Abstract
Meta released Muse Wonder and Muse Wonder Full on October 12, 2026. Both models carry the Active Memory and Active Focus mechanisms and support a one-million-token context window. On the Artificial Analysis Intelligence Index, Muse Wonder Full scores 56, tied with Claude Sonnet 5.5. On SWE-Marathon, a long-horizon software engineering benchmark, Full reaches a 57.5 percent completion rate, beating Claude Opus 5.5 at 46.3 percent and GPT-6 Astra at 42.5 percent by more than eleven points. This article reads the numbers, separates vendor claims from third-party data, and lays out the pricing, benchmark standing, and developer implications. All evaluation figures come from Artificial Analysis and Vals AI, dated October 12, 2026.
1. The Release and Its Core Mechanisms
Meta announced two flagship models on October 12. Muse Wonder is the standard tier. Muse Wonder Full is the high-capability tier. Both run on a one-million-token context window and accept multimodal input natively: text, high-resolution images, ultra-long video, and continuous audio. The long context does not carry an additional charge.
Two internal mechanisms drive the models.
Active Memory lets the model preserve key user preferences and context automatically across long sessions. Constraints introduced early in a conversation are not forgotten as the session grows.
Active Focus lets the model change strategy when it meets an obstacle, self-check after an error, and adjust its plan in real time when new requirements arrive.
The combination of the two explains the model's edge on long-horizon programming and professional work. The training background reinforces the point. Muse Wonder was built in the gigawatt-scale Prometheus data center, after fourteen months of infrastructure restructuring, and reinforced through a large volume of long-cycle, multi-file real programming tasks. The result is positioned as an executor that can carry a complete software project forward, rather than a model that only answers questions.
2. Standing on the Intelligence Index
The Artificial Analysis Intelligence Index v4.3.2 incorporates ten evaluations: AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPI, AA-Omniscence, and AA-LCR v1.1.
Muse Wonder Full scores 56, tied with Claude Sonnet 5.5. Muse Wonder scores 52. Compared with the previous generation, Muse Spark 1.3 at 48 points six weeks ago, Full gained eight points. Across the ten sub-tests, Full outperformed the previous generation on every item, reached the top three on six, and took first place on two.
The ranked field places the new models competitively. The order around the top is: Claude Sonnet 5.5 with thinking at 58, Claude Sonnet 5.5 with fallback at 56, Muse Wonder (Max) at 56, GPT-6 Astra (Max) at 54, Gemini 4.5 Ultra (Max) at 52, Wonder (Max) at 52, and GPT-6 (Max) at 51.
3. Coding Benchmarks
The coding suite shows where Muse Wonder is strongest. The table lists five evaluations across the Full, Wonder, Claude Fable 5.1, GPT-6 Astra, and Claude Opus 5.5 configurations, all at max reasoning.
| Benchmark | Full (max) | Wonder (max) | Fable 5.1 (max) | Astra (max) | Opus 5.5 (max) |
|---|---|---|---|---|---|
| MemoryBench (remembering context) | 85.5 | 85.6 | 79.0 | 81.4 | 81.5 |
| Terminal-Bench 4.0-AA (agentic terminal coding) | 57.1 | 52.0 | 52.0 | 59.1 | 59.6 |
| DeepSWE 1.1 (long-horizon coding) | 79.0 | 78.8 | 67.4 | 74.1 | 74.2 |
| SWE-Marathon (ultra-long horizon engineering) | 57.5 | 53.1 | 45.6 | 42.5 | 46.3 |
| ProgramBench (from-scratch programming) | 59.0 | 50.0 | 55.5 | 52.5 | 70.5 |
| Artificial Analysis Index | 56 | 52 | 53 | 53 | 58 |
The picture is mixed by task. Full leads on MemoryBench, DeepSWE, and SWE-Marathon. It trails on Terminal-Bench and ProgramBench, where Opus 5.5 still holds a clear margin. This asymmetry is normal for frontier models. No single model leads every evaluation.
4. Long-Horizon Programming: SWE-Marathon and DeepSWE
SWE-Marathon is the benchmark closest to real development work. The model must carry a software project through to completion: read requirements, modify multiple files, run tests, diagnose errors, and revise. The completion rates are:
| Model | SWE-Marathon Completion Rate |
|---|---|
| Muse Wonder Full | 57.5% |
| Muse Wonder | 53.1% |
| Claude Opus 5.5 | 46.3% |
| GPT-6 Astra | 42.5% |
Full leads Opus and Astra by more than eleven points. This is the headline figure of the release.
DeepSWE v1.1 covers repository repair. It includes 113 tasks across 91 code repositories and five programming languages. Full scores 79.0 percent, Wonder 78.8 percent, Opus 5.5 74.2 percent, and Astra 74.1 percent. Both Muse tiers outperform the incumbents on this repair workload.
5. Personal and Professional Agent Benchmarks
The model is evaluated not only on raw coding but on agentic work. The following table uses the same five model columns and covers personal and professional agent dimensions.
| Benchmark | Full (max) | Wonder (max) | Fable 5.1 (max) | Astra (max) | Opus 5.5 (max) |
|---|---|---|---|---|---|
| Personal Superintelligence Index | 59.1 | 54.3 | 49.9 | 54.5 | 55.2 |
| Harvey LAB-AA (legal agent work) | 13.1 | 10.0 | 6.4 | 8.6 | 4.2 |
| Vals Finance Agent V2 (financial research) | 59.1 | 56.6 | 58.9 | 53.5 | 58.6 |
| HealthBench Pro (health queries) | 62.3 | 60.5 | 62.1 | 65.3 | 65.6 |
| AutomationBench-AA (E2E business workflows) | 45.4 | 39.1 | 32.1 | 41.6 | 41.2 |
| GDP.pdf (professional work) | 33.2 | 30.8 | 27.6 | 34.2 | 30.6 |
| JobBench (professional tool use) | 75.3 | 70.5 | 73.1 | 66.1 | 75.4 |
| OSWorld V2.1 (agentic computer use) | 78.4 / 39.4 | 72.6 / 34.3 | 80.7 / 43.1 | 75.8 / 34.3 | 81.8 / – |
The legal result is the most striking. On Harvey LAB-AA, Full scores 13.1 against Opus 5.5's 4.2, a factor of 3.1. On financial research, Full leads at 59.1, above Opus at 58.6 and Astra at 53.5.
On the Vals AI combined board, covering finance, programming, legal, and tax, Wonder Full scores 64.57 and ranks fifth, above every OpenAI model on the board. On OSWorld, the agentic computer-use test, Full trails the leaders on partial success but records a competitive binary result.
6. Coding Agents Index
Coding-agent performance is measured through the Muse Code toolchain. Muse Code paired with Full scores 68, tying the combination of Claude Code with Sonnet 5.5 and exceeding Claude Code with Opus at 66. Muse Code with the standard Wonder scores 63. The previous generation, Muse Code with Spark 1.3, scored 54. The gain is fourteen points in one cycle.
On Terminal-Bench 4.0, the test requires the model to read files, run commands, and handle errors on its own. The previous Spark completed 33 percent of tasks. Full reaches 57 percent, a gain of 24 points.
7. Hallucination Rate and the Knowledge Tradeoff
The model makes an explicit trade between avoiding wrong answers and covering more knowledge. On a 6,000-question knowledge test from Artificial Analysis, Full's hallucination rate is 23 percent. Astra is 51 percent. Opus 5.5 is 59 percent. Full tends to decline to answer when uncertain, rather than fabricate a response.
The cost is knowledge breadth. Full's knowledge index is 30, below Astra at 43 and Opus at 46. The interpretation is scenario-dependent. For high-confidence requirements such as compliance review and legal analysis, the low hallucination rate is an advantage. For broad knowledge question answering, Opus and Astra still hold the edge.
8. Reasoning Efficiency
ProofBench, a mathematics proof test from Vals AI, illustrates the efficiency gain. Full solves 94 problems correctly, close to Opus at 100 and Astra at 99. The difference is in cost. Full reduces output tokens from 4.5 million to 2.7 million, a drop of 40 percent. Average solution steps fall from 120 to 62. The model reaches near-equivalent accuracy with substantially less reasoning.
9. Pricing and Availability
The two models ship through different channels.
Muse Wonder is available on the Meta Model API and in Muse Code. API billing is usage-based:
- Input: four dollars per million tokens
- Output: twenty dollars per million tokens
- Cached input: 0.20 dollars per million tokens
- Context window: one million tokens, no extra charge for long context
Muse Wonder Full is open only inside Muse Code subscriptions. It has no standalone API yet.
- Power Usage: fifty dollars per month, providing 20x the base usage
- Maximum Usage: two hundred dollars per month, providing 80x the base usage
Meta also plans to bring Wonder into the Muse personal assistant, with subscribers prioritized.
10. What This Means for Developers
The release resets the capability floor for frontier models. Long-horizon programming, terminal operation, and professional work are the core scenarios where developers use AI daily. These are precisely the three areas where Muse Wonder's advantage over Astra and Opus is most visible.
The competitive shift matters as much as the numbers. In six weeks, the score moved from 48 to 56. Meta's iteration pace is accelerating. The developer challenge is no longer "which model is better." It is "how to switch to the best model for each task quickly, while controlling API cost." Task-specific routing, budget monitoring, and rapid model switching now dominate the selection problem.
As the frontier expands to four serious players, teams increasingly run multiple models in parallel and route each workload to the best-performing endpoint. Unified access layers reduce the operational cost of this approach. 4sapi provides a single gateway to multiple mainstream models, letting developers use one API key across providers and compare output quality side by side, reducing the trial-and-error spend that comes with blind model switching.
11. FAQ Essentials
What is the core difference between Wonder and Full? Full leads on long-horizon programming (SWE-Marathon 57.5 percent versus 53.1 percent), legal work (a 3.1x gap on Harvey LAB-AA), and math reasoning, while using fewer tokens for equivalent tasks. The standard Wonder already exceeds GPT-6 Astra on repository repair and coding-agent tests and is the more cost-effective choice. Full is currently subscription-only.
What does the one-million-token context mean in practice? It can hold the full files of a medium-sized codebase, or hundreds of pages of technical documentation. Combined with Active Memory, the model tracks requirement changes across a conversation without forgetting earlier constraints. This is a structural advantage on SWE-Marathon-style benchmarks.
When will the Full API open? Meta has not published a timeline. Full is available through Muse Code at fifty or two hundred dollars per month. The standard Wonder API is fully open at four dollars input and twenty dollars output per million tokens.
Does the 23 percent hallucination rate mean weaker knowledge coverage? The rate counts the share of answered questions that are wrong. At 23 percent, Full is well below Astra at 51 percent and Opus at 59 percent, meaning it tends to abstain rather than guess. Its knowledge index of 30 is lower, so breadth is the remaining gap.
Which input modalities are supported? Text, high-resolution images, ultra-long video, and continuous audio, all natively, within the one-million-token window.
12. References
- Meta official blog (Muse Wonder release): https://research.meta.ai/blog/muse-wonder
- Artificial Analysis Intelligence Index: https://artificialanalysis.ai/models/muse-wonder-full
- Vals AI professional evaluation: https://www.vals.ai
- 4sapi API gateway (multi-model unified access): https://4sapi.com
International access: [https://4sapi.com](https://4sapi.com)
Domestic access: [https://4sapi.org](https://4sapi.org)
