Meta Muse Spark 1.3 Leads DeepSWE Coding Benchmark While Undercutting Rivals on Price

Meta Superintelligence Labs released Muse Spark 1.3 on September 2, posting the highest score on DeepSWE v1.1 — a contamination-free independent coding benchmark — while maintaining pricing that undercuts Anthropic's flagship model by roughly eight to one on standard input tokens. The release, available immediately in Muse Code and the Meta Model API, marks the fourth Muse Spark model in five months and the clearest evidence yet that Meta's AI rebuild under Chief AI Officer Alexandr Wang is now producing models competitive with the frontier on coding specifically. There is an important caveat: the benchmark table's strongest numbers, including the DeepSWE lead over Anthropic's Claude Opus 5, were generated using a max reasoning mode that was not publicly available at launch, creating a disclosure gap that separates what Meta measured from what developers can actually call.

Four Models in Five Months: What the Iteration Pace Means

The trajectory of Meta Superintelligence Labs' output under Wang is genuinely unusual. When the original Muse Spark launched in April 2026 — internally codenamed Avocado, built over nine months as a ground-up rebuild of Meta's entire AI stack — it scored approximately 43 on the Artificial Analysis Intelligence Index. Muse Spark 1.1 followed in July and pushed that to 51. Muse Spark 1.2 arrived in August at 57. Muse Spark 1.3 now sits at 61, confirmed by Artificial Analysis as an independent third-party evaluation. That is an 18-point climb in five months, at a pace of roughly four points per model generation, while the pricing stayed flat.

For context on what that index means: Artificial Analysis scores models across a composite of reasoning, coding, knowledge, and instruction-following tasks, with 61 placing Muse Spark 1.3 in a tie with GPT-5.6 Sol (max) and Grok 4.6 (high) near the top of the frontier — behind Claude Fable 5.1 (max) at 66 and Claude Opus 5 (max) at 63, but ahead of every Google model the index currently tracks, according to the same benchmarking organization.

Mark Zuckerberg announced the release on X, describing it as "frontier performance almost too cheap to meter" and "the biggest jump we've made so far on coding and agentic work." Wang, posting separately, called it the team's most capable model yet. The model is live in Muse Code, Meta's terminal-based coding agent, and through the Meta Model API, which accepts OpenAI-compatible SDK requests. A broader rollout to Meta AI, Instagram, and Facebook is described as forthcoming with no announced date.

What DeepSWE v1.1 Actually Tests: Why It Matters Here

Most model launches lean on SWE-bench variants for coding credibility. DeepSWE v1.1 is a different instrument, and understanding what makes it distinct is necessary for evaluating Meta's lead.

DeepSWE is an independent benchmark published by Datacurve, a research organization, with 113 software engineering tasks drawn from active open-source repositories across TypeScript, Go, Python, JavaScript, and Rust. Unlike earlier SWE-bench iterations, every task was written from scratch rather than adapted from historical pull requests, making benchmark contamination through pretraining essentially impossible — the model cannot have seen the solution during training because the solution did not exist until the benchmark was built.

The benchmark's prompts are short and behavior-focused rather than prescriptive. Rather than spelling out what functions to implement, a DeepSWE prompt describes what the resulting software should do, requiring the model-as-agent to explore the codebase, determine where to make changes, and implement a solution that passes program-based verifiers. The DeepSWE paper notes that this design separates models across a 69.8-point range — wider than SWE-Bench Pro's 29.7-point spread for the same models — making it easier to distinguish genuine capability differences. V1.1 introduced isolated verification: the agent's proposed changes are extracted as a git patch and evaluated in a separate container, independent of the agent's runtime environment.

On this benchmark, Meta reports Muse Spark 1.3 scored 75.4, against Claude Opus 5's 74.0 and OpenAI GPT-5.6 Sol's 73.0. That 1.4-point margin over Opus 5 is narrow enough that the confidence interval matters — DeepSWE's own methodology paper notes that cluster bootstrap intervals span several points even for well-separated models — but the direction is clear: Meta's model posted the highest reported score on a benchmark run by an independent organization, not by Meta itself. The Artificial Analysis Coding Agent Index, which combines DeepSWE with Terminal-Bench v2.1 and SWE-Atlas CodeBase QnA, lists Muse Spark 1.3 at the top of its coding composite.

Terminal-Bench 2.1 measures broad terminal-environment competence across software engineering, scientific computing, and other command-line domains. On that benchmark Muse Spark 1.3 scored 88.8, matching GPT-5.6 Sol and ahead of Opus 5's 86.7. On SWEAtlas CodeBase QnA, a codebase comprehension evaluation developed by Artificial Analysis, the model posted 59.4 against GPT-5.6 Sol's 53.5 and Opus 5's 52.7.

Read more: Verification, Not Capability, Is the Bottleneck Holding Agentic AI Back

The Max Reasoning Gap: What Benchmarks Show vs. What Developers Can Run

The most substantive disclosure issue in the Muse Spark 1.3 launch is one most coverage has passed over.

Meta's benchmark table, including the DeepSWE scores above, was generated using max reasoning mode — the highest inference compute setting available for the model. Meta's own launch blog stated clearly that this mode is "coming shortly after we finish additional safety testing," which means it was not publicly available at the time of publication. The modes that shipped on September 2 — low, medium, high, and xhigh — produce different outputs at different costs, and the xhigh mode that developers can actually call today is what earns the model its 61 on the Artificial Analysis Intelligence Index.

This is not a minor detail. Reasoning settings affect how much test-time computation a model applies to each response. The gap between xhigh and max is structurally the same kind of gap that separates, say, a model run with chain-of-thought from one run without it: it can be meaningfully larger than a clean generational improvement. Until max reasoning mode ships and independent benchmarking organizations run it under comparable conditions, the most-cited Muse Spark 1.3 numbers describe a capability ceiling, not the current developer experience.

The prior-generation parallel is instructive. Meta's Muse Spark 1.2 launch benchmark table compared its own max mode against competitor models' xhigh modes in some cases — a methodological discrepancy that independent analysis subsequently flagged. The same analytical caution applies here.

Where Opus 5 Still Leads: The Agent Benchmark Picture

Coding benchmarks are one dimension of the model evaluation space. Agentic task evaluation is another, and the picture there is more nuanced than Meta's launch framing suggests.

Meta's own benchmark table — not cherry-picked by critics — includes six agentic evaluations: GDPVal-AA v2 (knowledge-work benchmarks), JobBench (professional tool use), OSWorld 2.0 (computer use and desktop operation), AutomationBench (end-to-end business workflows), DeepSearchQA (agentic web browsing), and the Agentic IF Index (instruction following). On four of those six, Claude Opus 5 (max) leads Muse Spark 1.3. On the remaining two — DeepSearchQA and Agentic IF — GPT-5.6 Sol leads. Muse Spark 1.3 does not top any of the six agentic evaluations in Meta's own table.

On GDPVal-AA v2, which simulates knowledge-work tasks across a broad range of professional domains, Muse Spark 1.3 scored 1,754 against Opus 5's 1,824. On OSWorld 2.0, which tests whether a model can actually operate a desktop environment by clicking, typing, and navigating applications, it posted 66.9 against Opus 5's 72.1. On AutomationBench, which measures end-to-end completion of business workflows, Muse Spark 1.3 scored 49.4 against Opus 5's 50.3. The model is competitive — a meaningful second or third on each — but the claim of leading agent performance is not supported by Meta's own data.

This distinction matters practically. A developer building a coding pipeline — scaffolded tool use, code review, automated debugging — will find Muse Spark 1.3's profile favorable. A developer building a general-purpose agent that needs to browse the web for research, navigate desktop applications, or manage multi-step professional workflows will still find Opus 5's advantage on those dimensions meaningful. The two model families are not competing for the same use case with the same strength; they are competitive in coding-centric workloads and differentiated in broader agent workloads, and the pricing differential amplifies that distinction further.

Where Muse Spark 1.3 does lead, its lead is significant. The MRCR benchmark, which measures accurate retrieval from long contexts at spans from 256K to 1 million tokens, shows scores of 98.5 at the 256K–512K range and 98.1 at 512K–1M. GPT-5.6 Sol scores 91.5 and 73.8 at those same ranges — a gap of more than 24 points at the longest context span. Opus 5 did not publish a comparable MRCR result in Meta's table. That long-context retrieval advantage is functionally important for a use case that is becoming central to agentic coding: working across an entire codebase in a single thread, where the model must accurately recall constraints, interface definitions, and design decisions introduced hundreds of thousands of tokens earlier. Muse Spark 1.3's near-perfect retrieval scores suggest it handles that requirement substantially more reliably than competitors at longer spans, though independent replication of those figures has not yet been published as of this writing.

How the Efficiency Gains Work and Why Token Count Matters More Than Token Price

The model's most practically significant improvement is not a benchmark score but a behavioral change in how it approaches tasks.

Meta's internal engineering comparisons found that Muse Spark 1.3 completes coding workflows using approximately 20% fewer tool calls and 25% fewer output tokens than Muse Spark 1.2. The mechanism is explicit in Meta's blog: the model was trained specifically on long-horizon coding tasks with an emphasis on reducing unnecessary interaction turns and producing cleaner, less verbose output. The result is a model that, when given a multi-step objective, is less likely to ask redundant clarifying questions, less likely to generate interim summaries nobody requested, and less likely to produce boilerplate commentary around code blocks.

This matters for cost in a way that pure per-token pricing obscures. When a model uses 25% fewer tokens to complete the same task, a $1.25 per million input token rate effectively becomes $0.94 in real cost — before any caching or tiering is applied. Artificial Analysis estimates the cost per task at approximately $0.55 for Muse Spark 1.3 on the Intelligence Index, versus $0.95 for GPT-5.6 Sol at a comparable intelligence level — a 42% cost advantage even at the same per-token price tier.

The behavioral training is also visible in how the model handles ambiguity. Meta's release describes Muse Spark 1.3 as trained to ask clarifying questions when prompts are ambiguous, seek user assistance when genuinely stuck, and confirm before taking consequential or irreversible actions. For a long-horizon coding agent where a wrong decision at step three can cascade across the next twenty steps, the value of that calibration extends beyond politeness into actual reliability.

Meta trained the model across a "diverse set of harnesses to generalize to various agentic environments," per its blog — a training approach that distinguishes Muse Spark 1.3 from models optimized primarily for a single scaffold such as Muse Code itself. The practical implication: developers running Muse Spark 1.3 through OpenCode, through OpenRouter-connected tools, or through custom orchestration layers should see broadly consistent behavior, rather than performance that degrades sharply when the model is removed from Meta's own agent harness. Independent practitioners have noted, however, that running the model inside Muse Code specifically tends to produce lower token usage and fewer unnecessary turns than running it through third-party agents — suggesting that while cross-harness generalization has improved, the co-training with Muse Code still produces the most efficient execution on that surface.

The intelligence index gains from 57 to 61 also reflect training improvements in the post-training phase — the stage after large-scale pretraining, where model behavior is shaped through reinforcement learning, supervised fine-tuning, and preference optimization. The Muse Spark family's original foundation — codenamed Avocado and built through a nine-month full rebuild of Meta's AI stack — has not changed significantly in its base architecture between versions. What has changed with each iteration is the intensity and targeting of post-training, with 1.3 specifically emphasizing longer coding tasks, cleaner output style, and improved agentic self-management as the primary optimization objectives.

Read more: Codex Hits 20 Million Users as Claude Code's Lead in AI Coding Narrows

The Two-Tier Pricing Architecture and What It Actually Means

The pricing structure is the most widely discussed aspect of Muse Spark 1.3, and it deserves precision.

The standard tier — model ID muse-spark-1.3 — prices input tokens at $1.25 per million and output tokens at $4.25 per million. Cached input tokens drop to $0.15 per million. Prompts and completions sent through this tier are not used to train future Meta models. This compares favorably with Claude Opus 5 at $10 and $50 per million tokens and GPT-5.6 Sol at broadly comparable Opus-tier rates. Even on a raw per-token basis, Muse Spark 1.3 costs roughly 8x less per input token and approximately 11.8x less per output token than Anthropic's flagship.

The contributor tier — model ID muse-spark-1.3-contributor — prices input at $0.10 per million and output at $0.20 per million, with cached input dropping to $0.002 per million. That makes contributor-tier output approximately 21x cheaper than the standard tier and roughly 250x cheaper than Claude Opus 5 on output. The trade-off is explicit and non-negotiable: prompts and completions submitted through the contributor tier may be used to improve Meta's future models. This is not buried in terms-of-service fine print; it is the stated purpose of the tier.

That trade-off has real consequences for enterprise developers. Code is often the most confidential thing a company handles — proprietary algorithms, unreleased product logic, authentication flows, internal API schemas. Routing that through the contributor endpoint means Meta may train on it. NDA obligations, master services agreements, and data-processing terms with enterprise customers typically prohibit this. Wang told Axios that a "meaningful double-digit" percentage of developers using Muse Code have chosen the contributor tier, which is notable both as an adoption signal and as a governance reality: many of those developers are choosing the tier without necessarily having verified their data-use obligations first.

Safety Testing, the Gated Reasoning Mode, and What the Prior Breach Informs

The decision to gate max reasoning mode behind additional safety testing is not arbitrary, and the background context explains the caution.

Muse Spark 1.1 — the July 2026 model that opened Meta's paid API — was reported to have accessed an external company's systems during a cybersecurity evaluation, through an infrastructure misconfiguration at a third-party evaluation firm called Irregular. The same firm had been involved in a separate breach during Anthropic's evaluations days earlier, and subsequent reporting established that the issue was a shared evaluation-vendor infrastructure problem rather than isolated model misbehavior. Meta said it did not pause development to address the incident and cited comprehensive safety training for 1.3.

The 1.3 safety improvements Meta describes are specifically calibrated for agentic risk: stronger adversarial robustness and resistance to prompt injection attacks, improved calibration around what constitutes an irreversible action, and better awareness of its own capability limits. These are the properties that matter most when a model is given autonomous access to filesystems, APIs, and long-running code execution environments. The gating of max reasoning behind further testing reflects a reasonable judgment that the highest-capability configuration — the one that produces the benchmark-leading scores — merits a longer runway of evaluation before broad deployment.

Meta's blog also notes that Muse Spark 1.3 has "a better sense of what it can and can't do, what it knows and doesn't know, and when it hits hurdles instead of hallucinating outcomes." This is visible in at least one concrete measurement: on the Artificial Analysis AA-Omniscience test, which measures broad factual knowledge, the model's abstention rate increased — it is now more likely to decline answering when uncertain rather than confabulate an answer. That reduces hallucination rates while also reducing absolute knowledge scores, which means the model's raw knowledge breadth has not kept pace with its coding gains.

What the Open Weights Silence Means

Meta's post-Llama model strategy has been a constant balancing act between open ecosystem commitments and frontier monetization.

When Wang announced Muse Spark 1.3, Meta's blog described open weights for the Muse Spark line and a larger model as coming next, with Zuckerberg separately teasing a model codenamed Watermelon on X, with no dates committed for either. Meta has not decided whether to release Muse Spark 1.3 weights publicly, though it confirmed it still plans to release Muse Spark 1.2 weights when ready.

The EU AI Act adds a specific dimension to this. The regulation's open-source exemption under Article 53 applies to genuinely free and open-source general-purpose models — but that exemption is switched off for models classified as carrying systemic risk. A frontier model competing for the top of global intelligence benchmarks is almost certain to meet the systemic-risk threshold, which means if Meta does release Muse Spark 1.3 weights in Europe, it faces full Article 53 compliance obligations regardless of the open-source framing.

If Muse Spark open weights do arrive — whether for 1.2 or 1.3 — the implications for AI deployment costs are substantial. Self-hosted open weights at frontier-adjacent performance eliminate per-token API costs entirely. Organizations with the infrastructure to run multi-hundred-billion-parameter models could replace API dependence with a one-time hardware investment. That scenario would reset the competitive price floor for the entire developer market, well below even the contributor tier.

Watermelon and What the Next Competitive Cycle Looks Like

The Muse Spark releases from April through September 2026 represent a trajectory experiment: can you take a team built around Alexandr Wang's Scale AI background in data and evaluation infrastructure, rebuild Meta's entire AI stack, and reach frontier-level performance within a year? The answer, measured by the coding benchmarks, is yes — with the caveat that agent performance on general-purpose tasks still trails Anthropic's Opus 5 and that the strongest published numbers require a reasoning configuration not yet shipped.

What comes next is less clear. Wang has teased Watermelon as a substantially larger model. If the Muse Spark open weights release covers a model at 1.2 or 1.3 capability levels, it would be the most significant shift in the self-hosted frontier AI market since DeepSeek's open releases in early 2025 — not because the model itself would be exceptional, but because frontier-adjacent coding capability running on privately managed infrastructure would eliminate the per-token cost structure and the contributor-tier data trade-off simultaneously. Every enterprise that currently routes code through a cloud API because there is no capable self-hosted alternative would gain a new option.

The max reasoning mode for 1.3, once safety testing completes, remains the most important near-term disclosure. If it widens the DeepSWE gap over Opus 5 from 1.4 points to something substantially larger, or if it narrows the agent evaluation gap on GDPVal and OSWorld where Opus 5 currently leads by six to seven points, it will signal whether the Muse Spark coding story and the broader agent story are converging. If max reasoning delivers only incremental gains on agent benchmarks while delivering larger gains on coding, it will confirm that the architectural optimization choices embedded in Muse Spark 1.3's training are fundamentally specialized — excellent for developer workflows, competitive but second-tier for general autonomous operation.

For now, the most accurate description of Muse Spark 1.3 is a model that has earned the leading score on the most credible independent long-horizon coding benchmark available, at a per-task cost well below alternatives, with the important asterisk that its headline numbers describe a reasoning configuration that has not yet shipped. That combination — genuine benchmark leadership in coding, near-perfect long-context retrieval, and a cost structure that makes high-volume agent runs economically viable for the first time — is consequential even with the caveats. The AI coding agent market is moving fast enough that a four-week window between Muse Spark generations matters; how Anthropic and OpenAI respond to the DeepSWE result, and whether independent evaluators confirm or revise it, will be the next chapter.