode 2.0-AA, well behind Claude Haiku 5.5 at 34.9. If your agent performs substantial code merges into live repositories, test this benchmark closely on your own codebase before committing to production use.

Closed weights and vendor lock-in. Muse Ember is available only through Meta’s hosted infrastructure. No option exists to run it locally, quantize it, or use it without vendor dependency. This matters for teams prioritizing portability or on-premises deployment.

Long-context pricing tiers. Meta lists no long-context pricing tier for Muse Ember, even with its 1 million-token window. Haiku 5.5 charges $0.50 and $2.50 above 100K prompt tokens, and GPT-6 Luna doubles input rates above 272K input tokens. Clarify whether Meta’s rates remain flat across the full context range or if tiers appear as usage scales.

Common questions

How much does Meta Muse Ember cost compared to Claude Haiku 5.5?

Base token prices match: Meta Muse Ember costs $0.10 per million input tokens and $0.50 per million output tokens, identical to Claude Haiku 5.5. However, Muse Ember completes tasks using fewer tokens, resulting in roughly one-third the cost per task ($0.136 versus $0.411 at maximum reasoning effort). Cached input tokens cost Meta users $0.004 per million, widening the advantage on long-context workflows.

What is Muse Ember designed for?

Meta built Muse Ember for the repetitive work agents execute daily: tool calls, coding subtasks, routing, extraction, and long document parsing. Meta positions Muse Ember below its larger models, with Muse Wonder handling the hardest planning problems while Muse Ember handles high-volume execution at scale.

Does Muse Ember work with multimodal inputs?

Yes. Muse Ember natively processes text, images, audio, and video. This allows agents to read screenshots, charts, scanned documents, voice memos, and meeting recordings in a single call without a separate multimodal model. It also operates desktop applications directly from screenshots.

The conclusion

Muse Ember succeeds where it was designed to excel: running repeated, structured agent tasks at a cost that makes high-volume automation economically viable. For teams building agents that spend most tokens on many small steps, the three-times cost reduction per task represents meaningful progress. Performance on agent workflows at 37.8% on AutomationBench-AA and coding subtasks at 40.9 on Terminal-Bench 4.0 justifies the intended use case.

The model’s weakness on FrontierCode at 27.2 versus Haiku’s 34.9 means teams merging substantial amounts of code into live repositories should test thoroughly before relying on it. Closed weights and hosted-only availability lock users into Meta’s infrastructure, a tradeoff for the cost efficiency provided.

What to watch next: Meta states it has “even more efficient models in training.” If the company delivers Haiku-level reasoning at half the current Muse Ember cost, or matches Muse Spark performance at Muse Ember pricing, the agentic AI economics story will shift again. Early partner adoption at Box, Cline, and Oracle suggests the market is ready. Monitor whether these integrations report real cost savings in production or encounter token usage surprises that narrow the advertised advantage.

Related reading

• Soniox STT v6: Agent text input and echo filtering
• Gemini 4 Carbon vs GPT-6.1 Sol: Pricing and Performance
• Letterpress-1: Coding agent for growth engineering