Muse Spark 1.3 is Meta’s strongest Muse Spark release so far: the public xhigh model reaches 61 on the Artificial Analysis Intelligence Index, keeps a 1M-token context window, and passed all three medium-difficulty coding tasks in our first-response API test. The important caveat is that Meta’s headline benchmark chart uses max reasoning, while max was still awaiting broader release after additional safety testing when this review was prepared.
Muse Spark 1.3 Review: Quick Verdict
Our verdict
Muse Spark 1.3 is worth testing for coding agents, long-context work, and cost-sensitive automation. Its public xhigh result is already frontier-competitive, but do not quote Meta’s max scores as if every API user can reproduce them today.
Meta released Muse Spark 1.3 on September 2, 2026, positioning it for long-horizon agentic workflows, cleaner coding output, and more reliable collaboration. This continues the shift toward AI agents built around tool-using models. Meta says it asks clarifying questions when requirements are ambiguous, flags when it is stuck, and confirms before consequential actions. In internal comparisons with 1.2, Meta engineers observed about 20% fewer tool calls and 25% fewer tokens. Those are provider claims, not results from our test.
What changed from Muse Spark 1.2?
Artificial Analysis: 57 to 61 on xhigh.
Meta reports 75.4 vs 55.0 on DeepSWE, but compares 1.3 max with 1.2 xhigh.
The 1M window remains; high-band MRCR scores rise sharply in Meta’s chart.
Token rates stay flat, while AA measured $0.55 vs $0.40 per evaluated task.
Muse Spark 1.3 Benchmarks: Max and Xhigh Are Not the Same
Meta’s official chart is impressive. Muse Spark 1.3 max scores 88.8 on Terminal-Bench 2.1, 75.4 on DeepSWE v1.1, 98.1 on MRCR 512K-1M, and 66.9 on OSWorld 2.0. It ties GPT-5.6 Sol max on Terminal-Bench and leads the displayed competitors on the two long-context MRCR bands. Developers comparing the broader market can use our coding-agent overview as additional context.
META-REPORTED, MAX REASONING
| Patokan | 1.3 max | 1.2 xhigh | GPT-5.6 Sol max | Opus 5 max |
|---|---|---|---|---|
| Terminal-Bangku 2.1 | 88.8 | 82.9 | 88.8 | 86.7 |
| DeepSWE v1.1 | 75.4 | 55.0 | 73.0 | 74.0 |
| OSWorld 2.0 | 66.9 | 47.6 | 62.7 | 68.3 |
| MRCR 512K-1M | 98.1 | 55.5 | 73.8 | Tidak tersedia |
Different effort settings, agent harnesses, and OSWorld versions limit direct comparability. Meta also describes Agentic IF Index as an internal composite.
How to read the chart
1.3, GPT-5.6 Sol, and Opus 5 use max; 1.2 uses xhigh.
Named agents and fixed harnesses vary by evaluation.
The 1.3 and 1.2 runs use different environment versions.
Agentic IF Index is Meta’s composite, not a single public benchmark.
Artificial Analysis provides the cleaner view of what users can access now. Muse Spark 1.3 xhigh scores 61, up from 57 for 1.2. Its Terminal-Bench result is 85%, GDPval-AA v2 is 1709 Elo, and Tau3-Bench Banking is 47%. The limited-preview max variant reaches 62 overall. There are also regressions: AA-LCR falls from 83% to 79%, while AA-Omniscience Accuracy moves from 45% to 42% on xhigh.
Muse Spark 1.3 Pricing
Contributor
muse-spark-1.3-contributor
$0.10 input / $0.002 cached / $0.20 output per juta token.
Prompts may be used to improve Meta products.
Standar
muse-spark-1.3
$1.25 input / $0.15 cached / $4.25 output per juta token.
Meta says this tier is not used to improve its products.
The official token prices are unchanged from 1.2, but cost per task is not. Our all-in-one AI platform review explains why access price and workflow cost should be evaluated separately. Artificial Analysis measured $0.55 per Intelligence Index task for 1.3 xhigh versus $0.40 for 1.2, driven mainly by roughly 57% more input tokens on agentic evaluations. The contributor tier is exceptionally cheap, but its data-use term is a real product choice rather than a footnote.
Muse Spark 1.3 API Access
Meta’s official cookbook sekarang menggunakan muse-spark-1.3 as the default and documents a 1,048,576-token context window. The API is compatible with the OpenAI SDK. Muse Code and Meta Model API are confirmed current access routes. Meta’s model page also shows OpenRouter, although its button still pointed to a 1.2 URL when checked, so verify the route before deployment. For background on agent interoperability, see our guide to MCP and AI tools.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.meta.ai/v1",
api_key=os.environ["MODEL_API_KEY"],
)
response = client.chat.completions.create(
model="muse-spark-1.3",
messages=[{"role": "user", "content": "Review this patch for regressions."}],
)
print(response.choices[0].message.content)Hands-on Coding Tests: 3/3 Passed
We tested three medium-difficulty repository tasks through an API at temperature zero, using one response per task. Each response had to return complete files in structured JSON. We then applied those files inside disposable directories and ran public plus hidden checks. These are small controlled tests, not a replacement for Terminal-Bench or a production-repository trial.
5/5 tests; 18.4s; 2,473 tokens; $0.00905. The model found the unused processed-ID set, added an early duplicate guard, and wrote a regression test.
Type-check and two runtime checks passed; 47.6s; 4,945 tokens; $0.01661. Validation, persistence, and audit logging were separated without breaking the transaction boundary.
9/9 tests; 20.3s; 3,789 tokens; $0.01446. CSV compatibility, line-numbered errors, atomic output, dry-run safety, help, and docs all passed.
Total: 11,207 tokens and $0.04013. Times are end-to-end request times, not native Meta API latency or TTFT.
The result was better than a bare “3/3” suggests: all three first responses were reviewable and constraint-aware. The same evidence discipline matters when reading our GPT-5 review. Still, the sample does not measure repeated tool-call reliability, large-repository navigation, or Muse Code’s orchestration layer. Readers evaluating budget alternatives may also find our budget coding model comparison berguna.
Early Developer Reactions
Early reactions are mixed and too sparse to call consensus. In a Hacker News test, Simon Willison reported a 38-second, 4.2266-cent SVG run and judged the 1.3 result “definitely better” than his 1.2 comparison. Dax described the release as a reminder of how much he missed fast models. Sung Kim, after roughly an hour with Muse Code, said the model might be good but the Muse Code product “needs a lot of work.” That last comment is about the agent experience, not a clean model-only test.
Kelebihan dan kekurangan
What worked
- Strong public xhigh score
- Jendela konteks 1M-token
- OpenAI-compatible API
- All three coding tasks passed first response
- Very low contributor-tier token rates
What to watch
- Headline max results are not the default public experience
- Independent task cost rose vs 1.2
- Two independent metrics regressed
- Contributor data may improve Meta products
- Early Muse Code feedback is mixed
Who should use Muse Spark 1.3?
USE IT FOR
Coding-agent evaluations, large-context repository work, tool-driven automation, and cost-sensitive experiments where the data policy is acceptable.
SKIP OR WAIT WHEN
You require proven long-run production reliability, need max today, or cannot accept contributor-tier data use and find standard pricing uncompetitive.
Muse Spark 1.3 Review Verdict
Muse Spark 1.3 earns a recommendation for evaluation. It combines frontier-level public xhigh performance, strong long-context results, familiar API ergonomics, and aggressive pricing. Our three first-response coding tasks reinforce the case that it can follow multi-file constraints and produce testable patches.
Choose the standard tier for sensitive code. For another frontier comparison, see GPT-5 vs Claude 4. Treat max benchmark results as preview evidence until that reasoning level is broadly available. And evaluate Muse Code separately from the model API: an agent’s UX, tool loop, and repository control can succeed or fail independently of the base model.
Muse Spark 1.3 FAQ
What is Muse Spark 1.3?
Muse Spark 1.3 is Meta’s coding- and agentic-workflow model released on September 2, 2026, with a 1M-token context window and native image, video, and document perception.
How much does Muse Spark 1.3 cost?
The standard model costs $1.25 per million input tokens, $0.15 for cached input, and $4.25 for output. The contributor model costs $0.10, $0.002, and $0.20 respectively, but its data may be used to improve Meta products.
Is max reasoning available?
Meta said existing reasoning modes were available at launch and max would follow after additional safety testing. Check current API documentation before assuming general availability.
Did Muse Spark 1.3 pass the coding tests?
Yes. Its first responses passed all executable checks across our Python bug fix, TypeScript refactor, and JSONL feature tests. The sample contains only three controlled tasks and is not a general reliability rate.
Is Muse Spark 1.3 better than 1.2?
Mostly, based on the evidence available. Public xhigh intelligence rose from 57 to 61 and Meta reports large agentic and long-context gains, but Artificial Analysis also found higher task cost and small regressions in long-context retrieval and factual calibration.
Should I use the contributor or standard model?
Use standard for sensitive or proprietary prompts because Meta says that data is not used to improve its products. Contributor is far cheaper, but its prompts may be used for product improvement.
