01
Muse Spark 1.3 release and availability
Meta released Muse Spark 1.3 on September 2, 2026 and says it began rolling out that day in Muse Code and the Meta Model API. The update focuses on coding, tool use and longer-horizon agent work.
Meta says the reasoning modes that were already available remain usable, while a max-reasoning mode is still pending additional safety testing. Do not treat max reasoning as generally available until Meta updates the product surface you plan to use.
02
What changed in Muse Spark 1.3?
Meta says 1.3 is better at sustaining several workflows in one long thread, using tools to fill missing context, correcting gaps in its plan and retaining what it has learned before producing a final deliverable.
The company also says the model reduces unnecessary turns and token use versus Muse Spark 1.2. Those are vendor claims and should be validated on representative workloads before a production migration.
03
Context window and multimodal inputs
Vercel AI Gateway currently lists a 1 million-token context window for meta/muse-spark-1.3. Its model page documents text, image and PDF inputs, making the model relevant to repository-scale coding, document analysis and long agent histories.
A large context window is a capacity limit, not a guarantee of perfect retrieval or reasoning across every token. Teams should compare large-prompt workflows with retrieval, summarization and caching strategies on their own data.
04
Muse Spark 1.3 pricing on Vercel AI Gateway
At publication time, Vercel lists the standard model at $1.25 per million input tokens, $4.25 per million output tokens and $0.15 per million cached-input read tokens. These are Vercel AI Gateway rates, not a universal Meta API price.
Vercel also lists meta/muse-spark-1.3-contributor at $0.10/M input, $0.20/M output and $0.002/M cached-input reads. Vercel labels this lower-cost route as permitting submitted usage to improve Meta’s products, so the data-use tradeoff must be reviewed before sending confidential or regulated material.
05
Coding and agent use cases
Meta positions Muse Spark 1.3 for coding agents and long-running tool workflows. Practical evaluation targets include repository analysis, feature implementation, multi-file debugging, research agents and business-process automations that need to gather context before acting.
For agent systems, cost per million tokens is only one variable. Repeated context, cache-hit rate, retries, tool loops and first-attempt task success can change the real cost per completed workflow.
06
How to interpret Muse Spark 1.3 benchmark claims
Meta’s launch material presents Muse Spark 1.3 as a meaningful improvement on coding and agentic evaluations. Those results are useful evidence of the workloads Meta optimized for, but they remain vendor-published until comparable independent testing accumulates.
07
Current limitations and what to recheck
Meta’s public launch post does not consolidate every direct-API price, region, quota and provider-specific limit on one page. Max reasoning is not yet generally available, and third-party gateway pricing can change independently of Meta’s own API.
Muse Spark 1.3 should also be treated as an API and product release rather than an open-weight release unless Meta separately publishes downloadable weights. Recheck access, pricing, output limits, rate limits and data terms before a production commitment.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.