Meta Muse Spark 1.2

Muse Spark 1.2 is Meta’s proprietary Frontier model from August 5, 2026, designed for coding and agentic workflows, released by Meta Superintelligence Labs alongside the terminal agent Muse Code, with which it was co-trained. The cloud-only model under US jurisdiction (CLOUD Act) processes text, image, video, and audio with a context of 1,048,576 tokens (max. 131,072 output) and offers configurable reasoning effort up to ‘xhigh’.

Meta Version 1.2 Commercial use permitted Dense 1024 K Context $1.25 / $4.25 per 1M

Proprietary
Frontier
OpenRouter
Text
Vision
Audio
Video
Long Context
Agentic Orchestrator
Real-Time

Sovereign Risk: MEDIUM Meta is a US-based company and subject to the CLOUD Act; the model weights are proprietary and not publicly accessible (no self-hosting, no fine-tuning possible). Meta additionally offers a ‘Contributor’ pricing tier in which users agree, in exchange for significantly reduced costs, that their prompts may be used to train future Meta models — under this tier, the actual data risk increases considerably compared to the standard tier.

Tool-use details

Asset performance, reliability, and runtime profile

Name	P1	P2	Combined
EU License Research	100	40	64
HTTP Fetch & Extract	80	40	60
Tool Failure Handling (404)	40	40	37
Web Search & Tool Selection	100	80	91
URL Construction & Fetch	80	80	80
Multilingual Search & Synthesis	100	80	90

Name

Combined

EU License Research

100

HTTP Fetch & Extract

Tool Failure Handling (404)

Web Search & Tool Selection

100

URL Construction & Fetch

Multilingual Search & Synthesis

100

CrucibleMark evaluates tool use across 6 independent tests. Click a test name for details.

Reliability

Tool Call Valid: No

Retry: Not required

Hallucination: Not detected

Reliability measures how consistently a model actually executes tool calls: Tool Call Valid schema and format accepted, Retry Required successful only after retry, Hallucination Flag fabricated tools or parameters detected. All three green means production-ready.

Runtime profile

Call 1: 2.87; First Request

MCP: 2.18; Protocol Latency

Synthesis: 14.38; Response Generation

Total: 116.58; Sum of All Phases

Token: 18517; Input + Output

Cost: $0; Cost per Run

The runtime profile shows the latency and cost metrics for the model run: Call 1 First Request, MCP Protocol Latency, Synthesis Response Generation, Total sum of all phases, supplemented by Token input and output and Cost cost per run.

Deployment Verdict

Conditional deploy, because despite a solid overall score of 70.33 the model fails to produce a consistently valid tool-call trace and is therefore only viable in MCP pipelines under tight runtime supervision.

Tool-Execution Profile

The model demonstrates genuine tool-selection intelligence, but not protocol reliability at a consistent production level. In the Web Search & Tool Selection test — which checks whether the model distinguishes between search and direct retrieval without being prompted — it selects the correct tool with confidence. That argues against rigid pattern-matching behavior. It also performs reliably on the tool side in the Multilingual Search & Synthesis test. Execution is where it weakens: in the URL Construction & Fetch test, which measures correct URL derivation plus clean retrieval, performance is serviceable but not deterministic. HTTP Fetch & Extract confirms this picture. Retrieval works often; precision in the execution sequence does not always follow. The tool_call_valid: false finding is critical. This is not an indication of a planning deficit — it points to insufficient MCP reliability at the last mile. For orchestrated tool chains, that last mile is exactly what matters.

Synthesis Fidelity

How well does it consolidate tool results? Only reliably to a limited degree. P2 signals land at 40 in several tasks, including EU License Research and HTTP Fetch & Extract — precisely where tight consolidation of retrieved facts is expected. By contrast, Web Search & Tool Selection, URL Construction & Fetch, and Multilingual Search & Synthesis are noticeably stronger in their synthesis. The pattern is clear: solid research orchestration, inconsistent consolidation discipline.

Does it stay within the tool result or fall back on training? In the honeypot EU License Research — which checks whether current license restrictions are actually retrieved from web sources — it does not hallucinate. That is the most important trust finding. The model does not visibly drift into freely fabricated compliance answers, even when the consolidation remains too shallow.

Error Resilience

In the Tool Failure Handling (404) test, which measures transparent behavior when a retrieval fails, the model does not fabricate substitute content. That is acceptable for production. Answer quality remains tight at P2 40, however. It stays defensive rather than genuinely robust. For pipelines, that is preferable to creative error compensation.

Operational Profile

Total 116.58s. Call 1 2.87s. MCP latency 2.18s. Call 2 14.38s. Slow for the utility demonstrated. Cost per run: local.

Conclusion & Recommendation

Suitable for supervised agent pipelines where tool selection, research planning, and multilingual search matter more than strictly deterministic MCP execution. Not suitable for compliance, registry, or ETL pipelines where every tool call must be formally valid and every consolidation must stay tightly anchored to retrieved content. Anyone deploying it should make hard tool-call validation, output schema checking, and a downstream verifier mandatory.

This analysis was generated automatically based on benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and full methodology are documented in the GitHub project.