Résumez cet article avec :
Muse Spark 1.3 is Meta’s latest coding-focused model, built to compete with frontier models such as Claude Opus 5 and GPT-5.6 Sol. It combines strong coding and agentic performance with a 1M-token context window, making it particularly interesting for complex codebases and long-running development tasks. Its lower token pricing also makes it a compelling option for developers looking to balance model capability and cost. But benchmarks do not tell the whole story: the best model ultimately depends on your workload, context requirements, and the way your coding agent uses the model.
Meta has released Muse Spark 1.3, its latest reasoning model for agentic and coding workflows. The release comes with a notable claim: on Meta's published evaluation, Muse Spark 1.3 (max) outperforms Claude Opus 5 on the coding benchmarks reported for both models, while also achieving near-perfect scores on long-context retrieval.
For developers, the interesting part is not simply that another frontier model has launched. Muse Spark 1.3 combines strong software-engineering performance with a 1M-token context window and relatively low token pricing, making it a model worth testing alongside existing coding models.
At the same time, the benchmark results need some context. Meta's strongest numbers use the max reasoning configuration, and the comparison mixes results from Meta's own evaluation with results reported by model providers or benchmark leaderboards. Outside coding, Claude Opus 5 still leads on several agentic evaluations.
So how does Muse Spark 1.3 actually compare with Claude Opus 5 and GPT-5.6 Sol, and where does it make the most sense in a production AI stack?
Muse Spark 1.3 at a glance
Muse Spark 1.3 is Meta's latest model in the Muse Spark family, designed for long-horizon agentic work, coding, tool use, and multimodal workflows.
Meta says the model is designed to sustain longer tasks by tracking previous results, working through messy or conflicting information, correcting gaps in its own plan, and asking for clarification when necessary. It also improves multitasking and instruction following compared with earlier Muse Spark models.
For developers, the main specifications are:
Muse Spark 1.3 is therefore positioned less as a traditional code-completion model and more as a general-purpose reasoning model that can operate inside long-running agentic workflows.
That distinction matters when looking at its coding benchmarks.
How good is Muse Spark 1.3 at coding?
The strongest results from the release are concentrated in software engineering and code understanding.
Meta evaluates Muse Spark 1.3 across several categories, including professional work, computer use, web research, automation, software engineering, long-context retrieval, and instruction following.
For coding specifically, three benchmarks provide the clearest comparison with Claude Opus 5 and GPT-5.6 Sol:
Muse Spark 1.3 leads on two of the three benchmarks and ties GPT-5.6 Sol on Terminal-Bench.
The numbers are particularly interesting because these benchmarks measure different parts of software engineering rather than simply asking a model to generate a short function.
DeepSWE v1.1: long-horizon software engineering
DeepSWE is the most important coding result in the comparison.
The benchmark contains 113 software-engineering tasks across 91 repositories and five programming languages: TypeScript, Go, Python, JavaScript, and Rust. Tasks are written from scratch and evaluated using functional and regression tests.
Muse Spark 1.3 reaches 75.4%, compared with 74.0% for Claude Opus 5 and 73.0% for GPT-5.6 Sol.
The gap is relatively small, but the benchmark is designed around long-horizon engineering work: the model has to operate within a repository and produce a change that actually passes tests.
That makes DeepSWE more representative of coding-agent workflows than benchmarks focused only on isolated code-generation problems.
There is also an important methodological detail.
Meta ran Muse Spark 1.3 with the mini-swe-agent harness and obtained the comparison results for other models from the official DeepSWE leaderboard. The models are therefore not all necessarily being measured through an identical provider environment.
The result should be read as strong evidence that Muse Spark 1.3 is competitive on long-horizon software engineering, not as proof that it will outperform Claude on every real-world repository.
SWE-Atlas Codebase QnA: understanding an existing codebase
Muse Spark 1.3 also performs strongly on SWE-Atlas Codebase QnA, with a score of 59.4%.
Claude Opus 5 scores 52.7%, while GPT-5.6 Sol scores 53.5%.
SWE-Atlas focuses on codebase comprehension: before an agent can safely modify software, it first needs to understand how the existing system works.
The benchmark contains 124 tasks across 11 production repositories and evaluates repositories written in Go, Python, C, and TypeScript.
This matters because repository-scale development is rarely just “Write a function that does X.”
A coding agent may instead need to determine:
- where a behavior is implemented,
- which components depend on it,
- what assumptions existing code makes,
- which files need to change,
- and whether a proposed change could break another part of the system.
Muse Spark 1.3's result suggests that repository understanding is one of the areas where the model is particularly competitive.
Terminal-Bench 2.1: coding through the terminal
On Terminal-Bench 2.1, Muse Spark 1.3 scores 88.8%, tying GPT-5.6 Sol and exceeding Claude Opus 5's reported 86.7%.
Terminal-Bench evaluates agents completing tasks in terminal environments. The final container state is checked using an executable verifier rather than relying only on subjective evaluation.
For coding agents, this is an important distinction.
A model may produce syntactically correct code while failing to actually complete a task. Terminal-based evaluations test a broader workflow: navigating an environment, modifying files, executing commands, and leaving the environment in a correct final state.
Muse Spark 1.3's 88.8% result therefore places it among the strongest models in this particular category.
Muse Spark 1.3 is not just a coding model
The coding results are impressive, but they are only part of the picture.
Meta's broader evaluation shows that Claude Opus 5 remains ahead on several agentic workloads:
The pattern is useful.
Muse Spark 1.3's strongest competitive advantage is coding and long-context retrieval.
Claude Opus 5 remains stronger on several professional-work and computer-use evaluations, while GPT-5.6 Sol leads on the reported web-research and instruction-following evaluations.
This is a good example of why there is no single benchmark that can determine the best model for every AI application.
The 1M-token context window could be a major advantage
One of the most interesting features of Muse Spark 1.3 is not actually its coding score.
It is what happens when the context gets very large.
Meta evaluates the model on MRCR v2, a long-context retrieval benchmark. Unlike DeepSWE or Terminal-Bench, MRCR does not test coding or tool use. It measures whether a model can retrieve the correct information from a very large context.
Muse Spark 1.3 achieves:
These are extremely strong retrieval results.
And while MRCR itself is not a coding benchmark, the capability has an obvious application to software engineering.
Large repositories can contain thousands of files. Coding agents also accumulate context from:
- source code
- documentation
- issue descriptions
- previous tool calls
- terminal output
- test results
- generated patches
- earlier decisions
A large context window is only useful if the model can retrieve the relevant information reliably. Muse Spark 1.3's MRCR results suggest that this is one of its strongest capabilities.
For teams working with large repositories or long-running coding agents, this may be just as important as its DeepSWE score.
Muse Spark 1.3 vs. Muse Spark 1.2: how much did it actually improve?
There is an important caveat when comparing Muse Spark 1.3 with its predecessor.
Meta's methodology evaluates:
- Muse Spark 1.3 using max reasoning
- Muse Spark 1.2 using xhigh reasoning
So the headline comparison does not represent a pure model-generation comparison. It also includes a difference in reasoning configuration.
That matters when interpreting the large improvements reported on some benchmarks.
At the same time, Meta reports several practical improvements in Muse Spark 1.3. According to the company, the model uses approximately 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 in its internal engineering comparisons.
For an agentic coding workflow, fewer unnecessary tool calls can matter as much as a higher benchmark score.
Every additional model turn can add: latency, token costs, opportunities for an incorrect action, and more context that subsequent calls need to process.
This is why evaluating coding models purely by benchmark accuracy can miss an important part of production performance.
Muse Spark 1.3 pricing
Muse Spark 1.3 keeps the standard pricing of its predecessor:
Artificial Analysis estimates approximately $0.55 per Intelligence Index task for the xhigh configuration.
For comparison, its reported estimates are approximately:
These figures should not be confused with API token prices. The Intelligence Index task cost is an estimated workload-level metric, while the API prices above are the actual token rates.
For a direct token-price comparison, the models discussed in this article have the following standard API rates:
At the token level, Muse Spark 1.3 is significantly less expensive than both Claude Opus 5 and GPT-5.6 Sol. Its input price is 75% lower than Claude Opus 5 and 68.75% lower than GPT-5.6 Sol, while its output price is 83% lower and 78.75% lower, respectively.
The distinction becomes especially important for coding agents because a single user request can trigger many model calls and large amounts of generated context.
In practice, the right comparison is not simply “Which model has the lowest price per million tokens?". It is “How much does it cost to successfully complete my task?”
A model that requires fewer iterations, fewer tool calls, or fewer generated tokens can be cheaper even if its nominal token price is higher.
Which model should you use?
The benchmark results suggest a simple starting point.
Choose Muse Spark 1.3 when coding is the priority
Muse Spark 1.3 is particularly interesting for:
- long-horizon software engineering,
- repository-level coding,
- codebase understanding,
- terminal-based agents,
- and applications requiring very large contexts.
Its DeepSWE, SWE-Atlas, Terminal-Bench, and MRCR results make it a strong candidate for these workloads.
Choose Claude Opus 5 for broader agentic work
Claude Opus 5 remains ahead on several professional-work and computer-use benchmarks.
If your application combines coding with document generation, professional workflows, or GUI automation, it remains an important model to evaluate.
Choose GPT-5.6 Sol when browsing and instruction following matter
GPT-5.6 Sol leads the reported DeepSearchQA and Agentic IF Index comparisons.
It is therefore worth testing when web research or strict instruction following is a major part of your agent workflow.
The broader lesson is that model selection should follow the workload, not the leaderboard.
How to test Muse Spark 1.3 on your own workload
Public benchmarks are useful for narrowing down the candidates, but they cannot tell you which model will perform best on your specific application.
A better approach is to build a small evaluation set from real tasks your developers already encounter.
For coding, include several categories:
- Bug fixing: Can the model identify and fix an existing issue?
- Feature implementation: Can it add functionality without breaking existing behavior?
- Refactoring: Can it modify a codebase while preserving tests?
- Codebase Q&A: Can it locate and explain relevant implementation details?
- Test generation: Can it identify edge cases and produce useful tests?
Then measure more than whether the output “looks good.”
Track:
- Task success rate
- Test pass rate
- Input and output tokens
- Number of tool calls
- Latency
- Cost per successful task
- Human intervention required
For agentic coding, the strongest evaluation is often the simplest one:
Give different models the same issue in the same repository, let them work with the same tools, and measure whether they actually produce a working change.
Compare coding models through Eden AI
This is where a unified AI gateway becomes useful.
When new frontier models are released, testing them should not require rewriting an application's entire model layer.
With Eden AI, developers can access multiple AI providers through a unified API and switch between models without building a separate integration for every provider.
That makes it easier to run controlled comparisons:
import requests
url = "https://api.edenai.run/v3/chat/completions"
headers = {
"Authorization": "Bearer YOUR_API_KEY",
"Content-Type": "application/json"
}
payload = {
"model": "MODEL_NAME",
"messages": [
{
"role": "user",
"content": "Review this repository change and identify potential bugs."
}
]
}
response = requests.post(
url,
json=payload,
headers=headers
)
print(response.json())The important part of the evaluation is keeping everything except the model constant.
Use the same: prompt, repository, tool configuration, test suite, timeout, and success criteria. That lets you compare the models based on actual engineering outcomes rather than simply comparing marketing claims or isolated benchmark scores.
Muse Spark 1.3 vs Claude Opus 5: the practical verdict
Muse Spark 1.3 is a serious new contender in AI coding.
On Meta's reported comparison, it leads Claude Opus 5 on DeepSWE v1.1 and SWE-Atlas Codebase QnA and scores 88.8% on Terminal-Bench 2.1, tying GPT-5.6 Sol. Its near-98% MRCR scores also make its long-context performance particularly noteworthy.
But calling it a universal replacement for Claude Opus 5 would be premature.
Claude remains ahead on several broader agentic benchmarks, while GPT-5.6 Sol leads some web-research and instruction-following evaluations.
The more useful conclusion is this that Muse Spark 1.3 should be on the shortlist for any team evaluating coding models in 2026.
Its combination of coding performance, repository understanding, 1M-token context, and relatively competitive pricing makes it particularly attractive for long-running software-engineering agents.
And because the model landscape changes quickly, the best strategy is not to commit to a single model based on one benchmark.
Test the models against your own workloads, measure quality alongside cost and latency, and keep the ability to switch providers as new models arrive.
.png)