Meta Ships Muse Spark 1.2, Its First Model to Beat Opus 5 on an Internal Coding Benchmark

Meta shipped Muse Spark 1.2 on August 5, along with Muse Code, a terminal coding agent built to run on top of it. The two were co-trained together, which Meta says is why the pairing produces fewer retries and cleaner tool calls than bolting an agent onto an outside model.

The Muse Spark 1.2 benchmark numbers Meta published put the model ahead of GPT-5.6 Terra and Grok 4.5 on Terminal-Bench 2.1, at 82.9%. On DeepSWE 1.1, it scored 59.3%, up from 53.3% for Muse Spark 1.1 released in July. On Meta’s own internal coding benchmark, it placed second only to Claude Opus 5.

One catch worth flagging: Meta posted these results as image charts on the Muse Spark product page, without a published methodology. There’s no third-party verification yet on how the comparison models were run or at what settings.

Muse Spark 1.1’s coding gap is closing fast

Muse Spark 1.1 launched in July as an agent-first model. It topped tool-use benchmarks like MCP Atlas and JobBench but trailed on raw coding accuracy, landing behind both Claude Opus 4.8 and GPT-5.5 on SWE-Bench Pro and DeepSWE.

Muse Spark 1.2 closes that gap. The DeepSWE jump from 53.3% to 59.3% in under a month is the sharpest single-version coding gain Meta has posted for this model line. Whether that holds up against independently run evals is still an open question.

Muse Code enters a crowded field

Muse Code installs with a single command and takes on full engineering tasks: planning changes, writing code, and validating the result. Meta says parallel worker agents handle implementation while separate reviewer agents check the work in the background, and that every action is traceable.

That puts Muse Code up against Claude Code, Codex, and Cursor’s coding agents. Meta’s pitch is price. The standard Muse Spark 1.2 tier costs $1.25 per million input tokens and $4.25 per million output tokens. A contributor tier drops to $0.10 in and $0.20 out, in exchange for letting Meta use those prompts to improve its models.

The model carries a 1 million-token context window with built-in compression, meant to keep long coding sessions running in a single pass rather than resetting context partway through.

What’s next

Meta says larger models are already coming. That tracks with the pace of the last month: Anthropic shipped Claude Opus 5, OpenAI pushed GPT-5.6, and Moonshot AI’s Kimi K3 upended pricing expectations for open-weight models. Muse Spark 1.2 is Meta’s answer to all three at once — cheap, fast to deploy, and now closer on raw coding accuracy than its predecessor ever was.

Frequently Asked Questions

What is Muse Spark 1.2?

Muse Spark 1.2 is Meta’s latest coding-focused AI model, co-trained with its new Muse Code terminal agent for tighter tool integration and fewer failed retries.

How does Muse Spark 1.2 compare to Claude Opus 5?

On Meta’s internal coding benchmark, Muse Spark 1.2 placed second only to Claude Opus 5, according to Meta’s own vendor-reported figures.

How much does Muse Spark 1.2 cost to use?

The standard tier costs $1.25 per million input tokens and $4.25 per million output tokens. A contributor tier drops those prices to $0.10 and $0.20 in exchange for Meta using the prompt data to improve its models.

Is Muse Code available now?

Yes, Muse Code shipped in beta on August 5 alongside Muse Spark 1.2, accessible through the Meta Model API’s public preview.