Meta released Muse Spark 1.1 on July 9, a multimodal agentic model that lands right inside the frontier cluster. The pricing sheet comes out at $1.25 per million input tokens and $4.25 per million output tokens, which parks Muse Spark 1.1 at the aggressive end of that group. The pitch fits in one line: hit the frontier tier without paying the frontier bill.
Key Takeaways
- Muse Spark 1.1 is the new model from Meta Superintelligence Labs, multimodal and agentic, with a self-managed 1 million token context.
- On the Artificial Analysis Intelligence Index, it ties with GLM-5.2 max at 51 points, inside a cluster that also includes Grok 4.5 and Claude Opus 4.8.
- Cost per benchmark task drops to $0.26 against $0.37 for GLM-5.2, with the model shipping in public preview via the Meta Model API.
Muse Spark 1.1 lands in the frontier cluster
Meta pushed Muse Spark 1.1 into public preview on July 9, 2026, through the Meta Model API and the Thinking mode inside the Meta AI app. The model carries the Meta Superintelligence Labs stamp, the frontier arm that Zuckerberg has been rebuilding around large model bets.
On the Artificial Analysis Intelligence Index, Muse Spark 1.1 comes in at 51 points. That level ties with the max variant of GLM-5.2 and drops Meta into the same competitive pack as Grok 4.5, Claude Opus 4.8, GPT-5.5 and GLM-5.2. The tracker keeps that pack behind a single model, Claude Fable 5, which still tops the composite benchmarks.
What separates Muse Spark 1.1 from Muse Spark is the agent stack. Meta wrote in native primary-agent plus subagent orchestration, added MCP and custom-skill zero-shot support on unseen tools, and shipped a 1 million token context managed by the model itself on long sessions. The lab framed it as a model tuned for long agentic tasks, from heavy debugging to full codebase migration.
The step from Muse Spark to Muse Spark 1.1 also adds eight Intelligence Index points in three months according to the Artificial Analysis note. That kind of climb pulls Meta closer to the top labs of the year, without flipping the ranking outright.
Meta’s price sheet pressures the Grok 4.5 GLM-5.2 pack
Meta priced Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens. The sheet sits noticeably under the entry tickets on long workflows for Claude Opus 4.8 and GPT-5.5, at the tail of a cluster where price is becoming the real differentiator now that the Intelligence Index gap has narrowed.
The most striking number is the per-task cost. Muse Spark 1.1 burns fewer output tokens for the same benchmark load: Artificial Analysis tracks $0.26 per Intelligence Index task, against $0.37 for GLM-5.2. The gap is clean and documented on the same agentic run.
Meta arrives in the middle of a wider pricing squeeze on the lab side. Our reporting on Fable 5 moving from subscription to API-token billing this week already flagged how Anthropic is repricing its own stack. Xai had opened the front with Grok 4.5 undercutting Claude Opus at a third of the price. Meta pushes the same logic on a cluster-top model.
For product teams pricing agents in production, the trade-off shifts: at equivalent quality on orchestration, the provider who holds the cost curve on long sessions becomes the default choice. Muse Spark 1.1 slots straight into that shift without waiting.
More articles on Horizon
- Grok 4.5 Test: We Rate Musk’s Coding Promise
- Grok 4.5 Undercuts Claude Opus at Third the Price
- GPT-5.6 Test: We Rate Sol, Terra and Luna
Coding trails GLM-5.2, agent orchestration leads on long workflows
On pure coding, GLM-5.2 keeps the edge on SWE-bench Verified and Terminal-Bench, at the 87th percentile versus 82nd for Muse Spark 1.1. The gap is not huge, but it confirms Meta is not aiming for the top spot on classic coding leaderboards.
Where Meta is stacking chips is on multi-agent orchestration and tool use. Muse Spark 1.1 delegates to subagents in parallel, keeps an extended context across sessions that touch several tools, and generalizes zero-shot to MCP servers it has never seen. The lab reframed the story around enterprise workflow: multimodal perception on images, videos and PDFs, action on software interfaces, without scripting between each step.
The competitive roadmap now runs on two tracks. OpenAI just kicked off the global rollout of GPT-5.6 as a mass-coverage play. Meta answers with a model that targets unit cost on prolonged use and pairs API distribution with the consumer Meta AI app. Two distinct plays on the same segment.
The competitive impact runs on the coming weeks. Zhipu, whose GLM-5.2 scores just got challenged on the cost line by Muse Spark 1.1, will likely rework the API plan. Anthropic, whose Fable 5 still leads the composite, will probably recalibrate its Sonnet and Opus tickets against this new floor. The public debate around the AI “bubble” is drifting from the parameter race to the cost of production being held.
Muse Spark 1.1 is not the best model in the room. It is a frontier model priced at a level that forces the others to answer. On a market that keeps tightening around a cluster of narrow gaps, that kind of move counts for more than a “SOTA” stamped release.
Follow the story on Horizon.
