What it is
Muse Spark 1.3 is a closed-weights model from Meta Superintelligence Labs, released on September 2, 2026. The Muse Spark family launched in April 2026 as the successor to Meta's Llama line. Since this release Meta has added Muse Wonder (October 12), a larger hosted model at $4 input and $20 output per million tokens, and Muse Spark Open (October 15), an 816-billion-parameter open-weight version of Muse Spark. Meta's docs now pitch Muse Spark at high-volume, latency-sensitive or cost-sensitive work and Wonder at the hardest tasks.
It is built for agent sessions and coding, with a context of 1,048,576 tokens. It accepts text, images, video, audio and PDF, though Meta's model docs say audio understanding in 1.3 is not fully supported and recommend Muse Spark 1.2 or Muse Voice Transcribe for audio.
It is sold through the Meta Model API and Meta's Muse Code coding tool, and is listed on OpenRouter. Standard mode costs $1.25 per million input tokens and $4.25 per million output; Contributor mode costs $0.10 and $0.20 in exchange for letting Meta train on your prompts and outputs.
Our review
Muse Spark 1.3 is a sound agent model whose place in Meta's line-up changed within six weeks of release. On 12 October Meta launched Muse Wonder as its frontier model at $4 input and $20 output per million tokens. Three days later it released Muse Spark Open, an 816-billion-parameter mixture-of-experts model with 42 billion active parameters, under its own Muse Model License. Meta's docs now describe Spark as the choice for sub-agents, classification and tight tool loops rather than the hardest work.
Meta's case for 1.3 is behavioural, not numerical. It says the model asks clarifying questions when a prompt is ambiguous, confirms before consequential actions and admits when it is stuck, and that its engineers measured about 20% fewer tool calls and 25% fewer tokens than 1.2 on the same work. Those are internal comparisons. The benchmark chart is published as an image against GPT-5.6 Sol and Claude Opus 5, and TrueFoundry points to Meta's own caveat that its runs of rival models may not reflect their best performance.
The pricing is where the real decisions sit. Standard mode charges the same per token whether the context is nearly empty or nearly full, which few rivals do at 1M tokens. Contributor mode cuts the price by more than 90% in return for training rights over prompts, files and outputs, and the switch is a model ID suffix, muse-spark-1.3-contributor. Muse Code subscriptions default to Contributor mode, which any team pointing the tool at client code has to change on day one.
Two limits are easy to miss. Rate limits are pooled per team, 3,000 requests and 4 million tokens a minute across every key, so one batch job can starve an interactive product. And audio input went backwards in this version. For buyers whose objection was closed weights, Muse Spark Open is listed on OpenRouter at the same $1.25 and $4.25 and can also run on their own hardware, and Meta has published nothing that says how 1.3 compares with it.
Where it earns its place
A platform team running sub-agents beneath a larger orchestrator model
Handle the high-volume tool calls and classification steps of an agent pipeline
Meta's docs position Spark for exactly this beneath Muse Wonder, and the two share endpoints, so moving a step between them is a model ID change. Wonder costs 3.2 times as much per input token and 4.7 times as much per output token.
A software engineer at a legal-tech firm querying whole contract sets or codebases
Ask questions across hundreds of thousands of tokens without chunking
There is no long-context premium, so a 900,000-token prompt costs about $1.13 of input at Standard rates, or about $0.14 when it repeats a cached prefix at $0.15 per million.
A two-person start-up prototyping an agent on synthetic data
Keep API spend low through integration testing
Contributor mode makes 50M input and 50M output tokens cost about $15 instead of about $275. Meta trains on everything sent, the tier is capped at 150 requests a minute and it lacks the max reasoning level, so it should never see customer data.
A compliance officer at a software consultancy that has adopted Muse Code
Keep client source code out of Meta's training data
Muse Code subscriptions default to Contributor mode, so the setting must be switched off for every seat, and API traffic restricted to muse-spark-1.3 rather than the -contributor ID, ideally at a gateway, since the two differ only by suffix.
Before you commit
- Setup
- Create a team and API key at dev.meta.ai; the API accepts Responses, Chat Completions and Messages formats, and Muse Code installs on macOS or Linux with one curl command.
- Integrations
- Tool calling, structured output, prompt caching and web search grounding (billed at $2.50 per 1,000 queries); also listed on OpenRouter at the same token prices.
- Data and privacy
- Standard-mode content is not used for training; Contributor-mode content may be used to train future Meta models, and Muse Code subscriptions default to Contributor. Closed weights, so no self-hosting; Muse Spark Open is the self-hostable sibling.
- Cost at scale
- A million support replies a month at 3,000 input and 500 output tokens each is 3B input and 500M output tokens: about $3,750 plus $2,125, or $5,875 at Standard rates, less where prompts hit the cache.
- Limits
- 1,048,576-token context; 3,000 RPM and 4M TPM per team on Standard, 150 RPM and 3M TPM on Contributor, shared by every key. The max reasoning level is Standard-only, and audio input quality is degraded.
What's good
- $1.25/$4.25 per 1M tokens with no long-context premium
- 1M-token context; cached input at $0.15
- Trained to ask, confirm and admit limits in agent loops
- Responses, Chat Completions and Messages API formats
What's not
- Contributor mode, the Muse Code default, trains on your content
- Rate limits pooled per team, not per key
- Audio input degraded in this version
- Benchmarks only as an image, against older rivals
Best for
Teams that want a cheap, long-context model for sub-agents and tool loops and are content to send prompts to Meta's API.
Alternatives
Muse Spark 1.3: common questions
- Is Muse Spark 1.3 worth it?
- Cheap, long-context and well-mannered in agent loops, but squeezed from above by Muse Wonder and from the side by an open Spark at the same price. We rate it 4.0 out of 5.
- How much does Muse Spark 1.3 cost?
- Standard: $1.25 input, $0.15 cached input, $4.25 output per 1M tokens. Contributor mode: $0.10 / $0.002 / $0.20, and Meta may train on your content.
- What are the downsides of Muse Spark 1.3?
- Contributor mode, the Muse Code default, trains on your content; rate limits pooled per team, not per key; audio input degraded in this version; benchmarks only as an image, against older rivals.
- Who is Muse Spark 1.3 best for?
- Teams that want a cheap, long-context model for sub-agents and tool loops and are content to send prompts to Meta's API.
- How much does Muse Spark 1.3 cost for a team or in production?
- A million support replies a month at 3,000 input and 500 output tokens each is 3B input and 500M output tokens: about $3,750 plus $2,125, or $5,875 at Standard rates, less where prompts hit the cache.
- Is Muse Spark 1.3 safe for company data?
- Standard-mode content is not used for training; Contributor-mode content may be used to train future Meta models, and Muse Code subscriptions default to Contributor. Closed weights, so no self-hosting; Muse Spark Open is the self-hostable sibling.
- What are the best Muse Spark 1.3 alternatives?
- The closest alternatives we have reviewed are Muse Glimmer, GPT-6.1 Sol, Claude Sonnet 5.5, Gemini 3.8 Flash, Muse Wonder and Muse Spark Open.
Answers checked Oct 23, 2026.
Make Muse Spark 1.3? Show this rating on your site
