TL;DR
- Meta Muse Spark 1.3 (released September 2, 2026) uses approximately 20% fewer tool calls and 25% fewer tokens than version 1.2 on equivalent agentic coding workflows, according to Meta’s own evaluation.
- Priced at $1.25/M input and $4.25/M output tokens, it undercuts Claude Opus 5 ($5/$25) and GPT-5.6 Sol ($5/$30) by 4–7× on token costs.
- The headline benchmark wins come with a methodology problem: Meta compared its
maxreasoning mode against 1.2’sxhighmode — andmaxwas not publicly available at launch. - EU teams: Meta publishes no data processing agreement, no GDPR commitment, and no confirmed availability for the Meta Model API in Europe.
Who should care: Engineering leads evaluating frontier models for high-volume agentic pipelines — especially those tracking cost per task rather than cost per token.
Verdict: Test it — the pricing is genuinely compelling for US-based teams running thousands of agentic tasks per day, but EU/CH teams should hold off until Meta publishes a DPA.
What Changed Between 1.2 and 1.3
Meta released Muse Spark 1.3 on September 2, 2026, as the latest update to its flagship agentic coding model. The headline claim is straightforward: the model completes similar coding tasks with approximately 20% fewer tool calls and 25% fewer tokens than version 1.2. Meta’s own engineers describe it as “significantly faster and more efficient” on software engineering workflows.
The context window stays at 1,048,576 tokens — unchanged from 1.1. What changed is how the model decides when to call a tool. According to Meta’s release post, 1.3 is better at batching operations and avoiding redundant lookups, which is what drives the tool-call reduction. The model is closed-weight and runs through Meta Model API and the Muse Code editor.
Pricing is also unchanged from the previous version. Standard tier: $1.25 per million input tokens and $4.25 per million output tokens. A contributor tier at $0.10/$0.20 is available if you allow Meta to use your prompts and outputs for model training — a trade-off that most enterprise teams will decline on principle.
The Benchmark Numbers — and Who Actually Ran Them
Meta’s launch report included a comparison table pitting Muse Spark 1.3 against GPT-5.6 Sol and Claude Opus 5 across several benchmarks. The scores look strong for Meta on coding and long-context retrieval. But three caveats matter before you weight these numbers.
Caveat 1: Meta ran the evaluations. The comparison figures in Meta’s launch report come from Meta’s own test runs, not from independent labs. The benchmark scores for competitors in Meta’s table differ from what OpenAI and Anthropic report for their own models. Per Flowtivity’s analysis, “treat every benchmark table Meta publishes as a vendor’s own marketing until independent numbers land.”
Caveat 2: Unequal reasoning tiers. The most-cited benchmark improvement — OSWorld 2.0 jumping from 47.6 (1.2) to 66.9 (1.3) — compares 1.3 in max mode against 1.2 in xhigh mode. Eesel AI’s breakdown flags this as “kind of dishonest,” since max mode remained gated behind additional safety testing at launch and was not available to developers.
Caveat 3: Verbosity. Independent analysis found Muse Spark 1.3 generated approximately 120 million output tokens during benchmark evaluation, versus a field median of 72 million. That verbosity partially offsets the claimed token-efficiency gains.
Artificial Analysis, which evaluates models independently, ranked Muse Spark 1.3 sixth out of 636 models on its Intelligence Index, with a speed of 85 tokens per second at the standard endpoint.
With all that said, here is what Meta’s own evaluation table shows — treating it as vendor data, not ground truth:
| Benchmark | Muse Spark 1.3 (max mode) |
Claude Opus 5 (max mode) |
GPT-5.6 Sol (max mode) |
Notes |
|---|---|---|---|---|
| DeepSWE v1.1 (agentic SWE) | 75.4% | 74.0% | 73.0% | Meta internal eval; DataCamp reports Opus 5 at 68.8% on its own run |
| Terminal-Bench 2.1 | 88.8% | 89.1%† | 88.8% | †DataCamp figure; Meta’s table shows 86.7% for Opus 5 |
| Long-context MRCR v2 (256K–512K) | 98.5% | — | 91.5% | Meta internal eval; clearest win for Muse 1.3 |
| Long-context MRCR v2 (512K–1M) | 98.1% | — | 73.8% | Meta internal eval; large gap at 512K+ context |
| GDPval-AA v2 Elo (knowledge work) | 1,754 | 1,824 | 1,710 | Meta internal eval; Opus 5 leads on general agent knowledge |
Source: eesel AI citing Meta’s launch report; independent discrepancies noted from DataCamp.
Cost Per Task: Where the Real Advantage Sits
The pricing gap between Muse Spark 1.3 and its main competitors is large enough to be the primary reason to evaluate the model, even setting aside benchmarks.
For a typical agentic coding task — assume 10,000 input tokens (codebase context, instructions) and 20,000 output tokens (reasoning trace, code edits, tool outputs) — here is what each model costs at published list prices:
| Model | Input $/MTok | Output $/MTok | Task cost¹ | 1,000 tasks/mo |
|---|---|---|---|---|
| Muse Spark 1.3 (standard) | $1.25 | $4.25 | $0.097 | $97 |
| Claude Opus 5 | $5.00 | $25.00 | $0.55 | $550 |
| GPT-5.6 Sol | $5.00 | $30.00 | $0.65 | $650 |
| Claude Fable 5.1 | $10.00 | $50.00 | $1.10 | $1,100 |
¹ Estimated at 10K input + 20K output tokens per task. Prices from OpenRouter (Muse), Anthropic pricing page (Claude models), and DataCamp (GPT-5.6 Sol) as of September 2026. Token volumes are illustrative — actual task costs vary with context and output length.
The math changes if you take the contributor tier ($0.10/$0.20 per MTok). That same task costs $0.005 — but your prompt data is used to train future models. For teams running internal tooling with proprietary code, that is likely a non-starter.
One important counter-weight: Muse Spark 1.3 generated roughly 120 million output tokens per evaluation run versus the field median of 72 million, according to eesel AI. A model that takes longer to reason through a task may partially offset its lower per-token price. The net savings depend heavily on the task type and how verbose the model is in practice.
For Swiss & EU teams
Meta has published no data processing agreement and no GDPR commitments for the Meta Model API. Innfactory’s analysis notes that “EU availability is not confirmed by Meta” and that the API launches with no stated EU data residency. Teams subject to FADP or EU AI Act obligations on personal data should treat Muse Spark 1.3 as off-limits for production use until Meta publishes a DPA. The open-weight Muse Glimmer 30B (Apache 2.0) is the self-hostable alternative for EU workloads, though performance is a significant step down from the flagship model.
Verdict: What to Do With This Model
If you run high-volume agentic coding tasks in the US — CI agents, automated PR review, batch code generation — the cost-per-task gap is large enough to justify a serious evaluation. A team running 5,000 agentic tasks per month could save $2,250–$5,000 compared to Opus 5 or Sol, even if Muse Spark 1.3 turns out to be 10–15% worse on quality. That is a real business case.
The benchmark numbers are best ignored until an independent lab reprints them under consistent conditions. What matters more for adoption is whether the model behaves reliably in your specific agentic scaffold — particularly on long-context retrieval, where the MRCR v2 lead over GPT-5.6 Sol (98.1% vs 73.8% at 512K–1M context) is the single most credible signal in Meta’s own table.
Use it if: you have US-based infrastructure, cost-sensitive agentic workloads, and long-context retrieval is a key requirement.
Test it if: you want to benchmark against your own tasks before committing — and you have the engineering bandwidth to run the comparison rigorously.
Skip it if: you are in the EU or Switzerland, or if you need a data processing agreement before onboarding any new model API.
For a framework on running that comparison, see How to Evaluate a New Frontier Model Before Switching. For context on Meta’s broader Muse model family, our September benchmark of Muse Voice Transcribe covers how Meta is positioning the Muse line across modalities.
Further Reading
- Introducing Muse Spark 1.3 — Meta AI Research — the official release post; no pricing or independent evaluation, but includes the full benchmark table and API setup instructions.
- Meta Muse Spark 1.3: benchmarks, pricing, and what changed — eesel AI — the most thorough third-party breakdown, including the verbosity finding and the xhigh vs max mode methodology problem.
- Muse Spark 1.3 — Artificial Analysis — independent Intelligence Index ranking and speed benchmarks, one of the few non-Meta evaluations available at launch.
Your turn: Has your team run Muse Spark 1.3 in a real agentic pipeline yet — and does the tool-call reduction actually show up in your bills? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.
How this article was made: AI researched and wrote this article under the Adrian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.

