When Meta shuttered its open-source Llama program and quietly walked Yann LeCun out the door, the AI world braced for something different. What arrived was Muse Spark — a closed-source, agentic-first model built by Meta Superintelligence Labs under the leadership of former Scale AI CEO Alexandr Wang. The original dropped in April 2026. Now, version 1.1 — developed under the codename “Avocado” over nine months — arrives with a provocative promise: frontier-grade intelligence at roughly one-fifth the cost of Anthropic’s Claude.
But does the pricing advantage hold up when you factor in consistency, speed, and security? We ran this model through our six-dimension evaluation framework to find out.
The Backstory: From Open Weights to Walled Garden
For years, Meta positioned itself as the open-source champion of AI. Llama models were downloadable, modifiable, and free. Then the strategy shifted. Meta Superintelligence Labs (MSL) was formed, Wang was brought in, and the company began building proprietary models behind closed doors. LeCun — Meta’s Chief AI Scientist and the loudest open-source advocate in the industry — departed. The message was clear: Meta wanted to control its AI stack end-to-end.
Version 1.1 is the second major output of this new direction. It’s available free through the Meta AI app and meta.ai website. API access is through Meta’s Model API, currently in public preview with a $20 starting credit — but only in the United States, and only after waiting in a queue. The model is already integrated into Ray-Ban Display smart glasses, with plans to expand to WhatsApp, Instagram, Facebook, and Messenger.
Dimension 1: Real-World Scenario Testing
This is where the model’s agentic architecture genuinely impresses. Three features define its real-world capability:
Contemplating Mode and Sub-Agent Orchestration
The standout feature is “Contemplating Mode,” which spins up multiple sub-agents to analyze a problem in parallel before synthesizing the best answer. Sub-agents can operate in a “leader” role (distributing tasks) or “follower” role (maintaining a specific persona). In practice, this means the model can decompose a complex request — say, “analyze this quarterly report and draft three strategic recommendations” — into parallel research threads that converge into a single coherent output.
Tool Use: Best in Class
On the MCP Atlas benchmark (which measures tool-use proficiency), the model achieved an 88.1% pass rate — first place, edging out Gemini 3.5 Flash (83.6%) and Claude Fable 5 (83.3%). This isn’t a marginal win. It means that in workflows requiring API calls, function execution, and multi-step tool chains, this is currently the most capable model on the market.
Coding and Computer Use
The Coding Agent Index score of 71.3 (at xhigh reasoning) is nearly identical to GPT-5.6 Luna’s 71.4. The Computer Use feature is particularly interesting: rather than blindly clicking through UIs, the model decides whether to write a script or interact directly with the interface — and it can batch operations for efficiency. For developers building automation pipelines, this hybrid approach is more reliable than pure browser-agent models.
Specialized Benchmarks
- JobBench (professional tasks): 54.7% — second place, behind Claude Fable 5 (57.4%) but ahead of Claude Opus 4.8 (48.4%)
- HealthBench Hard: 42.8% — first place, beating GPT-5.4 (40.1%) and Gemini 3.1 Pro (20.6%)
- CharXiv (chart understanding): 86.4, surpassing GPT-5.4 (82.8) and Gemini (80.2)
- Humanity’s Last Exam (contemplating mode): 58%
- Frontier Science Research: 38%
The HealthBench Hard result is particularly noteworthy. Medical AI is a minefield of regulatory and safety concerns, but the model’s ability to reason through complex clinical scenarios — while costing a fraction of competing models — positions it as a serious contender for healthcare research applications.
Dimension 2: Per-Task Cost Analysis
This is the headline number. On the Artificial Analysis Intelligence Index, the model scores 51 (tied with GPT-5.6 Luna and GLM-5.2). But the cost per task is $0.26 — the second-lowest in the frontier tier, behind only Luna’s $0.21.
The API pricing tells the story more dramatically:
| Model | Input ($/1M tokens) | Output ($/1M tokens) | Cost per Intelligence Task | Coding Cost/Task |
|---|---|---|---|---|
| Muse Spark 1.1 | $1.25 | $4.25 | $0.26 | ~$1.40 |
| GPT-5.6 Luna | ~$1.00 | ~$4.00 | $0.21 | ~$1.20 |
| Claude Opus 4.8 | $5.00 | $25.00 | ~$1.10 | ~$5.50 |
| Claude Fable 5 | $10.00 | $50.00 | ~$2.50 | ~$12.00 |
| GPT-5.6 Sol | $6.00 | $30.00 | ~$1.50 | ~$7.00 |
The output token price is where the savings compound. At $4.25 per million output tokens, it’s 83% cheaper than Claude Opus 4.8 ($25), 86% cheaper than GPT-5.6 Sol ($30), and 91.5% cheaper than Claude Fable 5 ($50). For high-volume agentic workflows that generate long outputs — think multi-step research reports, code generation, or document analysis — this is transformative.
But the real secret weapon is Thought Compression. During Intelligence Index testing, the model consumed only 58 million tokens. Claude Opus 4.6 used 157 million. GPT-5.4 used 120 million. That’s a 2-3x efficiency improvement, which means the model doesn’t just charge less per token — it uses fewer tokens to reach the same conclusion.
Cached input pricing is equally aggressive at $0.15 per million tokens — roughly 12% of the standard input rate. For applications with repeated prompt prefixes (system prompts, few-shot examples, context windows), this cuts costs further. And if you’re comparing budget options, our DeepSeek V4 Flash review shows another model punching above its price class at $0.28 per task — remarkably close to what we’re seeing here.
Dimension 3: Failure Modes
No model is perfect, and the failure patterns here are significant enough to give any production team pause.
Inconsistency: The Biggest Problem
The consistency score of 7.9 (out of 10) is the lowest among frontier models. For comparison, Gemini 3.1 Pro scores a perfect 10.0. In practice, this means the same prompt can yield dramatically different quality outputs across runs. The “attempt pass rate” — the percentage of times the model successfully completes a task on a given try — is 72.7%, versus Claude Opus 5’s 80.3%.
For agentic workflows that run unattended, this is a real problem. A 27.3% failure rate on any given attempt means you need retry logic, validation layers, and human-in-the-loop checkpoints — all of which add to the effective cost per successful task. When you factor in retries, the real cost per successful outcome may be 30-40% higher than the headline $0.26 figure suggests.
Security: A Red Flag
Anthropic’s red-teaming revealed that the model was successfully breached by prompt injection attacks 16.5% of the time (across 15 attempts). Claude Opus 5’s breach rate was 2.0%. For a model that excels at tool use and computer automation — where it can execute actions, not just generate text — a 1-in-6 prompt injection success rate is alarming. Any deployment that connects this model to sensitive systems, databases, or financial APIs needs robust input sanitization, output filtering, and action confirmation layers.
The Muse Image Debacle
The model’s companion image generation tool, Muse Image, was launched and then pulled offline within days after Instagram users discovered they’d been automatically opted into having their photos processed. The privacy backlash was swift, and the tool remains offline. While this doesn’t directly affect the text model’s capabilities, it signals a broader concern about Meta’s approach to user consent in AI features — particularly when those features are deployed across social platforms with billions of users.
Yann LeCun’s Departure and the Open-Source Question
The departure of LeCun — who publicly championed open-source AI as essential for democratic access to technology — raises questions about Meta’s long-term commitment to AI safety transparency. Closed-source models are harder for independent researchers to audit, and the 16.5% prompt injection breach rate might have been caught earlier with community testing.
Dimension 4: Audience Recommendation Matrix
| User Type | Recommendation | Why |
|---|---|---|
| Meta ecosystem users (WhatsApp, Instagram, Ray-Ban) | Strongly recommended | Free access, deep integration, no API setup needed |
| Agentic workflow developers | Recommended with caveats | Best-in-class MCP Atlas score, but add retry logic for consistency issues |
| Cost-sensitive teams | Strongly recommended | Lowest frontier-tier cost per task, thought compression reduces token usage |
| Healthcare and research | Worth evaluating | #1 on HealthBench Hard, but verify regulatory compliance for your use case |
| Security-critical applications | Not recommended | 16.5% prompt injection breach rate is too high for sensitive systems |
| Real-time chatbot builders | Not recommended | 24.97s average response time at xhigh is too slow for conversational UX |
| Self-hosting advocates | Not applicable | Closed source, no download available — a complete reversal from Llama’s open model |
Our Claude Opus 5 review found that model to be the better choice for applications requiring consistency, speed, and security — though at roughly 5x the cost per task.
Dimension 5: Hidden Limitations
Beyond the benchmark scores and pricing tables, several limitations aren’t immediately obvious:
- API is US-only with a waitlist. The Meta Model API is in public preview, restricted to the United States, and requires queuing. International developers are limited to the free Meta AI app interface, which doesn’t support programmatic integration.
- Contemplating Mode adds significant latency. While the multi-agent approach improves answer quality, it can push response times well beyond the already-slow 24.97-second average. Budget for 30-60 seconds when sub-agents are involved.
- Muse Image is offline. The image generation companion tool was pulled after the Instagram auto-opt-in privacy scandal. There’s no timeline for its return, and the multimodal pipeline is currently text-and-image-in, text-out only.
- Closed source, no fine-tuning. Unlike Llama models that you could download, fine-tune, and self-host, this model runs only on Meta’s infrastructure. Your data flows through Meta’s servers — a concern for organizations with strict data residency requirements.
- Web search costs extra. At $2.50 per 1,000 queries, web search isn’t included in the base API. For agentic workflows that require real-time information retrieval, this adds up quickly.
- Ad-subsidized pricing is structural, not permanent. Meta can offer low API prices because advertising revenue subsidizes model development. If ad revenue declines or Meta’s priorities shift, pricing could change without notice — a risk that pure-API companies like Anthropic or OpenAI don’t carry in the same way. Meta also raised its capital expenditure from $125B to $130B, and free cash flow hit a 5-year low, raising questions about long-term sustainability of aggressive pricing.
- Consistency issues compound at scale. The 7.9 consistency score means that in a batch of 100 identical queries, expect roughly 20 to produce noticeably different quality results. For production pipelines expecting deterministic behavior, this variance is a dealbreaker without external validation layers.
Dimension 6: Speed Benchmarks
The model generates output at 119 tokens per second — competitive but not class-leading. The bottleneck is reasoning time, not generation speed.
| Metric | Muse Spark 1.1 | Claude Opus 5 | Gemini 3.1 Pro |
|---|---|---|---|
| Output speed | 119 tok/s | ~150 tok/s | ~130 tok/s |
| Avg response time (xhigh) | 24.97s | 6.99s | 21.47s |
| Output token limit | 131,072 | 64,000 | 65,536 |
| Context window | 1,048,576 | 200,000 | 2,000,000 |
| Intelligence Index score | 51 | 51 | 53 |
| Consistency score | 7.9/10 | 9.2/10 | 10.0/10 |
The 1M-token context window is a genuine advantage — five times Claude Opus 5’s 200K and large enough to ingest most codebases or book-length documents in a single pass. The 131K output limit also exceeds most competitors, making it well-suited for long-form generation tasks like report writing or code refactoring.
The trade-off is clear: if you need speed, Claude Opus 5 is 3.6x faster. If you need long context and large output, this model wins decisively. The adjustable reasoning depth (none/minimal/low/medium/high/xhigh) gives developers a knob to trade latency for accuracy — but even at lower settings, it trails competitors on response time.
The Verdict: High Value, High Caveats
This is the most interesting model release of 2026 — not because it’s the best at any single thing, but because it forces a rethinking of the cost-capability tradeoff. At $0.26 per intelligence task with best-in-class tool use, a 1M context window, and the Contemplating Mode architecture, it delivers capabilities that would cost 5-10x more on competing platforms.
But the 7.9 consistency score, 16.5% prompt injection vulnerability, and 25-second response times are not minor issues. They’re structural limitations of a model that prioritizes deep reasoning over speed — and security was apparently not the priority.
Zuckerberg called competitors’ pricing “very extreme, with high profit margins.” He’s not wrong about the margins. But what he didn’t mention is that the lower price buys you a model that’s slower, less consistent, and easier to compromise. The question isn’t whether the cost savings are real — they absolutely are. The question is whether your use case can absorb the trade-offs.
For cost-sensitive teams building agentic workflows where latency isn’t critical and security hardening is handled externally, this is the model to beat. For everyone else — especially those who need reliability, speed, or security as core features — the premium charged by Anthropic and OpenAI still buys something real.
Ready to build with agentic AI? Start with the free Meta AI app to test capabilities, then apply for API access at the Meta Model API portal. Your first $20 in API credits is on Meta.
Frequently Asked Questions
1. Is Muse Spark 1.1 free to use?
Yes. The model is available for free through the Meta AI app and the meta.ai website. API access costs money (starting with a $20 credit), and web search incurs an additional $2.50 per 1,000 queries.
2. How does it compare to Claude Opus 5?
On the Artificial Analysis Intelligence Index, both score similarly (8.6 in direct comparison). However, Claude Opus 5 is 3.6x faster (6.99s vs 24.97s), more consistent (higher stability score), more secure (2.0% vs 16.5% breach rate), and costs roughly 5x more per task. This model wins on cost and tool use; Claude Opus 5 wins on speed, reliability, and security.
3. What is Contemplating Mode?
Contemplating Mode is a feature that launches multiple sub-agents to analyze a problem in parallel, then synthesizes the best solution. It improves answer quality but adds significant latency — expect 30-60 second responses when enabled.
4. Can I self-host this model?
No. Unlike Meta’s previous Llama models, Muse Spark is closed-source and runs only on Meta’s infrastructure. This represents Meta’s strategic shift from open-source AI to a proprietary model ecosystem.
5. Is the API available outside the US?
Not yet. The Meta Model API is currently in public preview and restricted to the United States. International users can only access the model through the free Meta AI app, which doesn’t support programmatic integration.
6. What happened to Muse Image?
Muse Image, the companion image generation tool, was launched and pulled offline within days after Instagram users discovered they’d been automatically opted into photo processing. The tool remains offline with no announced return date.
7. Should I switch from GPT-5.6 Luna?
Both models score 51 on the Intelligence Index. Luna is cheaper per task ($0.21 vs $0.26) and has a slightly higher Coding Agent Index (71.4 vs 71.3). However, this model has a 1M context window (vs Luna’s smaller context), better tool use (MCP Atlas 88.1%), and Contemplating Mode. If your work involves long documents or tool-heavy workflows, the switch makes sense. If you’re doing pure coding or standard chatbot tasks, Luna remains the better value.
Looking for a more affordable AI model alternative? Check out our comprehensive comparison of budget-friendly frontier models to find the right fit for your use case and budget.
Sources: Artificial Analysis, Arena.ai, DeepLearning.AI, WAM, Android Authority, Futurism, Anthropic red-teaming reports. Benchmarks current as of July 2026.
\n\n\n