Meta Launches Muse Spark 1.1, a 1M-Token Coding Model

Meta Superintelligence Labs has publicly launched Muse Spark 1.1, a multimodal reasoning model built for agentic coding tasks, entering a market already occupied by OpenAI and Anthropic. The release lands the same week as OpenAI's GPT-5.6 and a new version of Grok from SpaceX AI, signaling an unusually competitive stretch for AI coding tools.

What Meta announced

Muse Spark 1.1 is available now in Thinking mode within the Meta AI app and on meta.ai. Meta has also opened a public preview of its new Meta Model API, giving developers direct access to the model.

The headline technical feature is a 1 million token context window, which the model can actively manage — a capability Meta positions as suited for agentic workflows that require processing large amounts of code or data at once.

On pricing, Meta lists Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens.

Meta says it conducted "extensive safety evaluations" before deployment, following what it calls its Advanced AI Scaling Framework, and states the model shows strong resistance to direct jailbreaks as well as indirect attacks such as prompt injection and untrusted-data exploits.

The timeline

April: The first version of Muse Spark was announced.
Tuesday: Meta unveiled Muse Image, a new AI image-generation model.
Thursday: Meta publicly launched Muse Spark 1.1 — the same day OpenAI released GPT-5.6.
This week: SpaceX AI released a new version of Grok.

Sources differ on some details of this timeline. Meta's own announcement describes Muse Image as launching "this week alongside Muse Spark 1.1," while TechCrunch reports the two products launched on different days — Muse Image on Tuesday, Muse Spark 1.1 on Thursday. There's also an inconsistency in reporting around Mark Zuckerberg: TechCrunch notes his last post on X was in July 2023, yet also quotes a more recent-sounding X post from him calling Spark "a strong agentic and coding model at a very low price." The report does not resolve this discrepancy.

Partner reactions

Several partners were quoted praising the model, though these are vendor perspectives rather than independent assessments:

Amjad Masad, CEO of Replit, said Muse Spark "packs a massive million-token context, full multimodal support, built-in search with citations, strong reasoning, and top-tier coding abilities."
Saoud Rizwan, CEO of Cline, said Meta is "building for serious agentic coding with strong tool use at a price viable for real coding workloads at scale."
Yashodha Bhavnani, VP of AI Products at Box, said Muse Spark "delivered enterprise capabilities competitive with today's leading frontier models."

What's missing

The report notably lacks independent verification on several fronts. There are no third-party benchmark results comparing Muse Spark 1.1 against GPT-5.6, Claude, or Grok. Pricing is stated but not benchmarked against competitor rates. Meta's safety claims — including jailbreak resistance and the specifics of its Advanced AI Scaling Framework — are not independently detailed or verified. There's also no data yet on developer adoption or usage volume since the public preview opened.

Why founders should care

For founders building or evaluating AI-assisted coding tools, this launch likely matters in a few ways, though with meaningful caveats:

The 1 million token context window may plausibly enable handling of larger codebases or documents in agentic workflows — but without independent benchmarks, real-world performance gains are uncertain.
The stated pricing ($1.25/$4.25 per million tokens) could represent a cost-competitive entry point for running coding workloads at scale, though founders should directly compare this against current OpenAI and Anthropic rates rather than relying on Meta's framing alone.
Partner endorsements from Replit, Cline, and Box may signal genuine enterprise interest, but they are vendor quotes, not neutral benchmarks — founders would be wise to treat them as directional signals rather than proof points.
Given the crowded field — Muse Spark 1.1 launching alongside GPT-5.6 and a new Grok version in the same week — founders evaluating agentic coding tools may benefit from tracking multiple releases before committing to a single provider, since differentiation in this space is still unsettled.

Risks to weigh

Meta is entering a market where OpenAI and Anthropic already have established agentic coding offerings, which could limit how much Muse Spark 1.1 differentiates itself. Additionally, safety claims — including jailbreak resistance — come directly from Meta without independent verification, and capability comparisons rely heavily on partner quotes rather than neutral third-party testing. Founders should factor these gaps into any evaluation before adopting the model for production workloads.