Muse Spark 1.3
Meta's proprietary agentic coding model for Muse Code and the Meta Model API
Context window
1M
Input / 1M tokens
$1.25
Output / 1M tokens
$4.25
Provider
Meta
Meta Model API, per 1M tokens: input $1.25, cached input $0.15, output $4.25, with no long-context premium; web search costs $2.50 per 1,000 queries. The Contributor endpoint (muse-spark-1.3-contributor) costs $0.10 input, $0.002 cached and $0.20 output in exchange for letting Meta use your prompts and responses to train future models. OpenRouter charges the same rates. · Data verified 2026-11-11
Muse Spark 1.3 (September 2, 2026) is an update to Meta's proprietary Muse Spark model, focused on agentic and coding work and on staying on task through long sessions. It has a 1,048,576-token context and accepts text, image and video input. Meta engineers measured about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 on the same tasks, at unchanged prices. In Meta's launch scorecard the max reasoning setting scored 75.4% on DeepSWE v1.1, ahead of GPT-5.6 Sol and Claude Opus 5 at max effort. Meta's larger Muse Wonder (October 2026) scores higher, and Muse Spark Open, released in October, is an open-weight model that shares its base.
How to access it
Meta Model API at https://api.meta.ai/v1 (model ID muse-spark-1.3, or muse-spark-1.3-contributor for the discounted Contributor mode) and Muse Code, whose subscriptions default to Contributor mode. Meta documents the API as compatible with the OpenAI and Anthropic SDKs. Also on OpenRouter as meta/muse-spark-1.3.
Strengths
- ✓75.4% on DeepSWE v1.1 at max reasoning (Meta-reported)
- ✓1M-token context at the same price at any context length
- ✓About 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 (Meta engineers' comparison)
- ✓Contributor endpoint at $0.10/$0.20 per 1M tokens for teams willing to share data with Meta
Best for developers who...
When to choose it (and when not to)
Reach for Muse Spark 1.3 when...
- →You want a coding-focused agent model with a 1M context at mid-range prices
- →You can accept Meta training on your traffic in exchange for the $0.10/$0.20 Contributor price
- →You already use Muse Code
Look elsewhere if...
- ✕You need Meta's strongest model (Muse Wonder scores higher on coding benchmarks)
- ✕You need open weights you can self-host (use Muse Spark Open)
- ✕Your code or data cannot be shared: use the standard endpoint, not Contributor mode
How to use it
- ›Keep a long task in one conversation: the 1.3 update targets sustained multi-step sessions
- ›Call the standard muse-spark-1.3 ID, not the Contributor ID, for confidential code
- ›Cache long repeated prefixes such as repository context to pay the $0.15 cached rate
Quickstart
Pythonimport os
from openai import OpenAI
client = OpenAI(
base_url="https://api.meta.ai/v1",
api_key=os.environ["MODEL_API_KEY"],
)
response = client.responses.create(
model="muse-spark-1.3",
input="Find the bug in this function and suggest a fix: ...",
)
print(response.output_text)Meta's own Model API quickstart uses the OpenAI Python SDK (pip install openai) pointed at api.meta.ai; Meta does not ship a separate Python SDK. Set MODEL_API_KEY to your Meta Model API key.
API model id: muse-spark-1.3
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| DeepSWE v1.1 | 75.4% (max) | Meta launch scorecard as reported by DataCamp and i-SCOOP; GPT-5.6 Sol (max) 73.0%, Claude Opus 5 (max) 74.0% |
| Terminal-Bench 2.1 | 88.8% (max) | Meta launch scorecard as reported by DataCamp; tied with GPT-5.6 Sol (max) |
| SWEAtlas CodeBase QnA | 59.4% (max) | Meta launch scorecard; GPT-5.6 Sol 53.5%, Claude Opus 5 52.7% |
| MRCR, 512K to 1M | 98.1% (max) | Meta launch scorecard long-context test; GPT-5.6 Sol 73.8% |
Compare Muse Spark 1.3
Compare Muse Spark 1.3 with any other model
Build a comparison →All model comparisons →