Meta Muse Spark 1.3: agentic coding, thought more efficiently
Meta shipped Muse Spark 1.3 in September 2026 — the next iteration of its closed-weight flagship from the Meta Superintelligence Labs, a little over a month after Muse Spark 1.2 (August 5). The emphasis stays on agentic coding, paired with Meta’s efficiency approach, thought compression. What’s newly solid are mainly two figures: the price and coding performance against the current competition.
What's new
- Release: September 2026, successor to Muse Spark 1.2 (August 5, 2026).
- Focus: agentic coding at the repository level, building on the co-training with Meta’s Muse code agent, parallel tool calls and the 1M-token context from 1.2.
- Coding (short tasks, third-party comparison): “xhigh” mode 73, “max” mode 72 — front of the field, just behind DeepSeek V4.1-Flash (74.2).
- Pricing: $1.25 input / $4.25 output per 1M tokens — well below Western frontier models such as Claude Opus 5 ($5/$25) or GPT-6 Astra ($10/$50).
What happened
Muse Spark is Meta’s break with the open Llama legacy: a closed-weight line from the Meta Superintelligence Labs, built for efficient reasoning via thought compression — a second training signal penalises the length of the chain of thought, so the model learns to reach a solution with fewer reasoning tokens. Version 1.2 put the focus squarely on agentic coding: a 1M-token context, co-training with the in-house Muse code agent, and parallel tool calls.
1.3 builds directly on that. The most striking new data comes from a third-party comparison: on short coding tasks, Muse Spark 1.3 scores 73 in “xhigh” mode and 72 in “max” mode — front of the field, but just behind DeepSeek’s V4.1-Flash (74.2). Pricing sits at $1.25 input and $4.25 output per million tokens.
Why it matters
The interesting part is the price-performance position. At $1.25/$4.25, Muse Spark 1.3 costs a fraction of what the Western frontier models charge for output — Claude Opus 5 sits at $5/$25, Claude Fable 5.1 and GPT-6 Astra at $10/$50 each. On short, well-scoped coding tasks, Muse Spark 1.3 plays at the front of the field without sitting at a frontier price.
For Meta, the pattern is notable: a closed line that iterates on a monthly cadence and deliberately prices itself between the cheap Chinese open-weight models and the expensive Western flagships. If you run agentic coding workloads at volume, that’s another option whose case rests mainly on the output price.
One caveat remains on the data. The coding figures come from a third-party comparison, not broad independent testing, and Muse Spark stays closed-weight — you can inspect neither the weights nor the training details. For auditable or documentation-bound processes that’s a selection criterion in its own right, independent of price.
What you can do now
If you automate agentic coding at volume: test Muse Spark 1.3 against your current model on a real task and weigh the saving on output price against the quality of your sample.
If you want to place the Muse line: background on thought compression, the closed-weight strategy and the contemplating mode is in the glossary entry → Muse Spark and in the fuller read in the lexikon → Muse Spark.