Meta introduced Muse Wonder on October 12, roughly three months after launching Muse Spark 1.1 and opening its paid Model API. The company positions the Muse Wonder line as capable on personal and professional agent tasks, claims it tops its internal MemoryBench memory benchmark, and says the Full version leads coding evaluations including DeepSWE and SWE-Marathon.

The model card includes Fable 5.1 as a comparison baseline, and Meta's announcement post publishes Fable 5.1 results alongside its own—a significant claim given that Fable 5.1 ranks among Anthropic's strongest offerings. Muse Wonder arrives in two flavors: a standard version accessible through the Model API and Muse Code, Meta's coding agent, plus a higher-capability Muse Wonder Full available only within Muse Code on premium subscription tiers.

Pricing gives Muse Wonder a clear advantage on paper. Meta charges $4 per million input tokens and $20 per million output tokens through its Model API. Anthropic released Fable 5.1 in September at $10 and $50 respectively, available via API and Claude's subscription plans. That puts Muse Wonder 60% cheaper per token. Cached input narrows the gap: Meta bills $0.20 per million cached input tokens versus $0.25 for Fable 5.1. For teams running agents continuously, matching performance at lower cost would represent substantial savings.

To evaluate whether Muse Wonder lives up to Meta's claims, the standard version was tested head-to-head against Fable 5.1 on three demanding agentic and memory tasks.

Test methodology

Muse Wonder ran through OpenRouter, served by Meta, while Fable 5.1 used Anthropic's API. Each model executed once per test at default reasoning settings; Meta's published comparisons employed maximum reasoning effort. A script evaluated every response against answer keys prepared and validated before either model encountered them. Each test ran once, so outcomes could shift on subsequent runs.

  • Agentic bug fix: A 27-module Python job-queue library containing three interacting bugs and one faulty test. Models used tools to read, search, edit, and run tests, capped at 25 turns.
  • Incident investigation: Logs, metrics, deploy history, and config history spanning 20 simulated services during a cascading outage. The root cause is a config change concealed behind two decoy deploys and a server with a clock 263 seconds slow.
  • Long-session memory: A 50-turn scripted chat about a webhook relay service, including 25 project rules with eight later changed and three only implied. Models then wrote config and code, followed by an eight-question recall quiz.

Results by task

Agentic bug fix

Both models identified the first bug in the correct file despite the error message pointing elsewhere. Both fixed the third bug, broke none of 22 regression tests, left the incorrect test untouched, and explained why it was wrong.

The second bug revealed the difference. Fable passed all six hidden tests. Muse passed five; its fix still forwarded the old attempt count to the job's failure event.

On cost, Muse consumed roughly twice as many input tokens, resulting in roughly double the expense despite lower per-token pricing. Both models' costs reflect prompt caching discounts.

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

Incident investigation

Both models scored 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 answer in six turns. Muse required 13 turns and examined more data overall, though it spent less total time waiting for model responses. Despite the lower per-token rate, Muse cost more due to elevated token consumption.

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

Long-session memory

The third test produced the starkest performance gap. Muse scored 25 of 25 on the final deliverable. It applied the latest values for all eight changed rules and captured all three implied ones, including writing code compatible with Python 3.8 because the client's infrastructure runs that version. It also scored 8 of 8 on the quiz and answered "unknown" to all three questions the conversation never addressed.

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

Following Anthropic's guidance, refused output was discarded while preserving the model's own messages in the chat history, so Fable retained access to every rule. An earlier attempt that retained refused turns in the history triggered 45 refusals, prompting a rerun. Anthropic recommends setting up a fallback to another Claude model for refused requests; this was not implemented because testing a different model would change the comparison. Anthropic's own Fable 5.1 benchmarks route cyber-flagged tasks to Claude Opus 4.8.

Since Fable declined the work, this test reveals little about its long-session memory capabilities. It does demonstrate how each model handles everyday backend code that a cyber filter may misclassify as security-related. Muse completed the job and scored 25 of 25.

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

Overall assessment

On agentic work, Muse Wonder performed at Fable 5.1's level, though Meta's own benchmark table shows the standard model trailing Fable 5.1 on certain agent tests, including OSWorld. It matched Fable on incident investigation and came within one hidden test on the bug fix.

Cost savings did not materialize in testing. 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, Fable 5.1 remains the stronger choice. It delivered slightly higher accuracy, required fewer turns, and cost less overall in these tests.

Muse Wonder serves as an alternative when Fable's cyber safeguards interrupt work. However, whether the refusal rate is reasonable compared to other models cannot be confirmed without comparative data across the broader model landscape.