# Muse Spark 1.2 Makes Meta’s AI Comeback a Model-and-Agent Systems Bet

**Muse Spark 1.2 returned Meta to frontier-model competition on August 5, 2026.** The more consequential move was launching Muse Code and training the model around its agent runtime. Artificial Analysis showed an Intelligence Index score of about 56.8 on August 7, up from 53.2 for Muse Spark 1.1. The larger change was in agentic work: its Agentic Index rose by roughly 9.6 points, while its Coding Index gained less than one point.

![The official Muse Spark 1.2 model page describes a one-million-token context window and long-running coding workflows.](https://s4.tenten.co/learning/content/images/2026/08/landing-page-1-5.png)

#### Three releases in four months changed Meta’s development cadence

Meta introduced the first Muse Spark in April 2026, released version 1.1 on July 9, and followed with 1.2 on August 5. Three releases in four months suggest that the group has moved from occasional model launches to a short model-and-product iteration loop.

Version 1.2 also has a narrower purpose than 1.1. The earlier release emphasized multimodal reasoning, tool use, computer use, coding, and a one-million-token context window. The new release directs more training compute toward coding, adds more varied environments, and targets repository understanding, debugging, and complete software workflows.

Muse Code arrived in beta on the same day. The terminal agent handles planning, edits, tests, and validation. That pairing complicates any attempt to reduce the release to one leaderboard number. The model matters, but task decomposition, context compaction, subagent management, and tool execution also affect the result.

#### Independent scores point to an agentic gain

The live Artificial Analysis data has already moved beyond the 54-point snapshot that circulated when the release appeared. On August 7, Muse Spark 1.2 scored about 56.8 on the Intelligence Index, compared with 53.2 for version 1.1. Both standard endpoints cost USD 1.25 per million input tokens and USD 4.25 per million output tokens, with a 1,048,576-token context window.

| Artificial Analysis metric | Muse Spark 1.1 | Muse Spark 1.2 | Change |
|---|---:|---:|---:|
| Intelligence Index | 53.2 | 56.8 | +3.6 |
| Coding Index | 71.3 | 72.2 | +0.9 |
| Agentic Index | 39.7 | 49.3 | +9.6 |
| Standard input/output per million tokens | USD 1.25/4.25 | USD 1.25/4.25 | No change |

The restrained reading is that 1.2 improved, with the clearest gain appearing in tasks that require search, tools, and multiple execution steps. That pattern matches Meta’s training description. The company used rejection-sampled Muse Code trajectories and optimized the model for goal conditioning, context compaction, subagents, and the Muse Code toolset.

#### Meta’s benchmarks measure systems as well as models

In Meta’s Terminal-Bench 2.1 chart, Muse Spark 1.2 with Muse Code scored 82.9%, up from 76.2% for version 1.1 with mini-swe-agent. It trailed Claude Opus 5 with Claude Code at 86.7%. On GDPVal-AA v2, Muse Spark 1.2 recorded an Elo score of 1,631, above the 1,371 result for version 1.1 and the 1,577 shown for GPT-5.6 Terra in Meta’s chart.

![Meta's official Terminal-Bench 2.1 chart shows Muse Spark 1.2 with Muse Code scoring 82.9 percent.](https://s4.tenten.co/learning/content/images/2026/08/landing-page-2-5.png)

Those results are useful, but they are not clean model-only comparisons. Meta’s methodology pairs each model with a different coding agent for Terminal-Bench: Muse uses Muse Code, Claude uses Claude Code, and GPT uses Codex. GDPVal-AA and MCP Atlas use provider harnesses instead. The measurements therefore describe deployable systems, not isolated base-model contributions.

That distinction reflects how coding agents now work. A runtime that preserves an event log, resumes after a crash, and prevents subagents from repeating research can change whether a long task finishes. Muse Code appends model calls, tool runs, approvals, and edits to a local log. Meta says this makes sessions replayable and restart-safe.

![Meta's official GDPVal-AA v2 chart shows an Elo score of 1,631 for Muse Spark 1.2.](https://s4.tenten.co/learning/content/images/2026/08/landing-page-3-5.png)

#### The lower-priced endpoint carries a data-policy decision

Meta offers two Muse Spark 1.2 endpoints. The standard endpoint does not use customer data to improve Meta products and charges USD 1.25 for input, USD 0.15 for cached input, and USD 4.25 for output per million tokens. The Contributor endpoint may use data to improve products and charges USD 0.10, USD 0.002, and USD 0.20 respectively. Both have a one-million-token context window.

Contributor output costs about 4.7% of the standard price. That gap should be treated as a data-governance trade rather than a routine discount. An enterprise must classify source code, customer records, and confidential material before routing work to an endpoint whose data can contribute to product improvement.

Muse Code remains in beta. It runs on macOS and Linux, can keep multiple background agents alive, and is designed for work that spans a session. Meta’s kernel-optimization case study involved more than 1,000 tool calls and runs lasting up to 24 hours. Development teams should evaluate completion rate, human correction time, recovery behavior, and cost per accepted task alongside token prices.

#### Meta has regained entry; retention is the next test

The release pace and measured capability put Meta back in the competitive band. Its strategy is more specific than rebuilding a universal chat assistant: it is entering through a coding agent whose runtime and model are developed together.

The next evidence must come from product performance. Muse Code has to complete real repository work reliably. Meta Model API has to establish capacity, support, and enterprise trust. Rapid model updates also need to avoid invalidating customer evaluations every few weeks. Benchmarks show that Meta has regained entry to the race; task economics and developer retention will determine the value of that position.

#### What is Muse Spark 1.2?

Muse Spark 1.2 is Meta’s model for coding and agent workflows, with a one-million-token context window. It is available through Muse Code, Meta Model API, and partner platforms.

#### How much better is Muse Spark 1.2 than version 1.1?

Artificial Analysis data on August 7 showed an Intelligence Index gain of about 3.6 points. The Agentic Index rose by about 9.6 points, while the Coding Index increased by about 0.9 points.

#### What is the difference between Muse Code and Muse Spark 1.2?

Muse Spark 1.2 is the model. Muse Code is the terminal coding agent that manages tools, background agents, event logs, context, and task recovery around that model.

#### Should enterprises send confidential code to the Contributor endpoint?

Price alone is insufficient. The Contributor endpoint may use data to improve Meta products, so enterprises should apply contractual review, data classification, and internal policy before routing confidential work to it.

#### Sources

- [Meta Research: Introducing Muse Code and Muse Spark 1.2](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2)
- [Meta Developer: Muse Spark 1.2 model and pricing](https://developer.meta.com/ai/models/muse-spark/)
- [Meta Research: Muse Spark 1.2 Evaluation Methodology](https://research.meta.ai/static/muse-spark-1-2-methodology)
- [Artificial Analysis: Muse Spark 1.2 evaluation](https://artificialanalysis.ai/models/muse-spark-1-2)

#### Author Insight

The recurring mistake in coding-agent evaluations is using a base-model score to predict an entire workflow. Context loss, failed-tool recovery, and duplicated subagent research often create the expensive failures. Muse Spark 1.2 matters because Meta is bringing those runtime problems into the model-training loop. That is a more interesting enterprise proposition than another chat interface.