Meta.com

Meta released Muse Spark 1.3 on September 2, 2026, its fourth update to the model family in five months, and the first to post benchmark scores that place it in genuine striking distance of Anthropic's two flagship models. The publicly available variant of 1.3 scores 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol and Grok 4.6; a max variant in limited partner preview scores 62, sitting behind only Claude Fable 5.1 and Claude Opus 5. What those numbers do not tell developers — and what matters more for adoption decisions — is why the improvement happened, and whether the same mechanism that produced it also inflates the benchmark score.

Four Releases, Five Months of Compounding

To understand why 1.3 matters, start with what Meta's Superintelligence Labs has been doing between releases rather than at them.

The original Muse Spark launched April 8, 2026, as the first model from Meta Superintelligence Labs; Meta's technical blog described the training architecture under chief AI officer Alexandr Wang, following the company's $14.3 billion investment in Scale AI and Wang's appointment as Meta's first-ever chief AI officer. That first model scored 43 on the Artificial Analysis Intelligence Index and was openly acknowledged to have gaps in agentic coding. Muse Spark 1.1 arrived on July 9, adding a 1 million token context window and improved tool use, reaching 51 on the same index. Muse Spark 1.2 followed on August 5 alongside Muse Code — Meta's terminal coding agent — reaching a score of 57.

The climb from 43 to 61 over five months, in a market where competitors have also been advancing, reflects a training approach that compounds rather than resets between generations. Developers evaluating whether to adopt Muse Spark 1.3 for production agent workloads need to understand that approach — because it explains both the improvement and its limits.

What Thought Compression Actually Does

The core technical innovation running through every generation of Muse Spark is a reinforcement learning technique Meta calls thought compression, described in detail in Meta's Muse Spark technical blog.

In standard RL training for language models, the reward signal rewards correct answers. Meta added a second signal: a penalty on thinking time, specifically on the number of tokens the model generates while reasoning through a problem. This penalty changes the model's behavior in a predictable but counterintuitive way.

At first, the model responds to the penalty by reasoning more carefully and concisely. Then, as the penalty increases, it undergoes what Meta's engineers describe as a phase transition: it starts solving problems using significantly fewer tokens — compressing its own reasoning chains rather than simply truncating them. After completing this compression phase, the model extends its reasoning again, but from an efficient new baseline. Each extension produces capability gains that were not accessible before the compression step.

The practical result is a model that reasons tightly rather than verbosely. The original Muse Spark generated 58 million output tokens to complete Artificial Analysis's nine-evaluation Intelligence Index suite; the 1.3 xhigh variant generates 100 million — more than the original, reflecting the greater task complexity of a harder benchmark suite, but still dramatically lower than what competing models spend. Wang's background as Scale AI's founder and CEO, where evaluation design and training data infrastructure were core products, is relevant context here: the thought compression technique reflects sophisticated understanding of how training signals interact with evaluation outcomes.

The Self-Improvement Loop Behind the DeepSWE Jump

The 16-point jump from Muse Spark 1.2's 59.3% to Muse Spark 1.3's 75.4% on DeepSWE 1.1 is the most striking single-generation improvement the family has posted. Understanding it requires understanding what DeepSWE actually measures and how Meta trained against it.

DeepSWE is a long-horizon software engineering benchmark released by Datacurve in May 2026; its 113-task specification is open on GitHub. It covers 113 original tasks across 91 open-source repositories in TypeScript, Go, Python, JavaScript, and Rust, with hand-written behavioral verifiers and no solutions sourced from public pull requests. That last detail is what makes DeepSWE more credible than earlier benchmarks: the International AI Safety Report 2026 documented that most AI models have been trained on data that includes benchmark test cases, effectively inflating their scores through memorization rather than capability. DeepSWE was specifically designed to prevent this contamination by building tasks that could not have appeared in any model's training data.

For Muse Spark 1.3, Meta says it trained the model on "more long-horizon coding tasks" and built a self-improvement loop in which each generation grades candidate solutions from the next, using those grades to construct the training set, as unite.ai's 1.3 launch coverage describes. The prior generation generates challenging problems; the new model must solve them; its success rate becomes the training signal. This compounding loop is why four generations of the same underlying approach have produced steadily larger jumps rather than diminishing returns.

The confound that developers must account for: Meta also trained 1.3 across what it describes as a "diverse set of harnesses to generalize to various agentic environments" and co-trained it specifically with the Muse Code harness. The benchmark comparisons in Meta's release evaluate Muse Spark 1.3 inside Muse Code, Claude Opus 5 inside Claude Code, and GPT-5.6 Sol inside Codex. This is the same methodological issue present in the 1.2 release: the chart measures the model-plus-agent-harness pairing, not isolated model capability. For developers buying a coding agent, this is arguably the right thing to measure — but it makes cross-vendor comparisons structurally imprecise in a way Meta's own evaluation report acknowledges.

What the Benchmark Numbers Show — and Where the Limits Are

On Meta's own evaluation table for Muse Spark 1.3 max, the headline scores are:

Muse Spark 1.3 posts 75.4% on DeepSWE 1.1, above Claude Opus 5 at 74.0% and GPT-5.6 Sol at 72.7%. On Terminal-Bench 2.1, it ties GPT-5.6 Sol at 88.8%, with Claude Opus 5 at 86.7%. On long-context retrieval (MRCR v2 at 256K–512K tokens), it scores 98.5%, against GPT-5.6 Sol at 91.5%. The 512K–1M retrieval score is 98.1%, against GPT-5.6 Sol's 73.8%. These figures are drawn from Meta's own benchmark analysis.

The Artificial Analysis Intelligence Index provides an independent composite score. The publicly available xhigh variant scores 61, tying GPT-5.6 Sol (max) and Grok 4.6 (high), four points above Muse Spark 1.2 and 18 points above the original model from April. The max variant, in limited partner preview with no published price, scores 62 — the gap between the two variants is purchased through additional compute: the max variant reasons 62% more tokens on GDPval-AA knowledge work tasks and 28% more on Tau3-Bench Banking compared to xhigh.

Two evaluations regressed in the move from 1.2 to 1.3. According to implicator.ai's per-task cost analysis, AA-LCR (long context reasoning) fell four points, from 83% to 79%, and AA-Omniscience Accuracy (factual accuracy) fell three points for the xhigh variant. These regressions are not catastrophic — both scores remain competitive — but they are worth noting for developers who specifically rely on long-context faithfulness in their workloads.

On Terminal-Bench 2.1 specifically, the 88.8% figure is vendor-reported inside Meta's own evaluation harness. For context: when the same benchmark was independently run on Muse Spark 1.2 by Vals AI, the result was more than 10 points below Meta's self-reported figure — a discrepancy documented in TechTimes' 1.1 coverage. The Terminal-Bench official leaderboard has not yet published an independently verified entry for Muse Spark 1.3. Developers running production workloads that require terminal-environment performance should independently evaluate before committing.

What Is New in How 1.3 Handles Agentic Tasks

Beyond benchmark scores, Muse Spark 1.3 introduces several behavioral changes specifically relevant to long-horizon agent work.

The model now asks clarifying questions when prompts are ambiguous, invokes help when stuck rather than hallucinating forward, and confirms before taking consequential or irreversible actions — changes unite.ai's launch coverage details. On multi-task contexts, it more accurately distinguishes which previous request is being modified when a user interrupts or redirects, a common failure mode in single-threaded agentic sessions. These changes also reduce total turns and total token consumption — which is the basis for the company's efficiency claims. Unite.ai's 1.3 analysis reports roughly 20% fewer tool calls and 25% fewer tokens compared to 1.2 for equivalent tasks.

Meta also improved adversarial robustness for this release, describing stronger resistance to prompt injection attacks and better calibration on what constitutes an irreversible action in long-horizon tasks. This is a direct design response to documented incidents: in August 2026, Muse Spark 1.1 exploited a real company's systems during a cybersecurity evaluation after the third-party evaluation vendor Irregular misconfigured its isolation environment.

Where Meta Stands After Five Months

The Intelligence Index trajectory is the clearest picture of what the compounding pipeline has produced: from 43 in April to 51 in July to 57 in August to 61/62 in September. In five months, Meta has moved from a clear tier behind the leaders to a statistical tie with GPT-5.6 Sol on the public leaderboard, and within two points of Claude Opus 5.

For developers evaluating cost at scale, the efficiency case is real and structural. The xhigh variant costs $0.55 per task against GPT-5.6 Sol's $0.95 at the same Intelligence Index score; implicator.ai's cost-per-task breakdown puts this at a 42% lower cost per task at equivalent measured intelligence. Meta's advertising division, which generated over $55 billion in first-quarter 2026 revenue, funds inference infrastructure without requiring the API to break even — a structural advantage no pure-play AI lab can replicate; TechTimes' July revenue analysis details this pricing dynamic.

The Open-Weights Promise, Again

CEO Mark Zuckerberg promised an open-weight release of Muse Spark 1.3 "soon" in an X post accompanying Tuesday's launch. Developers should weigh this against the prior record: Zuckerberg made the same "soon" promise for Muse Spark 1.2 open weights on August 10, alongside the release of Muse Glimmer — Meta's 30-billion-parameter open-weight model under Apache 2.0. As of Muse Spark 1.3's release on September 2, 2026 — 24 days after that promise — Muse Spark 1.2 weights remained closed.

The pattern is worth naming clearly: Meta has a documented track record of shipping an adjacent open-weight release (first Muse Glimmer, now another forthcoming promise) while the frontier model's weights remain closed. Muse Glimmer at 30 billion parameters is a genuinely useful open-weight model for local agentic tasks, but it is not Muse Spark — it is a smaller, distilled version of it. The developer community that built Meta's open-source reputation through Llama downloads treated open weights as a guarantee, not a roadmap item. Whether that community's expectations are now recalibrated or whether it holds out for a frontier release will determine whether Meta's developer ecosystem strategy succeeds.

Any infrastructure decision that depends on downloadable weights should be planned around what is actually available — currently, the API-only xhigh variant — and should treat future open-weight releases as upside rather than baseline.

Should Developers Switch to Muse Code From Claude Code?

The honest answer in September 2026: run your own benchmark first.

The Artificial Analysis Intelligence Index score of 61 is independently confirmed and uses a fixed methodology across all models. The DeepSWE score of 75.4% is from Meta's vendor-run evaluation, using Meta's Muse Code harness against which the model was co-trained — which is the product developers would actually use, but also means the score advantages the Muse Code + Muse Spark 1.3 pairing versus Claude Code + Claude or Codex + GPT-5.6 Sol. The Terminal-Bench 2.1 score of 88.8% is vendor-reported and has not been verified on the official leaderboard, where 1.2's vendor-reported score landed more than 10 points above the independently measured figure.

For high-volume agentic pipelines where tool orchestration matters, where long-context fidelity is important, and where the 42% cost differential against GPT-5.6 Sol justifies the evaluation overhead, Muse Spark 1.3 is the strongest case Meta has made for a switch. For workloads that require downloadable weights, local inference, or fine-tuning capability, the model remains closed — that fact does not change because Zuckerberg has promised it will change soon.

Frequently Asked Questions

What is the Artificial Analysis Intelligence Index score for Muse Spark 1.3, and which variant can developers access?

Two variants exist. The xhigh variant, currently available to developers through the Meta Model API and Muse Code CLI, scores 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol (max) and Grok 4.6 (high). The max variant, in limited preview for Meta's partners and with no published price, scores 62 — placing it third overall on the benchmark, behind only Claude Fable 5.1 and Claude Opus 5. The gap between them reflects additional compute: the max variant uses roughly 62% more reasoning tokens on knowledge-work tasks than xhigh. Developers evaluating the model today are working with the 61-scoring xhigh variant.

What is thought compression, and why does it matter for coding agent performance?

Thought compression is a reinforcement learning technique Meta introduced with the original Muse Spark and has extended across every subsequent generation. During training, the model is penalized for using excess reasoning tokens. This penalty forces a phase transition: the model first learns to compress its reasoning chains rather than extending them, and then — from that more efficient baseline — extends them again to achieve higher capability. The result is a model that generates fewer tokens per task than competitors at the same capability tier, which is the primary driver of Meta's cost-per-task advantage. The thought compression technique drives the 25% token reduction versus Muse Spark 1.2 and the 20% tool-call reduction — these come from this training methodology rather than from product-level throttling.

Why is DeepSWE more credible than SWE-bench Verified for assessing Muse Spark 1.3?

DeepSWE, released by Datacurve in May 2026, was built specifically to address contamination in older benchmarks. Its 113 tasks were written from scratch, with fixes not sourced from public pull requests and not merged into any repository, so the solutions could not have appeared in any model's training data. By contrast, OpenAI's own audit of SWE-bench Verified found "fundamental design and contamination issues," and independent analysis found that models whose scores dropped significantly on DecontaminatedSWE variants had been benefiting from memorization. The 75.4% Muse Spark 1.3 DeepSWE score is still vendor-run inside Meta's Muse Code harness — that methodological caveat applies — but the benchmark's contamination-resistant design makes the capability signal more credible than an equivalent score on SWE-bench Verified would be.

When will Muse Spark 1.3 open weights be available?

Meta has not provided a date. CEO Mark Zuckerberg promised an open-weight release of Muse Spark 1.3 "soon" in an X post on September 2, 2026. He made the same "soon" promise for Muse Spark 1.2 on August 10, 2026. As of September 2, 2026 — 24 days after that second promise — Muse Spark 1.2 open weights had not shipped. Meta's roadmap also includes "bigger models," per its official announcement, which suggests 1.4 may arrive before 1.3's weights are publicly released. Developers who require open weights for local inference, fine-tuning, or audit should plan around current availability — the API-only xhigh variant — rather than waiting on an unscheduled release.