What Is Muse Spark 1.3? 1M Context, 98.5% MRCR

Meet Muse Spark 1.3, a 1M-token coding agent

Muse Spark 1.3 is Meta's hosted multimodal reasoning model for coding and agentic workflows, released on September 2, 2026 with a 1,048,576-token context window and API prices from $0.10 per million input tokens on Contributor to $1.25 per million on Standard. It is available through Muse Code and the Meta Model API, where Meta positions it as the strongest Spark release so far. The model update focuses on long-horizon task continuity, coding quality, tool-use efficiency, and better instruction handling.

Key Takeaways

• Muse Spark 1.3 launched on September 2, 2026, through Muse Code and the Meta Model API.
• The model has a 1,048,576-token context window, or approximately 1M tokens.
• Meta reports approximately 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2.
• Standard API pricing is $1.25 per million input tokens and $4.25 per million output tokens.
• The Contributor tier costs $0.10 per million input tokens and $0.20 per million output tokens, but Meta uses Contributor prompts and completions to improve its products.
• Early community testing reports a 75.4% DeepSWE score and a 98.5 score on MRCR 256K–512K, although the supplied evidence does not include full evaluation methodology.

What Is Muse Spark 1.3?

Muse Spark 1.3 is Meta’s latest hosted model in the Muse Spark family. It is a reasoning-oriented model built for software development, autonomous agents, tool calling, and other workflows that require multiple steps rather than one text response.

Meta’s launch messaging describes it as a major improvement in coding and agentic work. Alexandr Wang called it Meta’s “most capable model yet,” while Mark Zuckerberg described it as offering frontier performance at very low cost. The official Muse Spark model page identifies the model within Meta’s developer ecosystem.

The release is confirmed, not a leak or preview-only announcement. Meta announced availability in Muse Code and the Meta Model API on September 2, 2026. The API documentation lists a public preview with expanded global access, although account-level and regional availability can still vary.

Muse Spark 1.3 is not currently an open-weights model. Meta has said that a Muse Spark open-weights release is coming soon, but no release date, model variant, or license has been confirmed.

Muse Spark 1.3 at a Glance

Specification	Muse Spark 1.3
Developer	Meta
Type	Hosted multimodal reasoning and agent model
Modality	Text, image understanding, and video understanding
Context window	1,048,576 tokens
Standard pricing	$1.25 per million input tokens; $0.15 per million cached input tokens; $4.25 per million output tokens
Contributor pricing	$0.10 per million input tokens; $0.002 per million cached input tokens; $0.20 per million output tokens
Availability	Muse Code and Meta Model API
Standard model identifier	muse-spark-1.3
Contributor model identifier	muse-spark-1.3-contributor
License	Not yet confirmed; weights are not currently published

The pricing and identifier details come from Meta’s Model API documentation. The Contributor tier is marked as used to improve Meta’s products, while the Standard tier is marked as not used for that purpose.

How Muse Spark 1.3 Works and What Makes It Different

Muse Spark 1.3 is designed to operate inside an agent loop. Instead of producing only an answer, it can maintain a task state, call tools, inspect results, revise its plan, and continue across multiple turns. Meta’s developer materials describe support for parallel tool calls, streamed tool-call arguments, reasoning that carries across turns, structured output, search grounding, image understanding, and video understanding.

Five practical labels capture the model’s reported behavior:

• Long-Horizon Continuity describes its ability to preserve requirements during extended coding or agent sessions.
• Tool-Call Efficiency describes the reported reduction in external actions needed to reach a result.
• Requirement Retention refers to holding onto constraints and project details across many turns.
• Calibrated Agency refers to asking clarifying questions, identifying when it is stuck, and recognizing limits.
• Consequential-Action Confirmation refers to requesting confirmation before taking actions with meaningful external effects.

These are descriptive anchors, not necessarily official Meta feature names. The underlying behaviors come from launch material attributed to Meta’s AI team and executives.

The clearest measured improvement over Muse Spark 1.2 is efficiency: Meta reports approximately 20% fewer tool calls and 25% fewer tokens on comparable long-horizon tasks. That does not mean every prompt will use 25% fewer tokens. It means Meta’s internal comparisons showed lower average usage across the evaluated workflows.

Muse Spark 1.3’s 1M-token context window is also central to its design. A long context can hold source files, test output, tool results, issue descriptions, and prior decisions without immediately compressing the entire session. Context size alone does not guarantee reliable retrieval, so long-context benchmark results need to be read separately from the nominal window.

What You Can Do With Muse Spark 1.3

The primary use case is software engineering. In Muse Code, the model can plan changes, edit files, run commands, and work through a repository-level task. Meta also positions the API for custom agents, multi-agent orchestration, web-grounded workflows, computer use, and GitHub automation.

Typical tasks include:

• Debugging a codebase while preserving project requirements across several iterations.
• Generating tests, reviewing pull requests, and explaining unfamiliar repositories.
• Building an autonomous coding agent around tool calls and structured output.
• Combining web search grounding with implementation work that depends on current information.
• Operating a browser or desktop interface through computer-use workflows.
• Producing self-contained HTML, SVG, or interactive prototypes from a detailed brief.

A third-party API test gave Muse Spark 1.3 and Muse Spark 1.2 the same one-shot brief: create three collectible 3D figurines in one self-contained HTML file. The test reported 22,474 tokens and $0.10 for 1.3, compared with 19,324 tokens and $0.08 for 1.2. That result is useful as a concrete workflow example, but it does not establish that 1.3 is cheaper on every task. {{EVIDENCE:2095248072154701847}}

For a broader look at the surrounding coding-agent stack, see the Muse Code release analysis covering persistent agents and restart-safe sessions.

How Muse Spark 1.3 Compares

The early benchmark picture is strong but uneven. A community report placed Muse Spark 1.3 first on DeepSWE 1.1 with 75.4%, ahead of Claude Opus 5 at 74.0% and GPT-5.6 Sol at 73.0%. The scores were reported by benchmark commentators rather than presented with full methodology in the supplied first-party posts. {{EVIDENCE:2095232917048000643}}

Reported evaluation	Muse Spark 1.3	Comparison result
DeepSWE 1.1	75.4%	Ahead of Claude Opus 5 at 74.0% and GPT-5.6 Sol at 73.0%, according to community reporting
MRCR 256K–512K	98.5	Higher than GPT-5.6 Sol at 91.5%, according to community reporting
MRCR 512K–1M	98.1	Higher than GPT-5.6 Sol at 73.8%, according to community reporting
JobBench	64.9	Slightly below Claude Opus 5 at 65.7%
OSWorld 2.0	66.9	Below Claude Opus 5 at 68.3%; above Gemini 3.8 Flash at 59.0%
AutomationBench	49.4	Slightly below Claude Opus 5 at 50.3%
GDPVal-AA v2	1,754	Above Gemini 3.8 Flash at 1,545
Terminal-Bench 2.1	88.8	Slightly below Gemini 3.8 Flash at 89.4%

The MRCR and agent-benchmark figures come from early benchmark reporting and a separate comparison with Gemini 3.8 Flash. They should be treated as reported results, not a complete independent model card.

Muse Spark 1.3 is therefore not a universal winner. It appears especially notable on coding and long-context retrieval, while Claude Opus 5 remains slightly ahead on several reported agent and computer-use evaluations. Readers comparing Google’s cost-focused model can also see the site’s Gemini 3.8 Flash versus Claude analysis.

Muse Spark 1.3 is best understood as a frontier-competitive coding and agent model whose strongest evidence currently comes from long-context and software-engineering tests.

Developers evaluating comparable hosted models can also try GPT-5.6 for a separate coding-model reference point.

Availability: How to Access Muse Spark 1.3

Muse Spark 1.3 is available through two official Meta channels:

1. Muse Code: Meta’s coding agent for terminal and continuous-integration workflows. The model is integrated into the product, so developers use Muse Code rather than implementing the complete agent loop themselves.
2. Meta Model API: A direct API for custom applications and agents. The documented base URL is https://api.meta.ai/v1, and the API is compatible with common OpenAI SDK patterns.

The Standard model identifier is muse-spark-1.3. The Contributor identifier is muse-spark-1.3-contributor. Meta’s documentation lists both identifiers and provides pay-as-you-go pricing.

The Contributor tier is substantially cheaper, but its terms differ. Meta labels that tier as used to improve its products, so teams should review data-handling requirements before sending proprietary source code, credentials, customer data, or other sensitive material.

A maximum-reasoning version was described as coming after additional safety testing. Its general availability, exact identifier, and pricing are not yet confirmed in the supplied official documentation.

What We Don't Know Yet

Several important technical details remain unpublished or insufficiently documented:

• Meta has not published parameter counts, architecture details, training-data composition, or a full technical report.
• The complete methodology behind the 75.4% DeepSWE result is not available in the supplied evidence.
• The evaluation conditions for MRCR, JobBench, OSWorld, and AutomationBench have not been independently reproduced here.
• It is not yet clear whether the reported maximum-reasoning configuration is available to all API users.
• Meta has not announced a date for Muse Spark open weights.
• The future open-weights license, hardware requirements, and supported inference stack are not yet confirmed.
• Regional rollout, account limits, rate limits, and differences between Muse Code and API access may vary.
• Community criticism argues that the long-context headline may reflect benchmark specialization. That concern is unconfirmed, but it makes broader real-world testing important.

The current evidence supports a strong release claim, not a final ranking across all engineering workloads. Results from cybersecurity, production repositories, long-running autonomous tasks, and error recovery remain especially valuable.

Frequently Asked Questions

What is Muse Spark 1.3?

Muse Spark 1.3 is Meta's hosted multimodal reasoning model for coding and agentic workflows. It launched on September 2, 2026, with a 1,048,576-token context window and access through Muse Code and the Meta Model API.

Is Muse Spark 1.3 open source?

No, Muse Spark 1.3 is not currently available as an open-weights model. Meta has said that a Muse Spark open-weights release is coming, but it has not published a date or license.

How much does Muse Spark 1.3 cost?

Muse Spark 1.3 costs $1.25 per million input tokens and $4.25 per million output tokens on the Standard tier. The Contributor tier costs $0.10 per million input tokens and $0.20 per million output tokens, and Meta says Contributor prompts and completions are used to improve its products.

How do you access Muse Spark 1.3?

You can access Muse Spark 1.3 through Muse Code or the Meta Model API. The API documentation lists the model identifiers muse-spark-1.3 and muse-spark-1.3-contributor.

What is Muse Spark 1.3's context window?

Muse Spark 1.3 has a 1,048,576-token context window, commonly described as a 1M-token context window. This supports long coding sessions, tool histories, files, and agent instructions in one working context.

Muse Spark 1.3 vs Claude Opus 5: which is better?

Muse Spark 1.3 is competitive with Claude Opus 5, but neither model is better on every reported evaluation. Early community testing puts Muse Spark 1.3 ahead on DeepSWE while Claude Opus 5 leads on several reported agent and computer-use benchmarks.

What to Watch Next

The next useful signals are Meta’s release date and license for Muse Spark open weights, the public rollout of maximum reasoning, and independent evaluations on real repositories and cybersecurity tasks. Updated API terms, regional availability, and broader cost-per-task measurements will show whether the reported efficiency gains hold outside launch-day tests.

Building similar coding and agentic workflows? On kie.ai you can try Claude Opus 5.5, Claude Sonnet 5.5, and GPT-6 Sol and Luna.