Muse Spark 1.1 + Meta Model API: Preference Share at Undercut Prices

July 9 was already the public unlock for GPT‑5.6 and the week of Grok 4.5. Muse Spark 1.1 is Meta’s answer on the paid coding/agent API front—not open weights, not “Llama 5,” but a closed MSL model sold through a Meta Model API. Pricing plus a top-10 Arena text slot is the independent story: Meta is buying preference share on blind chat while undercutting closed frontier APIs.

Meta on July 9, 2026 introduced Muse Spark 1.1, a major upgrade from the April Muse Spark release, and opened a public preview of the Meta Model API so developers can build on it. Meta positions 1.1 as a multimodal reasoning model built for agentic work—tool and computer use, coding on large codebases, multi-agent orchestration, and a 1M-token context window—available now in Thinking mode in the Meta AI app and on meta.ai. The same week Meta shipped Muse Image (July 7); Watermelon, a larger training run, remains unreleased.

What Meta shipped

From Meta’s primary blog “Introducing Muse Spark 1.1” (ai.meta.com, July 9) and contemporaneous Reuters/CNBC/Axios/NYT coverage:

Model

• Muse Spark 1.1: Multimodal reasoning model with “major gains” in tool/computer use, coding, and multimodal understanding vs Muse Spark.
• Agents: Zero-shot generalization to native tools, MCP servers, and custom skills; multi-agent planning with parallel subagents for latency; active management of a 1 million token context window (memory, retrieval, compaction).
• Computer use: Multi-app workflows; script automation when faster than click-by-click UI; batch action generation; adapts when mid-task context changes.
• Coding: Stronger on large/complex codebases; works with popular agentic coding harnesses. Meta cites large gains on internal Meta Internal Coding Bench.
• Multimodal: Visual-to-code, dense image/video captioning, perception + action loops.
• Safety: Evaluated under Meta’s Advanced AI Scaling Framework; claim of safe margins on chem/bio, cyber, and loss-of-control categories (per Meta evaluation report).

Access & packaging

• Meta AI / meta.ai: Available now in Thinking mode.
• Meta Model API: Public preview—first time external developers get broad API access (April Spark had been private/select partners only).
• Early partner quotes: Replit, Cline, Box, OpenClaw.
• Pricing (CNBC, Wang interview): $1.25 / $4.25 per million input/output tokens; $20 free credits for new API accounts; preview access reported as U.S.-first.
• Alexandr Wang (Meta CAIO): strongest model yet for agentic and coding work; Watermelon still training at much larger scale, later this year.

Claims vs checks

Source	Snapshot	Standing
Arena Text Overall	Board dated Jul 16, 2026	muse-spark-1.1 at 1493 ±8 (preliminary, ~5,729 votes) — #6 overall; only Fable 5 and several Claude Opus thinking variants sit clearly higher
Arena (official posts)	Launch week	Earlier mid-week reads placed Spark 1.1 roughly #17 before more votes moved it up—treat Jul 16 board as the firmer mid-month snapshot
Artificial Analysis	Mid-July cluster writeups	Top-15 Intelligence Index band; strong on MCP Atlas / agentic tool-use slices; not the #1 intelligence model (Fable/Sol still lead the 50+ club)

Internal Meta Coding Bench gains are vendor tables; AA and Arena are the independent cross-checks.

Limits

• API pricing cited via CNBC/Wang interview—confirm against live Meta console.
• Arena ranks moved within the launch week; Jul 16 is a dated snapshot.
• Watermelon remains unreleased; 1.1 is not the larger training bet.
• Preview reported U.S.-first.

Sources

• Meta AI Blog: “Introducing Muse Spark 1.1” (July 9, 2026)
• Meta: Muse Spark 1.1 Evaluation Report / methodology (linked from the launch post).
• Arena.ai Text Overall leaderboard snapshot (July 16, 2026); Arena / Meta social rank notes launch week.
• Artificial Analysis mid-July Intelligence Index cluster coverage (six labs above 50).
• Reuters, CNBC, Axios, New York Times same-day coverage of Muse Spark 1.1 and developer API preview (July 9, 2026).