Image: Unite.AI / unite.ai Meta Releases Muse Spark 1.3 with a Cheaper Training-Data Tier
Meta's agentic coding model gets stronger benchmarks, 20% fewer tool calls, and a new Contributor tier at $0.10 per million input tokens — 21x cheaper than the standard price, in exchange for your prompts and completions being used for training.
Meta released Muse Spark 1.3 on September 3, updating its closed-weights reasoning model for long agent runs and coding tasks. The headline numbers are 75.4% on DeepSWE 1.1 and 88.8% on Terminal-Bench 2.1, but the more interesting addition is a pricing tier priced at 21x less than standard in exchange for training access.
What changed in the model
Muse Spark 1.3 finishes the same coding work with about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2. For agentic tasks that involve many tool interactions across large codebases, that’s a meaningful efficiency gain: shorter sessions, lower per-task cost, less time waiting.
The model supports up to 1 million tokens of context (same as 1.2), handles text, images, and video, and is available in Muse Code (Meta’s IDE integration) and the Meta Model API on the day of release. Meta also says it scored 59.4% on SWEAtlas CodeBase QnA and 98.5% on long-context retrieval benchmarks.
On single-tool-call benchmarks Muse Spark 1.3 also beats OpenAI’s recently released GPT-6 Astra in some categories, though comparing across benchmark sets is tricky given how different test conditions can be.
The Contributor tier
The bigger business story is the Contributor pricing tier: $0.10 per million input tokens and $0.20 per million output tokens. Standard 1.3 pricing stays at $1.25 input and $4.25 output (identical to 1.2). The Contributor tier is 21x cheaper on input.
The catch: prompts and completions sent on the Contributor tier feed into Meta’s training data. This is clearly documented and opt-in by design (you select the tier), but it makes the tier a poor fit for anything involving proprietary code, internal tools, or sensitive data.
For developers building public-facing tools, experimenting with new workflows, or doing non-proprietary research, the economics look quite different. At $0.10 per million input tokens, the cost of running a long agentic session approaches near-zero.
What it means for the coding model landscape
The prior Muse Spark 1.2 was already competitive with Anthropic’s Claude Fable 5.1 on coding benchmarks. Meta’s Chief AI Officer Alexandr Wang called 1.3 the “biggest jump so far on model performance” in the Muse Spark line. That’s internal framing worth taking with some skepticism, but the DeepSWE and Terminal-Bench numbers are public and independently verifiable.
The Contributor tier pricing is the more novel move. If developers adopt it widely for non-sensitive tasks, it gives Meta an ongoing training data advantage that compounds over time. It also pressures competitors to consider similar programs or reduce prices.
Sources:
- Meta Launches Muse Spark 1.3, Citing Gains in Coding and Agentic Tasks — Unite.AI
- Meta AI Released Muse Spark 1.3 — MarkTechPost