Meta (Meta Superintelligence Labs)
Muse Spark 1.3
Seven of the fourteen benchmarks this site tracks carry an independent Muse Spark 1.3 score, including AA-AnalystAgent at 37.5 pass^5 (max). The clearest gain is LiveBench, +3.6 at a matched reasoning-effort tier, clearing that benchmark's noise band; Humanity's Last Exam's headline +3.2 mixes tiers against Muse Spark 1.2 and shrinks to +2.04 on a same-tier comparison. DeepSWE has no independent score yet to check Meta's self-reported 75.4% against.
Muse Spark 1.3 benchmarks and pricing, every number sourced: 7 tracked Muse Spark 1.3 benchmark scores (7 independently run, 0 still resting on a vendor’s own claim), priced at $1.25 per million input tokens and $4.25 per million output.
- Released
- 2026-09-02
- License
- proprietary
- Context window
- 1M tokens
- Knowledge cutoff
- Not disclosed
- Verified
- sources checked 2026-09-03–2026-10-15
- Succeeded
- Muse Spark 1.2 (2026-08-05)
- Parameters
- Not disclosed
- Architecture
- Not disclosed
Muse Spark 1.3’s verified record
Muse Spark 1.3’s most-compared rival is Kimi K3: 4 ties and 3 not callable across their 7 shared comparisons. Muse Spark 1.3 is priced at $1.25/$4.25 per 1M tokens (in/out) vs Kimi K3’s $3.00/$15.00.
Against the 189 head-to-head comparisons Muse Spark 1.3 shares with other tracked models:
A gap counts for Muse Spark 1.3 only where both sides were run independently and the benchmark still separates models. Losses are listed alongside wins on purpose — Muse Spark 1.3 trails on 25 of them.
HLE · no tools Reasoning
±2 is noiseLiveBench Composite score across 7 domains
±2.7 is noiseTerminal-Bench 4.0 · max Terminal ops
±12.4 is noiseAnalystAgent Spreadsheet & document analysis
±11.2 is noiseAgents' Last Exam Professional work
±3.2 is noiseNo verdict for Muse Spark 1.3 anywhere on Terminal-Bench 2.1 ( every independently confirmed comparison inside the noise band); GPQA Diamond ( saturated).
What changed from Muse Spark 1.2 to Muse Spark 1.3
The 4 benchmarks both models have been scored on, using the same variant each time. A raw Muse Spark 1.3 gain is not a real gain until it clears that benchmark’s own noise band, so each row below carries the verdict and not just the arithmetic.
| Benchmark | Muse Spark 1.2 | Muse Spark 1.3 | Change | Verdict |
|---|---|---|---|---|
| HLE | 45.46 | 48.7 | +3.2 | Setup-dependent |
| Terminal-Bench 2.1 | 80.15 | 84.3 | +4.1 | Setup-dependent |
| GPQA Diamond | 90.4 | 93.5 | +3.1 | Tainted |
| LiveBench | 78 | 81.6 | +3.6 | Real gap |
Muse Spark 1.3 API pricing
$1.25 in / $4.25 out per 1M tokens — official pricing source
What Muse Spark 1.3 costs per job
| Workload | Tokens in / out | Cost |
|---|---|---|
| One long chat turn | 100K / 10K | $0.168 |
| A codebase review | 1,000K / 100K | $1.68 |
| A day of agent work | 10,000K / 1,000K | $16.75 |
Computed from Muse Spark 1.3’s list rates above — cache discounts, its contributor tier (opts prompts/completions into meta's training data), and batch tiers are not applied.
Muse Spark 1.3 is one of 4 Meta (Meta Superintelligence Labs) models tracked on this site, at these official list prices.
| Model | In / 1M | Out / 1M |
|---|---|---|
| Muse Spark Open | $1.25 | $4.25 |
| Muse Wonder | $4.00 | $20.00 |
| Muse Spark 1.3 | $1.25 | $4.25 |
| Muse Spark 1.2(previous version) | $1.25 | $4.25 |
Muse Spark 1.3 benchmark scores
| Benchmark | Score |
|---|---|
| HLE(no tools)[1] Reasoning · ±2 is noise | |
| Terminal-Bench 2.1[2] Terminal ops · ±10.6 is noise | |
| GPQA Diamondsaturated[3] Expert science Q&A — not ranked at any gap size | |
| LiveBench[4] Composite score across 7 domains · ±2.7 is noise | |
| Agents' Last Exam[5] Professional work · ±3.2 is noise | |
| Terminal-Bench 4.0(max)[6] Terminal ops · ±12.4 is noise | |
| AnalystAgent[7] Spreadsheet & document analysis · ±11.2 is noise |
Who ran these numbers: 7 of 7 independent — artificialanalysis.ai (4), livebench.ai (1), snorkel.ai (1), vals.ai (1).
- HLE: AA's own run of Muse Spark 1.3 at max effort, text-only subset. AA's board read 49.1 on 2026-09-03 and 48.7 (underlying 0.4870) on 2026-09-29 — the max-tier figure was revised, without annotation on the board; AA's xhigh-tier config still reads 47.5% and is not tracked here.
- Terminal-Bench 2.1: AA's own run of 'Muse Spark 1.3 (max)'. AA's board read 85.8 on 2026-09-03 and 84.3 (underlying 0.8427) on 2026-09-29 — revised without annotation; AA's xhigh-tier config still reads 85.4%, now above max. Not yet on tbench.ai's official board, which lists only Muse Spark 1.1. Cross-check (2026-10-01): vals.ai's archived Terminus 2 table lists two rows, 'Muse Spark 1.3 Max' at 79.03 and 'Muse Spark 1.3' at 72.28.
- GPQA Diamond: AA's own run of Muse Spark 1.3 at max effort. AA's board read 93.8 on 2026-09-03 and 93.5 (underlying 0.9354) on 2026-09-29 — revised without annotation. Not on vals.ai's GPQA board, same as Muse Spark 1.2. AA's xhigh-tier config reads higher here, 94.14% — the one benchmark where the two tiers reverse.
- LiveBench: Board row 'Muse Spark 1.3 xHigh Effort' on the 2026-06-25 LiveBench release — same release round and same xHigh tier as Muse Spark 1.1 and 1.2's own tracked rows.
- Agents' Last Exam: Overall pass rate 32.2 (Codex, XHigh, the board's best effort for this model; score 55.8; est. cost $428). Per-effort results: XHigh 32.2, Max 29.6.
- Terminal-Bench 4.0: vals.ai board row — mini-swe-agent harness (single bash tool), pass@1 averaged over 3 full passes, raw value 24.747 at effort max. Artificial Analysis same-conditions run (Max): 33.33. Official board (grant-funded, Muse Code, xhigh effort, 5-trial pass@1): 14.55 ± 2.94 CI — 10.2 points below this run. vals.ai lists the max-compute build as its own model row (muse_spark_1_3_max); this site's existing record follows Artificial Analysis's '(Max)' mapping of the same model. The vals-vs-AA spread at nominally matched conditions is the largest in this column, and the official board ran the xhigh tier rather than max.
- AnalystAgent: AA's own run, board row 'Muse Spark 1.3 (Max)': analystAgent 0.375 = 37.5 pass^5. pass@1 is 56.5. Read from the Muse Spark 1.3 (Max) page on 2026-10-15.
Notes on the record
Launched 2026-09-02 (Meta developer page and AI Research blog), on the roughly monthly Muse Spark cadence (base 04-08, 1.1 07-09, 1.2 08-05). Meta describes it as trained on "months of broad adoption of Muse Code and Meta Model API" and does not say whether it is a new pretraining run or a post-training update to 1.2.
Pricing unchanged from 1.2: Standard $1.25 in / $4.25 out per 1M tokens ($0.15 cache hit) with a no-training-use guarantee; Contributor $0.10 / $0.20 ($0.002 cache hit), which opts prompts and completions into Meta's training data — the same ~12x/21x discount-for-privacy trade (developer.meta.com/ai/models/muse-spark/, checked 2026-09-03).
Artificial Analysis publishes two effort tiers for 1.3: max (its canonical page, tracked here) and xhigh (the tier 1.2 was tracked at). Max is higher on HLE (48.7 vs 47.5); xhigh is higher on Terminal-Bench 2.1 (85.4 vs 84.3) and GPQA Diamond (94.1 vs 93.5). Against 1.2, only LiveBench gives a clean tier-matched verdict: both at xHigh, +3.6, outside the band — a real gain. HLE's tracked +3.2 (48.7 vs 45.46) mixes tiers; matched xhigh-to-xhigh it is +2.04, barely over the 2.0 band. Terminal-Bench 2.1's +4.1 is inside its ±10.6 band — a tie. GPQA Diamond's +3.1 is moot; the benchmark is saturated.
Six tracked benchmarks carry an independent 1.3 score: HLE, Terminal-Bench 2.1 and GPQA Diamond (Artificial Analysis), LiveBench (81.6, xHigh), Agents' Last Exam (32.2, Snorkel, 2026-10-01), and Terminal-Bench 4.0 (24.75, vals.ai, 2026-10-08). LiveCodeBench, SWE-bench Verified, Toolathlon-Verified, DeepSWE, ARC-AGI-2 and HMMT had no 1.3 row as of 2026-09-03, nor AA-AnalystAgent as of 2026-09-29; Meta self-reports 75.4% on DeepSWE, unverified.
Weights: 1.2 shipped closed with an August 10 promise from Chief AI Officer Alexandr Wang that open weights were imminent, still unfulfilled. For 1.3, reporting describes the decision as undecided, citing the EU AI Act's Article 53 systemic-risk exemption. Wang's claim that 1.3 is "competitive with Claude Fable 5.1, better than GPT-5.6 Sol at coding" carries no benchmark number.
Compare with
FAQ
Has Muse Spark 1.3 been independently benchmarked?
Partially — seven of the fourteen benchmarks this site tracks carry an independently-run score: GPQA Diamond (93.5%), Humanity's Last Exam (48.7%), and Terminal-Bench 2.1 (84.3%), all from Artificial Analysis at its "max" reasoning-effort tier and each a few tenths to 1.5 points below the figures AA showed the day after launch, plus LiveBench (81.6 overall, xHigh Effort tier) and Snorkel's Agents' Last Exam (32.2 overall pass rate, Codex, added 2026-10-01), plus vals.ai's Terminal-Bench 4.0 (24.75) and AA-AnalystAgent 37.5 pass^5 (max). LiveCodeBench, SWE-bench Verified, DeepSWE, Toolathlon-Verified, ARC-AGI-2 and HMMT Feb 2026 have no Muse Spark 1.3 row on any board this site checked live — several of those have never scored any Muse Spark version at all, not just this one.
How does Muse Spark 1.3 pricing compare to Muse Spark 1.2?
It's identical: $1.25 per million input tokens and $4.25 per million output tokens on the Standard tier, confirmed side by side on Meta's own developer page. Meta also offers a cheaper Contributor tier at $0.10 in / $0.20 out — the same roughly 12x/21x discount as 1.2 — but that tier opts your prompts and completions into Meta's training data, unlike the privacy-preserving Standard tier.
Does Muse Spark 1.3 improve on Muse Spark 1.2's benchmark scores?
Only partly cleanly. LiveBench gains 3.6 points at a matched tier (both xHigh), clearing its noise band — a genuine improvement. Humanity's Last Exam's tracked +3.2 mixes tiers (1.3 at max, 1.2 at xhigh); matched xhigh-to-xhigh the gain is about 2.0, barely clearing that benchmark's band. Terminal-Bench 2.1 improves from 80.15% to 84.3%, but 4.1 points sits inside its wide (±10.6) band, so it counts as a tie. GPQA Diamond moves from 90.4% to 93.5%, but it is saturated here (trust grade D) and every comparison on it is tainted. DeepSWE — where 1.2 showed the clearest self-report/independent gap (Meta claimed 59.3%, the board measured 55%) — has no independent 1.3 score yet to check Meta's new 75.4% claim against.
Is Muse Spark 1.3 open source?
No, and its path there is less certain than 1.2's. Muse Spark 1.2 shipped closed with an August 2026 promise from Meta's Chief AI Officer that open weights were coming soon — still unfulfilled. For 1.3, reporting describes Meta as having explicitly not decided whether to release weights, citing the EU AI Act's Article 53: the exemption it grants genuinely open-source models does not apply to a model classified as carrying systemic risk, a bar frontier-scale models are likely to meet regardless of licence.
Does Muse Spark 1.3 replace Muse Spark 1.2, or sit alongside it?
It replaces it as Meta's current release, though 1.2 (and 1.1) remain listed and available on Meta's own developer page rather than being withdrawn. This site marks Muse Spark 1.2 superseded accordingly — it stays on this site for comparison, not deleted, but is no longer the current recommendation.
What is Muse Spark 1.3's best current-benchmark result?
An 84.3 on Terminal-Bench 2.1 (2026-09-29), with LiveBench at 81.6 behind it and HLE without tools at 48.7. The 93.5 GPQA Diamond figure printed above them belongs to a saturated board and is not ranked on this site.
Further reading
- GPQA Diamond leaderboard 2026 — Muse Spark 1.3 is one of the 23 models it compares.
- Models with 10M token context windows 2026 — Muse Spark 1.3 is one of the 40 models it compares.
Benchmark guides
- Humanity's Last Exam — what Muse Spark 1.3’s reasoning score on it does and doesn’t prove.