Meta has released three versions of Muse Spark in four months: the original in April, 1.1 on July 9, and 1.2 on August 5. Each targets a different strength. Understanding the differences matters because they share pricing tiers but serve different workloads.
The short version: 1.1 was the computer-use and multimodal reasoning upgrade. 1.2 is the coding-focused upgrade. They are not sequential replacements. They are specializations of the same base architecture.
Timeline
| Version | Release | Focus | Key improvement |
|---|---|---|---|
| Muse Spark | April 2026 | General reasoning | Original multimodal model |
| Muse Spark 1.1 | July 9, 2026 | Computer use, agentic tasks | Multi-app workflows, extended sessions |
| Muse Spark 1.2 | August 5, 2026 | Coding | Code generation, debugging, codebase understanding |
All three come from Meta Superintelligence Labs. Each version builds on the previous architecture but with specialized training emphasis.
What 1.1 added over the original
Muse Spark 1.1 launched July 9, 2026 with a focus on computer use and agentic tasks:
- Computer-use workflows: operates across multiple applications with changing on-screen information
- Extended session coherence: maintains context over long multi-step tasks without losing track
- Unfamiliar interface navigation: adapts to new applications without explicit instruction
- Tool use improvements: better at chaining multiple tools in sequence
- Multimodal understanding: stronger vision-language integration
The 1.1 update was Meta’s answer to Claude’s computer-use capabilities and Gemini’s multimodal agent features. It targets the “autonomous desktop worker” use case: filling forms, navigating between apps, extracting information from screens.
What 1.2 adds over 1.1
Muse Spark 1.2 launched August 5, 2026 with a specific coding focus:
- Code generation: scaled up training compute on coding tasks specifically
- Complex debugging: better at finding subtle bugs across multi-file projects
- Codebase understanding: processes large repositories and maintains architectural context
- End-to-end developer workflows: from spec to implementation to testing
- Training environment diversity: exposed to more programming languages, frameworks, and project structures
Meta explicitly describes 1.2 as “a coding-focused update to Muse Spark 1.1” rather than a full-generation upgrade. The base architecture is the same. The difference is where the additional training compute was directed.
Benchmarks
| Benchmark | Muse Spark 1.2 | Claude Opus 5 | Notes |
|---|---|---|---|
| Terminal-Bench 2.1 | 82.9% | 86.7% | Narrowly behind Opus 5 |
| DeepSWE 1.1 | 59.3% (3rd) | Higher | Without internet access |
Important context from kingy.ai’s verified analysis: the 82.9% Terminal-Bench score is the headline number Meta promotes. On DeepSWE 1.1 (113 coding tasks without internet access during grading), Muse Spark 1.2 drops to third place at 59.3%.
Yahoo Finance reported that Muse Code “loses to Anthropic on Meta’s own benchmark charts” for DeepSWE. This suggests 1.2 is strong on terminal-based agentic coding (its specialty) but less dominant on isolated coding without tool access.
No published benchmarks compare 1.1 directly to 1.2 on the same tasks. Meta has not released a side-by-side showing exactly how much coding improved from 1.1 to 1.2.
Pricing (unchanged between versions)
Both 1.1 and 1.2 use the same pricing tiers through the Meta Model API:
| Tier | Input/1M | Output/1M | Data sharing |
|---|---|---|---|
| Contributor | $0.10 | $0.20 | Meta trains on your prompts |
| Standard | $2.00 | $4.25 | Your data is not used for training |
The contributor tier is approximately 21x cheaper on output. The tradeoff: your prompts and completions become Meta training data. For sensitive/proprietary codebases, the standard tier is the only appropriate choice.
Source: MacRumors pricing breakdown
Which version to use when
Use Muse Spark 1.1 for:
- Desktop automation (filling forms, navigating apps)
- Multi-application workflows
- Tasks requiring screen understanding (screenshots, UI parsing)
- Long-running agentic sessions that span multiple tools
- Multimodal reasoning tasks (images + text + actions)
Use Muse Spark 1.2 for:
- Writing code
- Debugging (especially multi-file, cross-module issues)
- Large codebase navigation and understanding
- End-to-end development tasks (spec to tests)
- Terminal-based coding workflows (via Muse Code)
The practical distinction: if your task involves a keyboard and a code editor, use 1.2. If your task involves navigating a GUI, clicking buttons, or working across non-code applications, 1.1 may still be stronger.
Muse Code: the 1.2 showcase
Muse Code is Meta’s new terminal coding agent, launched alongside 1.2 specifically to demonstrate its coding capabilities. It runs on macOS and Linux, operates in your terminal, and can:
- Plan changes across a codebase
- Write and modify code in multiple files
- Run tests and validate results
- Make 1000+ tool calls in extended sessions (Interesting Engineering reported a 24-hour GPU optimization test)
- Understand repository structure and dependencies
Muse Code is powered exclusively by 1.2. There is no option to run it on 1.1. For the full guide, see our Muse Code Complete Guide.
What this means for Ray-Ban glasses users
Muse Spark 1.2 also powers the AI on Meta’s Ray-Ban smart glasses. The glasses received the update starting July 27, 2026, bringing smarter scene understanding, Threads integration, and neural text input.
For glasses users, the coding improvements are irrelevant. What matters is the general intelligence improvement that comes with more training compute: better answers to questions, better object recognition, better contextual suggestions.
For the full glasses impact analysis, see our coverage on Visible Compute.
The trajectory
Three models in four months signals Meta’s pace. Each release is not a replacement but a specialization:
- April: establish the base architecture
- July (1.1): dominate computer-use and agentic desktop tasks
- August (1.2): compete on coding (the most commercially valuable AI application)
This suggests more specialized versions are coming. A reasoning-focused variant, a research/science variant, or a creative variant would fit the pattern. Meta is building a family of task-specific models that share an architecture but diverge in training.
FAQ
Is Muse Spark 1.2 better than 1.1 at everything?
No. 1.2 is specifically better at coding. For computer-use, multi-app workflows, and general multimodal tasks, 1.1’s specialized training may still perform better. Meta has not published a direct comparison.
Do I need to switch from 1.1 to 1.2?
Only if your primary use case is coding. If you use Muse Spark for computer-use automation, screen understanding, or multi-app tasks, stay on 1.1 until Meta publishes comparison data.
Is the pricing the same?
Yes. Both versions use the same contributor ($0.10/$0.20) and standard ($2.00/$4.25) tiers.
How does 1.2 compare to Claude Opus 5?
82.9% vs 86.7% on Terminal-Bench 2.1. Opus 5 wins, but the gap is small. On DeepSWE (without internet), the gap is larger. On pricing, Muse Spark 1.2 contributor tier is dramatically cheaper ($0.20 vs $25 output), but with data-sharing.
Can I use both 1.1 and 1.2 through the same API?
Check the Meta Model API documentation for model selection. The API supports specifying which version to use per request, allowing you to route different task types to different specializations.
Is 1.2 the version on my Ray-Ban glasses?
Yes. The Ray-Ban glasses update (rolled out July 27 onward) runs Muse Spark updates. The glasses benefit from the overall intelligence improvements, even though the coding-specific capabilities are not relevant on glasses.
Sources
- Meta Research: Introducing Muse Code and Muse Spark 1.2
- Meta AI: Introducing Muse Spark 1.1
- kingy.ai: Meta’s 82.9% vs Verified Scores
- Yahoo Finance: Muse Code loses to Anthropic on DeepSWE
- MacRumors: Muse Code pricing tiers
- Interesting Engineering: 1000+ tool calls in 24-hour test
- eesel.ai: Muse Spark 1.1 analysis