Models
Muse Spark: what Meta's model is, what it costs, and how to use it free
Meta's first Muse model, from launch to version 1.3: what it's good at, how it scores on independent coding tests, what it costs, and where to use it without paying.
Muse Spark is the AI model family Meta Superintelligence Labs launched in April 2026, the first of its Muse models. The current version, Muse Spark 1.3 (September 2, 2026), is built for agents and coding, reads text, images, video and PDFs, and holds a 1M-token context. On Meta's API it costs $1.25 per million input tokens and $4.25 per million output, and in standard mode Meta doesn't train on what you send. You can use it free in the Meta AI app, in Meta's Muse agent app, and as one of the free models in Poly. Meta has since released a stronger family, Muse Wonder, and an open-weight Muse Spark Open, and independent coding tests put Spark 1.3 well behind Claude.
Key takeaways
- Maker: Meta Superintelligence Labs. Versions: April 8 (first), 1.1 July 9, 1.2 August 5, 1.3 September 2, 2026.
- Strengths: long agent runs with fewer tool calls; Meta says 1.3 uses about 25% fewer tokens than 1.2.
- API price: $1.25 / $4.25 per million tokens, cached input $0.15; a Contributor mode at $0.10 / $0.20 lets Meta train on your data.
- Coding scores: 24.8% at max effort on Vals AI's Terminal-Bench 4.0, against 64.1% for Claude Sonnet 5.5 and 36.4% for Haiku 5.5.
- Free: in the Meta AI and Muse apps, and as a free model in Poly.
| Muse Spark 1.3 | Claude Haiku 5.5 | Claude Sonnet 5.5 | |
|---|---|---|---|
| Price per million tokens (input / output) | $1.25 / $4.25 | $0.10 / $0.50 | $2 / $10 |
| Context | 1M | 1M | 1M |
| Terminal-Bench 4.0 (Vals AI) | 24.8% max, 10.6% default | 36.4% | 64.1% |
| LMArena WebDev | #19 (max) | #36 | #3 |
| Trains on API data | No (standard mode) | No | No |
| Free to use in Poly | Yes | Yes | Yes |
Sources: Meta's launch post, 1.3 post, models and pricing docs; Vals AI, LMArena and Artificial Analysis, opened October 29, 2026.
What is Muse Spark?
Meta introduced Muse Spark on April 8, 2026 as "the first in the Muse family of models developed by Meta Superintelligence Labs," and it powered the Meta AI app and meta.ai from launch (Meta). Each version since has leaned further toward agents and code:
- 1.1 (July 9) came with the public preview of Meta's Model API.
- 1.2 (August 5) launched alongside Muse Code, Meta's terminal coding agent.
- 1.3 (September 2) is the current one. Meta's engineers saw "~20% fewer tool calls and ~25% fewer tokens" than 1.2 on the same work, and Meta labels it for "agentic workflows and coding."
Muse Spark 1.3 takes text, images, video, audio and PDFs (Meta suggests 1.2 for audio) and answers in text, with a 1,048,576-token context.
The rest of the Muse family: Muse Wonder and Wonder Full (October 12), which Meta calls its frontier models; Muse Glimmer (August 10), small and open-weight for devices; and Muse Spark Open (October 15), an open-weight model with 816 billion parameters, 42 billion active, under Meta's Muse Model License. Muse Spark 1.3 itself is closed; Meta lists its open release only as a plan.
How good is Muse Spark at coding?
Independent results for Muse Spark 1.3:
- Vals AI, Terminal-Bench 4.0 (long agent tasks, October 29): 24.75% at max effort, 28th of 58; 10.61% at default effort. For comparison, Claude Opus 5.5 65.15%, Sonnet 5.5 64.14%, Muse Wonder 49.49%, Claude Haiku 5.5 36.36% and GPT-6 Luna 13.64%.
- LMArena WebDev (people voting on web apps, October 27): 19th at max effort (1656), ahead of Haiku 5.5 (36th) and GPT-6 Luna (41st), behind Sonnet 5.5 (3rd).
- Vals AI, Vibe Code Bench: 85.9% at max effort, 15th of 122.
- Artificial Analysis Intelligence Index: 36 at max effort, level with Haiku 5.5 and above GPT-6 Luna's 28.
The picture: Muse Spark 1.3 is a capable everyday model for building web apps and running long jobs cheaply, at about Haiku 5.5's level, and well short of Claude Sonnet and Opus on hard agent tasks. Turning its effort up to max more than doubles its Terminal-Bench score. For Meta's strongest coder, look at Muse Wonder, which Cursor now calls Meta's flagship.
How much does Muse Spark cost?
On Meta's Model API, per million tokens (Meta):
- Standard mode (
muse-spark-1.3): $1.25 input, $4.25 output, $0.15 cached input. "Your content is not used to train Meta models." - Contributor mode (
muse-spark-1.3-contributor): $0.10 input, $0.20 output, $0.002 cached, in exchange for Meta's "permission to use your content to train future Meta models." Muse Code subscriptions default to it.
There's no long-context surcharge and no listed free API tier; it's pay as you go, with limits of 3,000 requests and 4 million tokens a minute in standard mode. Standard Muse Spark sits between Claude Haiku 5.5 ($0.10 / $0.50) and Sonnet 5.5 ($2 / $10) on price (Claude API pricing).
How can you use Muse Spark for free?
- The Meta AI app and meta.ai, which Muse Spark has powered since launch. Meta doesn't say which version runs there now.
- Meta's Muse app, a personal agent on iOS, Android, muse.ai and inside WhatsApp, "powered by Muse Spark" and "free for most of what people need," launched in the US on September 8 and in Singapore in late October.
- Poly (usepoly.co): Muse Spark 1.3 is one of the free models every account starts with, alongside Claude Sonnet 5.5 and Haiku 5.5, GPT-6 Luna and Grok Build. Poly calls Meta's API in standard mode, so Meta doesn't train on your work, and nobody needs a Meta account or API key.
Can you use Muse Spark in coding tools?
Yes. Meta's API speaks Anthropic's Messages format at https://api.meta.ai, and its docs cover Claude Code, Codex, OpenCode, Goose and its own Muse Code (Meta). Cursor lists Muse Spark 1.3 as a built-in model billed outside its included usage, with 300K context by default and 1M in Max Mode (Cursor).
In Poly, Muse runs inside Claude Code in the same room as Claude, GPT and Grok: each member picks their own model, everyone sees every turn, and the room keeps one shared timeline. When a conversation moves between Muse and Claude, Poly tells the model who wrote each earlier answer, so neither claims the other's work (multi-model AI coding).
Muse Spark vs Claude
- Hard agentic coding: Claude Sonnet 5.5 or Opus 5.5, by a wide margin.
- Cheap, long-running everyday work: Muse Spark 1.3 or Claude Haiku 5.5, which score alike; Haiku is cheaper per token, Muse Spark has the larger context for files and video.
- Self-hosting: neither; Muse Spark Open is Meta's downloadable model (the best open-weight models for coding).
More comparisons in the best AI model for coding.
Common questions
What is Muse Spark?
Meta's first Muse AI model, from Meta Superintelligence Labs, launched April 8, 2026. The current version, 1.3, is built for agents and coding, reads text, images, video and PDFs, and has a 1M-token context.
Is Muse Spark free?
In Meta's apps and in Poly, yes: it powers the Meta AI and Muse apps, and Muse Spark 1.3 is one of Poly's free models. On Meta's API it costs $1.25 per million input tokens and $4.25 output.
Is Muse Spark good at coding?
It's a capable everyday model, about level with Claude Haiku 5.5. On Vals AI's Terminal-Bench 4.0 it scored 24.8% at max effort, against 64.1% for Claude Sonnet 5.5 and 49.5% for Meta's newer Muse Wonder.
Does Meta train on Muse Spark API data?
Not in standard mode, where Meta says your content isn't used to train its models. The cheaper Contributor mode gives Meta permission to train on it, and Muse Code subscriptions default to that mode.
Is Muse Spark open source?
Muse Spark 1.3 is closed. Meta released a separate open-weight model, Muse Spark Open, on October 15, 2026, under its Muse Model License, which has conditions for large commercial services.