Muse Spark 1.2 and Muse Code

2026-08-05 · Publisher : Meta · Domain : Artificial Intelligence

Meta launches Muse Code, a terminal coding agent, together with Muse Spark 1.2, available in the Meta Model API with two pricing tiers depending on the use of data for training.

Meta launches Muse Code, a terminal coding agent, together with Muse Spark 1.2, available in the Meta Model API with two pricing tiers depending on the use of data for training. ### Context Muse Spark 1.1 (July 2026) opened up the API. Agentic coding has become the main field of competition with Claude Code and Codex, and Meta is responding with its own agent. ### What the model brings - Muse Code: terminal agent (beta, macOS and Linux) with persistent background agents that keep the repository context. - Isolated parallel git worktrees so as not to touch the developer's working copy. - Local append-only event log for replayable execution that can be resumed after a restart. - Built-in skills /plan, /grill and /goal. - Muse Spark 1.2 as the default model, in the API with broader worldwide access. ### Architecture and specifications Size, architecture and knowledge cutoff: not disclosed. Availability: Muse Code and Meta Model API, including OpenRouter according to Wikipedia. License: proprietary; Meta promised on August 10 an open weights release of Muse Spark 1.2 “soon”, with no date; no Hugging Face repository existed at the start of September according to a secondary source. ### Cost and access | Tier | Input | Cached input | Output | Data | |---|---|---|---|---| | Contributor | $0.10 | $0.002 | $0.20 | Used for training (60 requests/min) | | Standard | $1.25 | $0.15 | $4.25 | Not used (3,000 requests/min) | Prices per million tokens, recorded by VentureBeat and Wikipedia (the Meta post that was read does not cite them). ### Benchmarks | Benchmark | Muse Spark 1.2 | Claude Opus 5 | GPT-5.6 Terra | Grok 4.5 | |---|---|---|---|---| | Terminal-Bench 2.1 | 82.9% | 86.7% | 81.8% | 81.6% | | DeepSWE 1.1 | 59.3% | 65.0% | 64.8% | not disclosed | | Meta internal benchmark | 70.6% | 79.4% | 65.4% | not disclosed | Source: charts from Meta's announcement as read by VentureBeat, scores published by the vendor (not read directly). ### Versus closed models On Terminal-Bench 2.1, Muse Spark 1.2 (82.9%) is ahead of GPT-5.6 Terra (81.8%) and Grok 4.5 (81.6%) but remains behind Claude Opus 5 (86.7%). On DeepSWE 1.1, it stands at 59.3% versus 65.0% for Opus 5 and 64.8% for GPT-5.6 Terra. The contributor tier is at $0.20 per million output tokens, versus $4.25 at the standard tier: the gap comes from handing data over to Meta, a compliance criterion to assess. No open weights are available: the promise of openness for version 1.2 has not been kept at the time of writing. ### Key takeaways Meta enters agentic coding with a model close to the leaders on Terminal-Bench, at a very low price but in exchange for data sharing. Sovereignty remains conditional on the future release of the weights.

Tags : LLM, Meta, Muse, Poids fermés, Code, Agents, API