GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Meta Opens Muse Spark 1.1 to Developers at $1.25/M — Leads 4 of 12 Agentic Benchmarks

Meta Superintelligence Labs has released Muse Spark 1.1, a multimodal reasoning model built for agentic tasks, alongside the public preview of its first developer-facing API. The Meta Model API is live today in the US, OpenAI SDK-compatible, and starts at $1.25 per million input tokens and $4.25 per million output tokens. New accounts receive $20 in free credits.

That pricing undercuts Claude Sonnet 5 ($2/$10 introductory) by a substantial margin, and sits far below Claude Opus 4.8 and GPT-5.5. Three months after Muse Spark’s initial April launch with no API access, Meta has moved from a consumer-first release to a direct play for developer budgets.

Benchmark Results Across 12 Tests

Meta published a comparison table covering Muse Spark 1.1, the original Muse Spark, Gemini 3.1 Pro, Claude Opus 4.8, and GPT-5.5 across three categories: agents, coding, and multimodal.

Where Muse Spark 1.1 leads:

BenchmarkMuse Spark 1.1Opus 4.8GPT-5.5
MCP Atlas (tool use)88.1~79~80
JobBench (professional tool use)54.748.438.3
HLE with Tools62.157.9—
Finance Agent v257.253.951.8

The JobBench margin is the headline result. At 54.7 against Opus 4.8’s 48.4 and GPT-5.5’s 38.3, it is the widest gap in the entire release table.

Where Muse Spark 1.1 trails:

BenchmarkMuse Spark 1.1Opus 4.8GPT-5.5
Terminal-Bench 2.180.082.783.4
SWE-Bench Pro61.569.2—
DeepSWE 1.153.359.067.0
OSWorld-Verified (computer use)80.883.4—
Toolathlon-Verified75.676.2—

Coding remains Muse Spark 1.1’s weakest area. The 7.7-point gap on SWE-Bench Pro against Opus 4.8 is meaningful. On DeepSWE 1.1 — a long-horizon agentic coding test — Muse Spark 1.1 scores 53.3 against GPT-5.5’s 67.0, a 13.7-point deficit. That said, the original Muse Spark scored just 10.0 on the same test, so the internal trajectory is steep.

On VALS-AI’s independent evaluation, Muse Spark 1.1 ranks fourth overall, while being the best-priced model in that tier.

Muse Spark 1.1 leads 4 of the 12 benchmarks in Meta’s table. Opus 4.8 leads 5, GPT-5.5 leads 3.

Architecture and Agentic Design

The model is built to operate both as a main orchestrator and as a subagent. In orchestrator mode, it gathers context, builds a plan, and delegates to parallel subagents. As a subagent, it runs tasks and escalates when appropriate. Meta says it zero-shot generalises to new native tools, MCP servers, and custom skills without targeted fine-tuning.

Muse Spark 1.1 self-manages its 1M-token context window, deciding what to retain, retrieve, and compress during long agentic runs. This is positioned as central to multi-session project work.

The Closed Weights Decision

Muse Spark 1.1 ships without open weights. Like its predecessor, it is a proprietary closed model — a departure from the Llama strategy that made Meta a hero in the open-source AI community. No Apache 2.0 weights are forthcoming, at least at launch. The positioning is as a commercial API product, not a foundation for community fine-tuning.

Muse Image, Meta’s image generation model, is not available through the API at launch.

Why This Matters

Before today, Meta had no public developer API for any frontier-class model. Muse Spark 1.1 changes that, and does it at pricing that forces a comparison against Sonnet 5 — not against the models it beats in specific agent benchmarks. Developers building tool-heavy or finance-adjacent agents have a credible lower-cost option for the first time. For pure coding workloads, Opus 4.8 and GPT-5.5 remain ahead.