Meta's Muse Wonder Challenges Anthropic's Claude in Agent Benchmarks, But Token Efficiency Lags

On October 12, Meta introduced Muse Wonder, a new addition to its model lineup arriving roughly three months after Muse Spark 1.1 opened the company's paid Model API. According to Meta's claims, the Muse Wonder series excels at personal and professional agent applications, achieves top scores on Meta's internal MemoryBench test, and the Full version leads on coding benchmarks including DeepSWE and SWE-Marathon.

Meta positioned Muse Wonder directly against Anthropic's Claude Fable 5.1, one of the leading models in the market. The company included Fable 5.1 in its model card comparisons and published Fable 5.1 benchmark scores alongside its own, a significant statement given Fable 5.1's standing among Anthropic's top-tier offerings.

The standard Muse Wonder model is accessible via Meta's Model API and Muse Code, the company's coding agent. A more capable variant, Muse Wonder Full, remains exclusive to Muse Code's premium subscription tiers. Pricing for Muse Wonder on the Model API stands at $4 per million input tokens and $20 per million output tokens. Anthropic released Claude Fable 5.1 in September at $10 and $50 respectively through its API and paid plans. This pricing structure gives Muse Wonder a 60% per-token advantage. When accounting for cached input, Meta charges $0.20 per million cached input tokens versus $0.25 for Fable 5.1, potentially delivering substantial savings for teams deploying agents continuously.

Testing methodology

To evaluate whether Muse Wonder matched Meta's performance claims, the standard version was tested head-to-head against Fable 5.1 on three demanding agentic and memory tasks. Muse Wonder ran through OpenRouter using Meta's infrastructure, while Fable 5.1 ran through Anthropic's API. Each model executed once per test using default reasoning settings, though Meta's published comparisons employed maximum reasoning effort. A grading script evaluated all responses against pre-verified answer keys. Each test ran once, meaning results could differ on subsequent runs.

• Agentic bug fix: A 27-module Python job-queue library containing three interacting bugs and one incorrect test. Models used tools to read, search, edit, and run tests with a 25-turn limit.
• Incident investigation: Logs, metrics, deploy history, and config history for 20 simulated services during a cascading outage. The root cause was a config change obscured by two decoy deploys and a server with a 263-second clock skew.
• Long-session memory: A 50-turn scripted conversation about a webhook relay service with 25 project rules, eight subsequently modified and three only implied. At completion, each model generated config and code, then answered eight recall questions.

Results by task

Agentic bug fix

Both models correctly identified the first bug's location despite misleading error messages. Both resolved the third bug, maintained all 22 regression tests, left the incorrect test untouched, and articulated why it was wrong. Performance diverged on the second bug: Fable passed all six hidden tests while Muse passed five, as its fix still transmitted the old attempt count to the job's failure event.

Muse consumed roughly double the input tokens, resulting in approximately double the cost despite its lower per-token rate. Both models' costs reflect prompt caching discounts.

• Muse: 15 turns, 269,615 input tokens, 4,627 output tokens, $0.75, 55.9 seconds
• Fable: 14 turns, 138,213 input tokens, 3,132 output tokens, $0.39, 68.5 seconds

Incident investigation

Both models achieved perfect scores of 10 of 10. Each identified config change CFG-4470 as the root cause, eliminated both decoy deploys, detected the slow clock on the checkout service, and listed all nine affected services in the correct sequence.

Fable reached the solution in six turns. Muse required 13 turns and examined more data overall, though it spent less cumulative time awaiting model responses. Despite its lower pricing, Muse incurred higher costs due to elevated token consumption.

• Muse: 13 turns, 327,587 input tokens, 8,811 output tokens, $1.03, 69.1 seconds
• Fable: 6 turns, 108,128 input tokens, 8,593 output tokens, $0.89, 88.3 seconds

Long-session memory

The third test revealed the most substantial performance gap. Muse scored 25 of 25 on the final deliverable, applying the latest values for all eight modified rules and correctly implementing all three implied requirements, including writing code compatible with Python 3.8 to match the client's server environment. Muse also achieved 8 of 8 on the quiz, answering "unknown" to all three questions the conversation never addressed.

Fable scored zero on this test. On turn 5, Anthropic's API returned a refusal categorized as "cyber." That turn requested reading the relay's target URL from an environment variable. Fable ultimately refused 34 of 50 turns, including the final deliverable and quiz. The project was a standard webhook relay with no connection to cybersecurity. This proved particularly notable because Anthropic indicated that Claude Code users should experience approximately 60% fewer cyber safeguard interventions per session with Fable 5.1 compared to Fable 5. Fable 5 never blocked this testing work for such reasons.

Following Anthropic's guidance, refused outputs were discarded while preserving the model's own messages in the conversation history, ensuring Fable retained access to every rule. An earlier attempt that retained refused turns in the history generated 45 refusals, prompting a rerun. Anthropic recommends implementing a fallback to another Claude model for refused requests; this was not done to avoid testing a different model. In Anthropic's own Fable 5.1 benchmarks, cyber-flagged tasks were routed to Claude Opus 4.8.

Because Fable declined the work, this test provides limited insight into its long-session memory capabilities. It does demonstrate how each model handles routine backend code that a cyber filter might misclassify as security-related. Muse completed the assignment and scored 25 of 25.

• Muse: 50 turns, 93,977 input tokens, 9,848 output tokens, $0.46, 143.2 seconds

Overall assessment

On agentic tasks, Muse Wonder demonstrated performance comparable to Fable 5.1, though Meta's own benchmark table indicates the standard model trails Fable 5.1 on certain agent tests, including OSWorld. It matched Fable on the incident investigation and fell short by one hidden test on the bug fix.

Cost savings did not materialize in these tests. Muse consumed more tokens per task, totaling $2.25 against Fable's $1.85. Fable's total includes $0.57 and 119,187 input tokens from the memory session it refused; on the two tests both models completed, Muse cost $1.78 to Fable's $1.28.

For agentic debugging and investigation work, Fable 5.1 remains the stronger choice. It demonstrated slightly superior accuracy, required fewer turns, and delivered lower total costs in these tests.

Muse Wonder presents a viable alternative when Fable's cyber safeguards interrupt development workflows. However, without comparative data across other models, it remains unclear whether the frequency of such blocking is reasonable relative to competitors.