Meta Ships Muse Spark 1.2 Coding Model and Muse Code Agent, Undercuts Rivals with $0.20 Output Pricing

Meta has released Muse Spark 1.2, an upgrade focused on coding performance, along with Muse Code, the company's first dedicated coding agent. The release positions Meta to compete primarily on price rather than raw capability, with a discounted tier that trades cheaper tokens for user data.

What changed in Spark 1.2

According to Meta, Spark 1.2 improves on code generation, debugging, and reasoning over large codebases compared to Spark 1.1, which shipped earlier this year. The company put more training compute toward programming tasks and scaled up the number of training environments, focusing on long-running work such as generating entire repositories or conducting independent research.

Meta says the model was trained to plan several steps ahead, work toward a fixed goal, and compress prior context rather than truncate it during extended sessions. Some training data came from Spark 1.1 itself: the predecessor generated programming tasks and instruction templates, then scored candidate solutions against them. Meta claims this self-generated data pipeline helps Spark 1.2 follow complex instructions more accurately than its predecessor.

Benchmark gaps

Meta compared Spark 1.2 against Grok 4.5, Claude Opus 5, GPT-5.6 Terra (not OpenAI's stronger Sol model), and Gemini 3.6 Flash on Terminal-Bench 2.1, DeepSWE v1.1, and a set of 440 internal codebase tasks. Spark 1.2 shows a clear improvement over Spark 1.1 but does not consistently close the gap to top performers.

Meta's own methodology documentation acknowledges the test setup was not tuned for competing models and may not reflect their best possible results. Independent leaderboards support this: Opus 5 scores roughly two percentage points higher elsewhere than in Meta's published charts. Notably, Meta's methodology document confirms Kimi K3 was tested, but the model is absent from published benchmark results. On Terminal-Bench 2.1, K3 trails Opus 5 by a narrow margin and sits well ahead of Spark 1.2. DeepSWE results also can't be directly compared to the public leaderboard, since each model ran inside its own agent framework rather than a standardized harness.

Muse Code: Meta's answer to Claude Code and Codex

Muse Code, Meta's new terminal-based coding agent, installs with a single command and includes planning commands familiar from competitors: "/plan" generates a plan requiring user approval, while "/goal" drives the agent toward a fixed objective. A less common addition, "/grill," stress-tests a plan for weaknesses before execution.

The agent's distinguishing feature is how it manages sub-agents. Rather than spinning up for a single subtask and shutting down, Muse Code's helper agents stay active for an entire session and report findings on their own initiative, which Meta says reduces redundant research. The tool also supports fast-resume after crashes: it logs every model call, approval, and file change to a local protocol file, and claims to resume exactly where it left off rather than re-reading the full session context.

Pricing: discount for data

Standard pricing matches Spark 1.1: $1.25 per million input tokens and $4.25 per million output tokens. A new discount tier drops output pricing to $0.20 per million tokens, but requires users to share their data for model training. For comparison, Western competitors charge between $10 and $30 per million output tokens, while Chinese providers start around $0.18. Kimi K3 costs $3 per million input tokens and $15 per million output tokens, though cache hits reduce input costs to $0.30 per million.

Meta's stock dropped 10% last week. The company generates 98% of its revenue from advertising.

What this means

Meta's benchmark presentation omits its most relevant competitor, Kimi K3, despite testing against it — a gap that undercuts confidence in the comparative claims. The real story is pricing: a $0.20 per million output token tier, contingent on data sharing, undercuts most Western providers by an order of magnitude and matches the cheapest Chinese offerings. For a company earning nearly all its revenue from advertising, that data-for-discount trade may matter more than winning coding benchmarks outright.