Strengths
- Very cheap: $0.10 per million input tokens and $0.50 per million output tokens, with cached input at just $0.004.
- It scores 36 on the independent Artificial Analysis index, level with Claude Haiku 5.5 and ahead of GPT-6 Luna (28).
- Meta says it is strong at agent tasks: 40.9% on Terminal-Bench 4.0, against 39.2% for Haiku 5.5 and 16.4% for GPT-6 Luna.
- A 1 million token context and answers of up to 512K tokens. It understands text, images, video, audio and PDFs.
- It runs in Standard mode only, where Meta does not train on your content.
Weaknesses
- It is brand new, and apart from the Artificial Analysis index all results come from Meta. Some of the tests are Meta's own internal ones.
- It trails on the demanding FrontierCode 2.0 coding test even by Meta's figures: 27.2% versus 34.9% for Claude Haiku 5.5.
- Above 200K input tokens the whole request is billed at three times the price.
- It is closed, with no downloadable weights. Meta publishes neither a country list for the API nor an EU data processing option.
Best for
- helper agents and tool calls at scale
- sorting and routing requests
- pulling data out of long documents
- coding subtasks
Not ideal for
- the hardest coding and planning, for which Meta recommends Muse Wonder
- data that must stay in the EU
How to use it in Europe
Developers use it through the Meta Model API, in Standard mode only, where Meta does not train on your content. It is also included in a Muse Code subscription. Meta does not mention it for the Muse app or the Meta AI assistant. Meta does not publish a country list for the API or an EU data processing option, so check this before using it with personal data.
Price compared with similar models
Benchmark results
| Benchmark | Score | Source |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0.1 (max) | 36 (#13 of 52) | artificialanalysis.ai ↗, November 2026 |
| Terminal-Bench 4.0 (max, vendor-reported) | 40.9% | research.meta.ai ↗ (vendor-reported), November 2026 |
| Coding Agent Index-AA (max, vendor-reported) | 46 | research.meta.ai ↗ (vendor-reported), November 2026 |
| OSWorld 2.1 partial credit (max, vendor-reported) | 69.5% | research.meta.ai ↗ (vendor-reported), November 2026 |
| AutomationBench-AA (max, vendor-reported) | 37.8% | research.meta.ai ↗ (vendor-reported), November 2026 |
Benchmarks cover only a narrow slice of ability and vendors often report them at their best settings. Treat them as a rough guide.
More models from this vendor
- Muse Wonder: Meta's strongest model for long-running agents and coding, still released as a preview.
- Muse Spark 1.3: Meta's balanced model with a 1-million-token window and a fair price. The Muse Spark line also powers the Meta AI assistant.
- Muse Spark Open: Meta's large open model that a company can run on its own servers, so data never leaves its infrastructure.
- Muse Glimmer 30B: Meta's small open model under Apache 2.0 that runs on a single graphics card, even offline.
- Llama 4 Maverick: The larger model in the newest Llama generation, from April 2025. Cheap through providers, but its license excludes people and companies in the EU.
- Llama 4 Scout: The smaller Llama 4 from April 2025, very cheap through providers. Its license grants no rights to people or companies in the EU.
Similar models from other vendors
- Claude Haiku 5.5 (Anthropic): Claude's cheapest and fastest model, built for classification, data extraction and large volumes of simple tasks.
- DeepSeek V4.1 Flash (DeepSeek): DeepSeek's cheap, fast open-weight model with a 1 million token window. In the app and API, though, your data goes to China.
- Gemini 3.5 Flash-Lite (Google): A very fast, low-cost Google model for large volumes of simple tasks, such as pulling data out of documents.
- Mistral Small 4 (Mistral AI): Mistral's cheap, fast open model under Apache 2.0 for everyday tasks through the API or on your own server.
- GPT-6.1 Terra (OpenAI): OpenAI's fast, cheaper model for working with images, operating a computer and everyday office work.
- Qwen3.8-Flash (Alibaba (Qwen)): A very cheap, fast Qwen model for high volumes of tasks. Unlike Max and Plus, it runs in Frankfurt with processing kept in the EU.