Meta released Muse Spark 1.3 on September 2, 2026, just one day after Google’s Gemini 3.8 Flash launch and two days after Anthropic’s Claude Fable 5.1 and Mythos 5.1 debuts. The timing is no accident. This is a three-way race for the best coding and agentic AI model, and Meta just made it dramatically cheaper to run at the top level.
Muse Spark 1.3 reaches roughly Claude Opus 5 level performance on key benchmarks and even beats Opus 5 on long-horizon coding tests. On the contributor API tier, it costs $0.10 per million input tokens and $0.20 per million output tokens, which undercuts Qwen, DeepSeek, and MiMo class alternatives by a wide margin. The tradeoff is speed of thought: reasoning cannot be fully disabled, and deep reasoning modes can take a long time before responding.
Three Flagship Launches in Three Days
The first week of September 2026 has been extraordinary for AI releases. Anthropic launched Claude Fable 5.1 and Mythos 5.1, its most advanced models for coding and scientific research, on August 31. Google followed with Gemini 3.8 Flash on September 1. Meta answered on September 2 with Muse Spark 1.3.
Each vendor is targeting the same workload: autonomous agents that can code, use computers, browse, and complete multi-step professional tasks. Meta’s pitch is different from the others. Instead of only chasing the top benchmark score, Meta is chasing top benchmark scores per dollar. As we covered in our Muse Code versus Claude Code breakdown, Meta’s strategy is an extremely cheap alternative to Claude Code and rival agent platforms.
Muse Spark 1.3 is rolling out now in Muse Code and the Meta Model API, according to the official Meta announcement. Previously available reasoning modes shipped on day one, while the new max reasoning mode arrives shortly after additional safety testing is complete.
Benchmarks: Around Opus 5 Level, Ahead on Coding
Meta published a four-model scorecard comparing Muse Spark 1.3, Muse Spark 1.2, GPT-5.6 Sol, and Claude Opus 5 across agent, coding, instruction following, and long-context tasks. The honest summary is that Opus 5 still leads on most agentic benchmarks, but Spark 1.3 closes the gap sharply and wins the coding categories that matter most to developers.
On DeepSWE v1.1, a long-horizon agentic coding benchmark, Muse Spark 1.3 scores 75.4, ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0. On Terminal-Bench 2.1, it ties GPT-5.6 Sol at 88.8 and edges past Opus 5 at 86.7. On SWEAtlas CodeBase QnA, it scores 59.4 against 52.7 for Opus 5 and 53.5 for Sol.
On the six agent benchmarks, Opus 5 in max mode still leads on four (GDPVal-AA v2, JobBench, OSWorld 2.0 and AutomationBench), while GPT-5.6 Sol takes DeepSearchQA and the Agentic IF Index. Spark 1.3 is close behind rather than far behind, which is the big change from version 1.2. JobBench moves from 61.6 to 64.9 against Opus 5 at 65.7, and OSWorld 2.0 computer use jumps from 47.6 to 66.9 against Opus 5 at 68.3.
The most dominant results are in long context retrieval. On MRCR 256K to 512K, Spark 1.3 hits 98.5 against 91.5 for Sol, and on MRCR 512K to 1M, it holds 98.1 against 73.8 for Sol. That makes it unusually strong for codebase-scale work inside its 1M token context window, which accepts text, images, video, and PDF input.
Independent tracking points the same way. Artificial Analysis gives Muse Spark 1.3 an Intelligence Index of 62 in max mode and 61 in xhigh, roughly on par with Claude Fable 5 and ahead of GPT-5.6 Sol, with an Agentic Index of 59 in max mode, on par with Claude Opus 5. Meta also says the model uses around 20% fewer tool calls and 25% fewer tokens than version 1.2 in engineering comparisons, so benchmark gains come with lower real task cost.
- Coding is the headline win: first or tied first on all three published coding benchmarks.
- Agent tasks are close losses: within 1 to 2 points of Opus 5 on JobBench, OSWorld, and AutomationBench.
- Long context is dominant: near-perfect retrieval out to 1M tokens, far ahead of GPT-5.6 Sol.
- Efficiency improved: about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2.
| Benchmark | Spark 1.3 (max) | Spark 1.2 (xhigh) | GPT-5.6 Sol (max) | Opus 5 (max) |
|---|---|---|---|---|
| DeepSWE v1.1 (agentic coding) | 75.4 | 55.0 | 73.0 | 74.0 |
| Terminal-Bench 2.1 | 88.8 | 82.9 | 88.8 | 86.7 |
| SWEAtlas CodeBase QnA | 59.4 | 46.2 | 53.5 | 52.7 |
| GDPVal-AA v2 (knowledge work) | 1754 | 1615 | 1710 | 1824 |
| JobBench (tool use) | 64.9 | 61.6 | 45.4 | 65.7 |
| OSWorld 2.0 (computer use) | 66.9 | 47.6 | 62.7 | 68.3 |
| MRCR 512K to 1M (long context) | 98.1 | 55.5 | 73.8 | Not reported |
Pricing: The Contributor Tier Changes Everything
Muse Spark 1.3 keeps the same standard Meta Model API price as version 1.2: $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens. That standard tier does not use prompts or completions to train Meta models.
The story is the contributor tier, published as muse-spark-1.3-contributor: $0.10 per million input tokens, $0.002 per million cached input tokens, and $0.20 per million output tokens. That is about 12.5 times cheaper on input and about 21 times cheaper on output than the standard tier, in exchange for letting Meta use the data to improve its products.
At those contributor rates, routine agent work costs cents rather than dollars, which is why it undercuts budget rivals including Qwen, DeepSeek, and MiMo alternatives for comparable capability. Web search grounding is billed separately at $2.50 per 1,000 queries, with no long-context multiplier, so the 1M context window does not carry a price penalty.
The Catch: Reasoning Takes Time and Cannot Be Switched Off
The main downside is latency from reasoning. Muse Spark supports adjustable reasoning levels, including minimal, low, medium, high, and xhigh, with max arriving after safety testing. Minimal is the lowest setting. Reasoning cannot be disabled entirely, so even simple queries go through some reasoning passes.
In practice that means deep modes can take a long time thinking before they respond. Independent task testing found wall times over 100 seconds per accepted result on high and xhigh profiles for hard multi-step tasks, compared with much faster responses from lighter models such as Gemini 3.8 Flash. For long-horizon coding and research, that tradeoff is worth it, because accuracy rises sharply with reasoning depth. For quick chat or simple completions, it can feel slow.
Meta’s guidance is straightforward: use minimal or low reasoning for fast iteration, xhigh for hard agentic work, and max when it becomes available for the most challenging coding and agent tasks. The model is also trained to ask clarifying questions when prompts are ambiguous, confirm before consequential actions, and track what it has learned across messy single-threaded workflows.
How to Get It: Muse Code, API, and the Subscription Question
Muse Spark 1.3 is available through two routes: Muse Code, Meta’s coding agent CLI for macOS and Linux, and the Meta Model API for direct integration. Install runs from a single shell command published in the Meta launch post.
- Muse Code CLI: free to install, usage billed per token through the Model API at standard or contributor rates.
- Meta Model API: muse-spark-1.3 for standard terms, muse-spark-1.3-contributor for discounted training-data terms.
- Reasoning modes: existing modes available now, max reasoning coming after safety testing.
On subscriptions, the documented picture is narrower than rumors suggest. Muse Code has no flat monthly plan, no seat price, and no documented free quota: it is purely token billed, and you attach a payment method before running it. We could not verify a generally available Meta One subscription for Muse Spark access or a free-use tier inside the Meta app from official documentation at the time of writing, so treat those as unconfirmed until Meta publishes terms. An early sign-up credit was reported during the Muse Spark 1.1 era, but nothing on the live pricing pages confirms an ongoing credit.
Do Not Use the Contributor Tier for Sensitive Data
This is the most important warning in this article. Anyone handling sensitive data, including client code, customer records, medical or legal material, or proprietary repositories, should not use the contributor endpoint. The discount exists because Meta may use submitted prompts and outputs to improve its models.
For that work, use the standard tier, where Meta states data is not used for training. Even then the price stays competitive with other models at this capability level: $1.25 input and $4.25 output per million tokens is in line with flagship agent pricing, while the contributor tier is the option that breaks the price curve for personal projects, experiments, and non-sensitive workloads.
Frequently Asked Questions
Is Muse Spark 1.3 better than Claude Opus 5?
It depends on the task. Spark 1.3 beats or ties Opus 5 on DeepSWE v1.1, Terminal-Bench 2.1, and SWEAtlas CodeBase QnA, plus long-context retrieval. Opus 5 still leads on most general agent benchmarks such as GDPVal, JobBench, and OSWorld. Overall it is fair to call Spark 1.3 roughly Opus 5 level, with a coding-first skew.
How much does Muse Spark 1.3 cost?
Standard API pricing is $1.25 per million input tokens, $0.15 for cached input, and $4.25 for output. Contributor pricing is $0.10 input, $0.002 cached, and $0.20 output. Web search grounding costs an extra $2.50 per 1,000 queries.
Is Muse Spark 1.3 open weight?
No. It is proprietary and available only through Muse Code and the Meta Model API. Meta has signaled a Muse Spark open weights release on its roadmap, but it has not shipped yet.
Can I use Muse Spark 1.3 for free?
There is no documented free tier or recurring quota for Muse Code or the Model API. The contributor tier is so cheap that hobby use costs only cents, but it still requires a payment method, and it grants Meta training rights.
What are the reasoning levels?
Adjustable levels include minimal, low, medium, high, and xhigh, with max coming after safety testing. Minimal is the lowest option, and reasoning cannot be turned off completely, so expect longer waits on harder modes.
Bottom Line
Muse Spark 1.3 does not win every benchmark, but it wins the value argument decisively. Opus-level coding, near-Opus agent performance, a 1M context window, and contributor pricing at $0.10 and $0.20 per million tokens make it the cheapest way to run frontier-class coding agents on non-sensitive work. Just keep client data on the standard tier, pick lower reasoning for quick tasks, and save xhigh and max for the jobs that actually need deep thought.
