Muse Spark 1.3
Muse Spark 1.3 is Meta's model for agents and coding, released on Sep 2, 2026, and it is the model Meta points to when it says the Muse agent runs on Muse Spark. It costs US$ 1.25 per million input tokens and US$ 4.25 per million output, with a 1,048,576 token window and no long-context surcharge. In Meta's own table, it leads on coding and long context, but it does not win any of the six agent tests on the primary metric. The weights are closed.
Muse Spark 1.3 in official numbers
Data published by the maker in its documentation and announcement. The source of each item is at the end of the page.
- Maker
- Meta (Meta Superintelligence Labs)
- API identifier
muse-spark-1.3; themuse-spark-1.3-contributorversion is cheaper and lets Meta train on your data- Release
- Sep 2, 2026
- Input
- US$ 1.25 per million tokens
- Output (includes reasoning)
- US$ 4.25 per million tokens
- Cache read
- US$ 0.15 per million; automatic prefix caching, no cache write line in the price table
- Contributor tier
- US$ 0.10 input, US$ 0.20 output and US$ 0.002 cache, without max effort
- Long context
- no surcharge: the same price with an empty or a full window
- Context window
- 1,048,576 tokens
- Max output
- not published
- Batch API
- not published; the documentation only offers background requests
- Knowledge
- date not published by Meta
- Reasoning
- always on; minimal, low, medium, high, xhigh or max (max on the Standard tier only); without the parameter, the model decides
- Input and output
- text, image, video, PDF and audio (partially supported) in; text out
- Weights
- closed; Meta lists opening the Muse Spark weights on its roadmap, with no date
- Rate limits
- 3,000 requests and 4 million tokens per minute, per team (Standard tier)
- Where it runs
- Meta Model API (Responses, Chat Completions and Messages, compatible with the OpenAI and Anthropic SDKs) and Muse Code
- Countries
- the API works where Meta AI and billing are available; Brazil is not on the restricted territories list, but Meta does not publish a list of supported countries
Where Muse Spark 1.3 does well, where it falls behind and who it is for
Where it does well
- Long context: 98.5 on MRCR from 256 thousand to 512 thousand tokens and 98.1 from 512 thousand to 1 million, against 91.5 and 73.8 for GPT-5.6 Sol, in Meta's table.
- Long coding work: 75.4% on DeepSWE v1.1, ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0).
- Price: below Grok 4.7 (US$ 2 and US$ 6) and Claude Sonnet 5.5 (US$ 2 and US$ 10), with no extra charge for long context.
- Leaner than 1.2: about 20% fewer tool calls and 25% fewer tokens, in comparisons by Meta engineers.
- Easy to plug in: it accepts the OpenAI and Anthropic SDKs and connects to Claude Code, Codex and OpenCode.
Where it falls behind
- Agents: it does not win any of the six agent tests on Meta's primary metric. On GDPVal-AA v2 it scores 1,754 against 1,824 for Opus 5; on AutomationBench, 49.6 against 50.3.
- Instruction following: 57.8 on Meta's internal index, below GPT-5.6 Sol (60.5) and Opus 5 (59.1), on a test nobody outside can check.
- Previous-generation rivals: the table leaves out Claude Fable 5.1, released one day earlier, and GPT-6 Astra and Opus 5.5, which came later.
- Gaps: max output, knowledge cutoff and a Batch API do not appear in the documentation; audio is only partially supported.
- A prompt you do not see: Meta injects a hidden system prompt and developer message into every request. It does not bill for them, but they take up the window.
Who it makes sense for
- Agents that read a lot of material at once, such as a whole repository or a long contract, at a single price.
- Coding in Muse Code or in the agent the team already uses (Claude Code, Codex or OpenCode), pointed at Meta's API.
- Prototypes with data that can go into training, on the Contributor tier at US$ 0.10 and US$ 0.20 per million.
- Anyone who wants to build something like the Muse agent through the API, on the same model.
When to pick another model
- Office work agents and process automation: in Meta's own table, Opus 5 and GPT-5.6 Sol come out ahead.
- Sensitive data on the Contributor tier, where Meta trains on your prompts and completions. On Standard it does not.
- Audio understanding: Meta itself warns that quality may degrade on 1.3 and points to 1.2 or Muse Voice Transcribe.
- Running on your own servers: the weights are closed. Meta's open option is Muse Glimmer, a smaller model.
What Meta published at launch
Muse Spark 1.3 came out on Sep 2, 2026, with max effort already on the API and in Muse Code. Meta says it sustains longer work, asks for clarification when a request is ambiguous and confirms before consequential actions. On safety, it mentions stronger resistance to prompt injection, with no published figure.
The announcement table, published as an image, compares 1.3 at max with 1.2, GPT-5.6 Sol and Claude Opus 5. 1.3 leads the five coding and long-context rows, with a tie on Terminal-Bench 2.1. On the six agent rows, on the primary metric, Opus 5 leads four and GPT-5.6 Sol leads two.
| Muse Spark 1.3 max | Muse Spark 1.2 xhigh | GPT-5.6 Sol max | Opus 5 max | |
|---|---|---|---|---|
| GDPVal-AA v2 office work, Elo |
1,754 | 1,615 | 1,710 | 1,824 |
| JobBench professional tool use |
64.9 | 61.6 | 45.4 | 65.7 |
| OSWorld 2.0, partial computer use, primary metric |
66.9 | 47.6 | 62.7 | 68.3 |
| OSWorld 2.0, binary full task, secondary metric |
32.0 | 17.9 | 27.3 | 31.4 |
| DeepSearchQA web research |
90.3 | 85.9 | 93.1 | 90.4 |
| Agentic IF Index instruction following, internal test |
57.8 | 46.2 | 60.5 | 59.1 |
| AutomationBench business workflows |
49.6 | 38.2 | 46.7 | 50.3 |
| MRCR 256K-512K long context |
98.5 | 66.3 | 91.5 | not measured |
| MRCR 512K-1M long context |
98.1 | 55.5 | 73.8 | not measured |
| DeepSWE v1.1 long coding work |
75.4 | 55.0 | 73.0 | 74.0 |
| SWE-Atlas Codebase QnA code understanding |
59.4 | 46.2 | 53.5 | 52.7 |
| Terminal-Bench 2.1 terminal coding |
88.8 | 82.9 | 88.8 | 86.7 |
Table from the Muse Spark 1.3 announcement. The dot marks the best in each row; Terminal-Bench 2.1 is a tie between Muse Spark 1.3 and GPT-5.6 Sol. "Not measured" cells are the ones where Meta published no score for Opus 5.
How to read Meta's table
The methodology, in a separate PDF, says Meta uses for each model the highest comparable score among its own run, the official leaderboard and the maker's figure. Where they can be checked, the figures match: Opus 5's 1,824 on GDPVal-AA v2 is the one in Anthropic's system card, and GPT-5.6 Sol's 88.8% on Terminal-Bench 2.1 is OpenAI's. The caveats are elsewhere:
- OSWorld 2.0: the primary metric, by Meta's own definition, is the partial score, where Opus 5 leads; the binary one, where 1.3 leads, is secondary.
- DeepSWE: 1.3 ran with mini-swe-agent, and the others came from Datacurve's official leaderboard, under different conditions.
- Test version: on AutomationBench (v3 task set), Meta gives GPT-5.6 Sol 46.7 and Opus 5 50.3; OpenAI and Anthropic published 18.1% and 26.9% for the same models. Different versions do not compare.
Muse Spark, the Muse agent and Muse Code: what each one is
Muse Spark is the model family from Meta Superintelligence Labs: the first came out on Apr 8, 2026, 1.1 on Jul 9, 2026, with the Meta Model API in public preview, 1.2 on Aug 5, 2026, together with Muse Code, Meta's terminal coding agent, and 1.3 is the fourth since April.
Muse is Meta's personal agent, launched on Sep 8, 2026 in the United States. The announcement says it runs on Muse Spark and links that name to the 1.3 post, without stating the version. Muse Glimmer, from Aug 10, 2026, is the open-weight model: 30 billion parameters, trained on Muse Spark's outputs, under the Apache 2.0 license, to run on your own machine.
Pricing and the fine print
Meta charges the same for 1.3, 1.2 and 1.1 on the Standard tier. Reasoning is billed as output, including at max effort, which spends more tokens. Web search costs US$ 2.50 per thousand queries.
- Standard or Contributor: Contributor charges US$ 0.20 for output instead of US$ 4.25, but Meta may train future models on your data, and the limit drops to 100 requests per minute.
- Zero data retention: only for qualified accounts, on request, and it turns off files, web search and background requests.
- Muse Code subscription: three plans, with 5 and 20 times the basic usage, with no price in the documentation.
| Muse Spark 1.3 Meta | Grok 4.7 xAI | Sonnet 5.5 Anthropic | Opus 5.5 Anthropic | |
|---|---|---|---|---|
| Input | US$ 1.25 | US$ 2 | US$ 2 | US$ 4 |
| Output | US$ 4.25 | US$ 6 | US$ 10 | US$ 20 |
| Cache read | US$ 0.15 | US$ 0.50 | US$ 0.20 | US$ 0.20 |
| Context window | 1,048,576 | 500 thousand | 1 million | 1 million |
| Long context | same price | doubles from 200 thousand | same price | same price |
Meta (Standard tier), xAI and Anthropic price tables, checked on Oct 1, 2026.
Examples of using Muse Spark 1.3
A team puts the code of a large system into the window and asks where each business rule lives before touching anything. That is what SWE-Atlas Codebase QnA measures, and there 1.3 scored 59.4 in Meta's table.
A year of support conversations goes into a single request, with no extra charge for the size, and the agent finds what the customer said in each contact. MRCR measures this retrieval: 98.1 between 512 thousand and 1 million tokens.
A team builds the integration with made-up data on the Contributor tier and moves to Standard when customer data comes in, because there Meta does not train on the content.
Where Muse Spark 1.3 stands out and where it falls behind, by the numbers
Muse Spark 1.3 has published results on 18 tests compiled by Epoch AI. The yardstick here is its generation: its position among the models released between Dec 6, 2025 and Dec 1, 2026 that took the same test. Against the whole archive, which includes 2023 models, a recent release lands near the top of almost everything, and that says nothing.
Where it stands out
Tests where it sits in the top third of its own generation.
Where it falls behind
Tests where it sits in the bottom half of its own generation.
See all 18 tests, one by one
| Test | Score | In its generation | Overall | Test record |
|---|---|---|---|---|
| AIME 2024 and 2025 (OTIS) | 99.2% | #15 / 78 (tied with 1) | #15 / 168 | 100.0% GPT-5.5 |
| Dtbench | 96.5% | #10 / 66 (tied with 1) | #10 / 148 | 98.9% Claude Opus 5.5 |
| Lmca | 53.87 | #14 / 66 | #14 / 121 | 68.23 Claude Opus 5.5 |
| CritPt | 26.0% | #17 / 77 | #17 / 112 | 32.3% GPT-5.6 Sol |
| SciCode | 59.7% | #9 / 75 | #9 / 104 | 66.9% Claude Opus 5.5 |
| Frontiermath Tiers 1 3 V2 | 74.4% | #22 / 67 | #22 / 86 | 93.7% GPT-6 Astra |
| ALE-Bench | 1,688 | #10 / 60 | #10 / 82 | 2,951 GPT-6 Astra |
| WebDev Arena | 1,652 | #11 / 70 | #11 / 82 | 1,820 Claude Opus 5.5 |
| Chess puzzles | 37.0% | #34 / 62 (tied with 1) | #36 / 80 | 76.0% GPT-6 Astra |
| Game puzzles | 25.0% | #34 / 60 (tied with 1) | #35 / 77 | 84.0% GPT-6 Astra |
| ProofBench | 58.0% | #21 / 67 (tied with 2) | #21 / 74 | 100.0% Claude Fable 5.1 |
| Frontiermath Tier 4 V2 | 46.3% | #20 / 57 (tied with 2) | #20 / 67 | 100.0% GPT-6.1 Sol |
| Apex Agents | 57.8% | #15 / 48 | #15 / 50 | 82.4% Gemini 4 Argon |
| Gdp Pdf | 27.6% | #10 / 37 | #10 / 37 | 35.0% GPT-6.1 Sol |
| FrontierSWE | 25.8% | #13 / 20 | #13 / 20 | 65.5% GPT-6 Astra |
| CursorBench | 41.6% | #10 / 16 | #10 / 16 | 57.8% Claude Opus 5.5 |
| Weirdml V3 | 7.0% | #9 / 10 | #9 / 11 | 42.2% GPT-6 Astra |
| Btf3 | 14.1% | #5 / 8 | #5 / 8 | 14.9% GLM-5.3 |
For each test, the model's best effort level counts, as in the ranking. Source of the scores: Epoch AI, under CC BY 4.0.
What to compare Muse Spark 1.3 with
The models from other labs closest in capability and released in the same period, at most two per lab. Click any of them to see its page.
| Model | Lab | ECI | ELO | LiveBench | Release |
|---|---|---|---|---|---|
| Gemini 3.8 Flash | Google DeepMind | 157.20 | 1,551 | 75.8 | Sep 2, 2026 |
| Qwen 3.8 Max | Alibaba | 156.88 | 1,536 | 78.5 | Aug 2, 2026 |
| Muse Spark 1.3 | Meta AI | 156.76 | 1,528 | 81.6 | Sep 2, 2026 |
| GPT-5.3 Codex | OpenAI | 156.76 | n/a | n/a | Feb 5, 2026 |
| GPT-5.4 | OpenAI | 156.72 | 1,519 | 78.0 | Mar 5, 2026 |
| Claude Sonnet 5 | Anthropic | 156.36 | 1,491 | 76.0 | Jun 30, 2026 |
| Claude Opus 4.7 | Anthropic | 156.32 | 1,532 | 76.5 | Apr 16, 2026 |
Muse Spark 1.3 or GPT-5.4?
GPT-5.4 is the model from another lab closest to Muse Spark 1.3 in capability. Across the 13 tests both took, this is the score:
| Test | Muse Spark 1.3 | GPT-5.4 |
|---|---|---|
| Game puzzles | 25.0% | 37.0% |
| Chess puzzles | 37.0% | 48.0% |
| WebDev Arena | 1,652 | 1,463 |
| CritPt | 26.0% | 23.4% |
| Apex Agents | 57.8% | 52.4% |
| Frontiermath Tier 4 V2 | 46.3% | 49.0% |
| Frontiermath Tiers 1 3 V2 | 74.4% | 78.6% |
| SciCode | 59.7% | 56.6% |
| ALE-Bench | 1,688 | 1,607 |
| Lmca | 53.87 | 51.97 |
Unweighted score: each test counts one point, whether math or coding. The dot marks who did better on each row.
What Muse Spark 1.3 costs to deliver each task
Price per token is what is on the price sheet; cost per task is what lands on the invoice, because it depends on how many tokens the model spends. Each point is an effort level, published by whoever ran the test and compiled by Epoch AI. The further up and to the left, the more score per dollar.
CursorBench
The score scale does not start at zero.
ProofBench
The score scale does not start at zero.
Other models from Meta AI
The models from the same lab closest in time, newest first.
| Model | Release | ECI | ELO | LiveBench | Access |
|---|---|---|---|---|---|
| Muse Spark Open | Oct 15, 2026 | no ECI | n/a | n/a | Open weights (restricted use) |
| Muse Wonder | Oct 12, 2026 | no ECI | 1,540 | n/a | API access |
| Muse Spark 1.3 | Sep 2, 2026 | 156.76 | 1,528 | 81.6 | API access |
| Muse Spark 1.2 | Aug 5, 2026 | 154.90 | 1,509 | 78.0 | API access |
| Muse Spark 1.1 | Jul 9, 2026 | 154.00 | 1,516 | 75.3 | API access |
| Muse Spark | Apr 8, 2026 | 152.13 | 1,508 | n/a | API access |
| Llama 4 Maverick | Apr 6, 2025 | 132.20 | n/a | n/a | Open weights (restricted use) |
| Llama 4 Scout | Apr 5, 2025 | 129.65 | n/a | n/a | Open weights (restricted use) |
| Llama 3.3 70B | Dec 6, 2024 | 127.34 | n/a | n/a | Open weights (restricted use) |
Have you used Muse Spark 1.3? Rate it
Users' opinion, not measurement: the rating is kept apart from the ranking and does not move anyone's position. One vote per browser, no sign-up, and you can change your rating later.
What people ask about Muse Spark 1.3
How much does Muse Spark 1.3 cost?
On the Standard tier, US$ 1.25 per million input tokens, US$ 0.15 for cached input and US$ 4.25 for output, with no long-context surcharge. On the Contributor tier, where Meta may train on your data, it costs US$ 0.10, US$ 0.002 and US$ 0.20. Web search costs US$ 2.50 per thousand queries.
Does Muse Spark 1.3 have open weights?
No. Meta only says that opening the Muse Spark weights is on its roadmap, with no date. Its open model today is Muse Glimmer, with 30 billion parameters, distilled from Muse Spark, under the Apache 2.0 license.
How is Muse Spark 1.3 related to the Muse agent?
Muse is Meta's personal agent, launched on Sep 8, 2026 in the United States. The announcement says it runs on Muse Spark and links that name to the 1.3 post. Muse Spark 1.3 is the model; Muse is the product that uses it.
Can I use Muse Spark 1.3 in Brazil?
Meta says the API is available in most countries where Meta AI works and billing is supported. Brazil is not on the restricted territories list in the geographic policy, updated on Sep 18, 2026, but Meta does not publish a list of supported countries. The Muse agent, for now, launched only in the United States.
Is Muse Spark 1.3 better than Claude and GPT?
In Meta's table, it beats Claude Opus 5 and GPT-5.6 Sol on coding and long context and does not lead any of the six agent tests. Newer models such as Claude Fable 5.1, GPT-6 Astra and Opus 5.5 were left out. The independent scores are at the top of this page.
Where does Muse Spark 1.3 rank among AI models?
In the ROO3 AI Benchmark, ordered by the Epoch AI ECI, Muse Spark 1.3 ranks #21 out of 267 models, with an ECI of 156.76. The ECI is a composite index that combines several hard benchmarks into a single scale. The position changes when a new model comes out, and the page is refreshed six times a day.
Is Muse Spark 1.3 better than GPT-5.4?
Across the 13 tests both models took and Epoch AI compiled, Muse Spark 1.3 comes out ahead on 9, behind on 4 and tied on 0. It is a test-by-test count with no weighting: a win in math counts the same as a win in coding. To choose between the two, look at the tests that resemble your own use.
What is Muse Spark 1.3 best at?
Against models of the same generation (released up to 9 months before and up to 3 months after it), the best relative results of Muse Spark 1.3 are: SciCode (#9 of 75); WebDev Arena (#11 of 70); Dtbench (#10 of 66). The yardstick is its position among the models of that period that took the same test.
Where do the numbers on this page come from?
From the same three sources as the ranking, all openly licensed: Epoch AI (ECI and individual tests, CC BY 4.0), Arena (human-preference ELO, CC BY 4.0) and LiveBench (overall average, CC BY-SA 4.0). ROO3 does not measure any model. When the page shows a figure published by the maker itself, it appears in the written analysis, labeled as the maker's, and never enters the ranking.
Where each number comes from
- Meta, Muse Spark 1.3 announcement (with the benchmark table as an image)
- Meta, Muse Spark 1.3 evaluation methodology (PDF)
- Meta, Meta Model API models (documentation)
- Meta, Meta Model API pricing and rate limits (documentation)
- Meta, reasoning (documentation)
- Meta, Meta Model API and Muse Code overview (documentation)
- Meta, token counting and injected context (documentation)
- Meta, Muse Code subscriptions (documentation)
- Meta, zero data retention (help center)
- Meta, countries where the Model API works (help center)
- Meta, Meta Model API geographic use policy
- Meta, Muse agent announcement (Sep 8, 2026)
- Meta, first Muse Spark announcement (Apr 8, 2026)
- Meta, Muse Spark 1.1 and Meta Model API announcement
- Meta, Muse Code and Muse Spark 1.2 announcement
- Meta, Muse Glimmer announcement
- OpenAI, GPT-6 Astra announcement (GPT-5.6 Sol AutomationBench score)
- Anthropic, Claude Opus 5.5 announcement (Opus 5 AutomationBench score)
- xAI, pricing (Grok 4.7)
- Anthropic, pricing (Sonnet 5.5 and Opus 5.5)
Epoch AI, "AI Benchmarking Hub". Published at epoch.ai. Data under the CC BY 4.0 licence, reordered by roo3.co.
The Arena (LMArena) leaderboard, from the public lmarena-ai/leaderboard-dataset on Hugging Face, under the CC BY 4.0 licence. We are not affiliated with the Arena.
LiveBench, by the LiveBench team, published at livebench.ai under the CC BY-SA 4.0 license. The overall average is recomputed by roo3.co with the same formula as the website (average of the categories), and those numbers stay under CC BY-SA 4.0. We are not affiliated with LiveBench.
Read also: Muse, Meta's AI agent: what changes for anyone who sells online
← See the full ranking