Model Card | Muse Spark 1.2

🌏 中文版

Model Information

Key Capabilities

• First release paired with Meta's own code agent "Muse Code" (beta), designed for long-running multi-agent collaboration: every sub-agent's trajectory, tool call, and steer/cancel action is observable and replayable
• GDPval-AA v2 (knowledge work evaluation) Elo jumps 260 points to 1631, ranking 5th among all models tested by Artificial Analysis, surpassing Claude Opus 4.8 Max at 1588
• Terminal-Bench v2.1 improves from 78% (1.1) to 80%; τ³-Banking (agentic tool use) from 25% to 27%
• No long-context surcharge: billing rate stays the same regardless of how much of the context window is used; prompt caching is enabled automatically with no extra setup

Benchmark Results

⚠️ AA Intelligence Index and GDPval-AA v2 scores come from independent testing by Artificial Analysis (Meta provided access before the official release). Terminal-Bench v2.1 and τ³-Banking scores are from Meta's official announcement and await independent reproduction.

Comparison with Previous Generation and Competitors

Compared to Muse Spark 1.1, the biggest improvement is in knowledge-work agent tasks: GDPval-AA v2 Elo jumped from ~1371 to 1631, a 260-point gain. This benchmark tests how well a model uses shell access and web browsing to perform real knowledge-work tasks (e.g., preparing presentations, analyzing reports) and is considered by Artificial Analysis to be the most representative measure of general agentic performance. Terminal-Bench and τ³-Banking saw only modest 2pp gains, suggesting this generation's training focused heavily on knowledge-work scenarios rather than pure coding.

Against competitors, Muse Spark 1.2's AA Intelligence Index (57) still trails Claude Opus 5 Max (63.05), GPT-5.6 Sol Max (60.93), and Kimi K3 Max (59.70), and sits slightly below Qwen3.8 Max (58.08). However, in cost per Intelligence Index task, Muse Spark 1.2 comes in at just $0.40 — the cheapest among all models scoring 56 or above. The nearest competitor in the same score range, GPT-5.6 Terra Max ($0.51), costs 22% more.

Pricing remains at $1.25/$4.25, identical to Muse Spark 1.1, effectively an implicit price cut given the performance gains. But in practice, because 1.2's responses are longer (output token usage increased 78% on the same benchmark), the actual cost of running a full Intelligence Index suite rose 36.6% — the rate card didn't change, but the bill did.

What This Means for Agent Developers

The real architectural signal here is Muse Code: this is the first time Meta has shipped a "model + agent product that runs the model" together, rather than just releasing an API endpoint. Muse Code emphasizes observability in multi-agent collaboration — every sub-agent trajectory, every tool call, every steer/cancel action can be replayed, which is critical for debugging multi-agent systems.

• If you're building knowledge-work agents (report generation, data analysis, presentation creation): the massive GDPval-AA v2 Elo improvement is a signal worth evaluating as a replacement in your pipeline, especially in budget-sensitive scenarios — it's the cheapest option in its score range
• If you're building multi-agent systems and struggling with debugging: Muse Code's event log (every sub-agent, every tool call replayable) addresses the common pain point of "something went wrong but you can't tell which step," worth studying for its observability design
• Not ideal for: pure hardcore coding scenarios — Terminal-Bench v2.1 only improved by 2pp, limited competitiveness against coding-focused models like GPT-5.6 Sol Max and Claude Opus 5 Max; also watch out for actual bills rising due to longer responses, even when the rate card hasn't changed

Takeaway

Muse Spark 1.2's rate card is letter-for-letter identical to its predecessor, but because the model's responses grew 78% longer, the actual cost of running the same benchmark suite rose 36.6%. This is a reminder that "pricing unchanged" does not mean "costs unchanged" — when evaluating model upgrades, look at real-world token consumption, not just per-token rates.

References

• Meet Muse Spark 1.2 and Muse Code, the first coding agent from Meta — Meta AI Developer Blog
• Muse Spark 1.2 — Meta for Developers
• Muse Spark 1.2: Improved Agentic Performance at Higher Cost per Task — Artificial Analysis
• Muse Spark 1.2 (xhigh) — Intelligence, Performance & Price Analysis — Artificial Analysis
• Meta Muse Spark 1.2 pricing: the rate card that never changed — eesel AI