Terminal coding
Near frontierTerminal-Bench 2.1 (Meta harness)
Meta-published Muse Code harness comparison (Aug 5, 2026). Muse Spark 1.2 is 2nd behind Opus 5. Independent verified TB 2.1 entry for 1.2 was not public at launch.
SPIRITT
Muse Spark 1.2Muse Spark 1.2 is built for coding agents
Released August 5, 2026 with Muse Code, Muse Spark 1.2 is Meta Superintelligence Labs' coding-focused update to Muse Spark 1.1. It is co-trained for repository-scale software engineering: plan changes, write code, validate results, and run multi-step agent work. On SPIRITT, you pick it in the model picker and run it inside a fully equipped agent workspace.
Launch week on X
Coding harnesses, near-frontier benches, contributor pricing, and independent AA coverage. Real posts from the Muse Spark 1.2 and Muse Code launch.
Official launch card
From Meta Superintelligence Labs: Muse Spark 1.2 ships with Muse Code, a terminal coding agent for macOS and Linux built around persistent background agents, parallel sub-agents in isolated git worktrees, and built-in verification.
What Meta built it for
A coding-focused model update, not a general multimodal refresh. Meta scaled training compute on coding tasks and expanded environment diversity so Muse Spark 1.2 can plan repository changes, write code, debug, and validate end-to-end developer workflows.
Co-trained with Muse Code: persistent background agents, parallel sub-agents, worktree isolation, and crash-safe execution. Available through the Meta Model API with a 1M-token context, plus a cheaper Contributor tier if you opt into data use.
Coding agentsMuse CodeRepository-scale SETerminal + tools1M contextMeta Model API
AA Intelligence Index
Source: Artificial Analysis (Aug 5, 2026). Muse Spark 1.2 scores 54, up from 51 on 1.1 and 43 on 1.0. Near GPT-5.5 and Grok 4.5, behind current frontier leaders.
Where Muse Spark 1.2 is strong vs other models
This release is about coding agents. Head-to-head charts below mix independent AA scores with Meta's published Muse Code harness comparisons. Vendor harness numbers are labeled as such.
Deep agentic coding
DeepSWE 1.1 (Meta harness)
Meta-published comparison. Muse Spark 1.2 improves on 1.1 (53.0 → 59.3) but trails Opus 5 and GPT-5.6 Terra on this board.
Internal coding
Near frontierMeta Internal Coding Bench
Meta internal agentic coding bench from the 1.2 launch materials. Muse is 2nd behind Opus 5 and ahead of GPT-5.6 Terra.
Agentic knowledge work
WinnerAA Intelligence Index (generation jump)
Independent AA. Within the Muse Spark line, 1.2 leads 1.1 and 1.0. Absolute frontier leaders still score higher overall (see ranking chart).
Epistemics
WinnerAA-Omniscience Index
Independent AA. Score rose 18 → 22 vs 1.1, driven by lower hallucination (38% → 28%) and more abstention when unsure.
Efficiency tradeoff
Cost per AA Intelligence task
Lower is better. AA places Muse Spark 1.2 near the intelligence/cost frontier at about $0.40 per task. That is higher than Muse Spark 1.1 (~$0.26): more agentic quality, higher cost per task.
API price
Standard API input price ($ / M tokens)
Meta Model API list prices at launch: standard Muse Spark 1.2 is $1.25 / $4.25 per M input/output. Contributor tier is much cheaper ($0.10 / $0.20) if you opt into data use. Compare to typical flagship list prices around launch.
Context
Context window (tokens)
1M-token context on Muse Spark 1.2 via Meta Model API, matching the 1.1 long-context design for large repos and long agent traces.
Sources: Meta research post Introducing Muse Code and Muse Spark 1.2 (Aug 5, 2026) and evaluation methodology PDF; Artificial Analysis Muse Spark 1.2 article and Intelligence Index; Meta developer model page pricing. Terminal-Bench / DeepSWE / Internal Coding numbers above are Meta-published harness comparisons unless labeled AA. Independent verified Terminal-Bench 2.1 for Muse Spark 1.2 was not public at launch. On SPIRITT you run the model inside a full agent workspace, not only a terminal CLI.
How It Works
From zero to Muse Spark 1.2 running real work in a cloud agent environment
01
Open a workspace
Open a workspace and land in a fully equipped cloud computer: browser, files, terminal, integrations, and memory. No local setup. No thin chat box pretending to be an agent.
02
Pick Muse Spark 1.2
Open the model picker and choose Muse Spark 1.2. Same model shipped for agentic work, now inside a workspace that already has tools, browser control, and durable context.
03
Build or automate
Tell it what to ship or what to run. Muse Spark can code, call tools, drive the browser, coordinate multi-step work, and keep going while you step away. The point is not another chat window. It is an agentic environment where Muse Spark actually does the job.
Try Muse Spark 1.2 in an agentic environment
A real cloud environment, model picker, tools, and memory. Run coding-agent work end to end, not just another chat window.
Questions
Frequently asked questions
01What is Muse Spark 1.2?+
Muse Spark 1.2 is Meta Superintelligence Labs' August 2026 coding-focused update to Muse Spark 1.1. It powers Muse Code, Meta's terminal coding agent, and is available through the Meta Model API with a 1M-token context. On SPIRITT you can select it in the model picker and run real agent work in a cloud workspace.
02How is Muse Spark 1.2 different from Muse Spark 1.1?+
1.2 is a coding-specialized jump, not a broad multimodal refresh. Meta scaled coding training compute, shipped Muse Code as a co-trained harness, and improved agentic coding boards (Terminal-Bench, DeepSWE, internal coding). Independent AA Intelligence rose from 51 to 54. Cost per AA task also rose (~$0.26 → ~$0.40), so you pay more per intelligence task for stronger agentic coding.
03What is Muse Spark 1.2 best for?+
Repository-scale software engineering and coding agents: planning multi-file changes, writing and debugging code, terminal workflows, and multi-step validation loops. It is strongest when paired with a real harness (Muse Code, SPIRITT workspace tools, browser, files), not single-turn chat.
04What are some of the main limitations of this model?+
It still trails Opus-class leaders on Meta's own Terminal-Bench and DeepSWE charts. Some launch numbers are vendor-harness results, not independently verified leaderboard runs. Standard API pricing is higher than the Contributor tier, and Contributor requires opting into data use. Treat benches as signals and test on your own repos.
05What is a good example of using Muse Spark 1.2 well?+
Open a SPIRITT workspace, pick Muse Spark 1.2, and hand it a multi-file engineering goal: migrate a module, fix a failing suite, or ship a feature across a large repo with tools, terminal, and browser available. That is the co-trained coding-agent loop Meta optimized for.
06Where is Muse Spark 1.2 best to test for full agentic capabilities?+
SPIRITT Workspaces. Pick Muse Spark 1.2 in the model picker and run real work in a fully equipped cloud environment: tools, browser, files, terminal, memory, and durable sessions. That is where the model can act as an agent, not just chat.
Buy from builders who use what they sellBuilt using