The result: Muse Spark 1.3 returned 11/36 basic verdicts with accepted proof, ahead of Muse Wonder’s 9/36 and Muse Spark Open’s 1/36. All used High effort. Spark 1.3 and Wonder each reached 4/12 on multi-host intrusions; Spark Open reached 1/12. More expensive did not produce a higher proof count here.

The task: An alarming PowerShell command could be an attack or routine administration. Each agent chooses searches, inspects records and submits a classification with citations. A correct verdict gets the classification right. A verdict with proof also meets the required supporting-record checks.

Basic investigations: Muse trails GPT and Gemini

Basic investigations · 36 cases/model · DFIR 0.6 · One pass
Model Effort Correct verdict Verdict with proof
Muse Spark 1.3 High 80.6%29 / 36 30.6%11 / 36
Muse Wonder High 75.0%27 / 36 25.0%9 / 36
Muse Spark Open High 66.7%24 / 36 2.8%1 / 36
GPT-6.1 Sol High 72.2%26 / 36 69.4%25 / 36
GPT-6 Astra High 77.8%28 / 36 69.4%25 / 36
Gemini 4 Argon High 94.4%34 / 36 80.6%29 / 36
Claude Opus 5.5 High 91.7%33 / 36 33.3%12 / 36
Claude Sonnet 5.5 High 91.7%33 / 36 30.6%11 / 36

Spark 1.3 matched Claude Sonnet 5.5’s 11 accepted-proof verdicts and finished one behind Claude Opus 5.5. GPT-6.1 Sol and GPT-6 Astra each reached 25/36; Gemini 4 Argon reached 29/36. These are observed counts from one pass, all at High effort.

Intrusions: Spark 1.3 and Wonder finish level on proof

Multi-host intrusion investigations · 12 cases/model · DFIR 0.6 · One pass
Model Effort Correct verdict Verdict with proof
Muse Spark 1.3 High 83.3%10 / 12 33.3%4 / 12
Muse Wonder High 91.7%11 / 12 33.3%4 / 12
Muse Spark Open High 83.3%10 / 12 8.3%1 / 12
GPT-6.1 Sol High 83.3%10 / 12 41.7%5 / 12
GPT-6 Astra High 66.7%8 / 12 41.7%5 / 12
Gemini 4 Argon High 83.3%10 / 12 33.3%4 / 12
Claude Opus 5.5 High 100.0%12 / 12 41.7%5 / 12
Claude Sonnet 5.5 High 83.3%10 / 12 33.3%4 / 12

Wonder got 11/12 intrusion classifications right, but only four met the proof checks. Each Muse model completed 47/48 workflows. Its one invalid basic outcome stays in the denominator; completing a workflow does not guarantee a correct or supported answer.

Did the Muse agents collect enough evidence?

This checks whether the agent retrieved the records needed for the verdict, separately from its final citations. Spark 1.3 acquired sufficient basic evidence in 21 cases but returned only 11 verdicts with accepted proof.

Sufficient evidence acquired · Separate from final proof
Model Effort Basic cases Intrusion cases
Muse Spark 1.3 High 21 / 36 4 / 12
Muse Wonder High 12 / 36 7 / 12
Muse Spark Open High 1 / 36 1 / 12

Cost:

Full 48-case workload · USD · Standard rates checked October 16, 2026
Model Effort Measured provider pin Recorded spend Current Standard estimate Workflows completed
Muse Spark 1.3 High meta $0.61–$0.64 $9.61–$10.21 47 / 48
Muse Wonder High meta $22.65 $22.65 47 / 48
Muse Spark Open High modal $18.21 $18.22 47 / 48
GPT-6.1 Sol High openai $4.83 $4.83 48 / 48
GPT-6 Astra High openai/flex $15.51–$26.66 $31.01–$53.31 48 / 48
Gemini 4 Argon High google-ai-studio $21.69–$23.59 $21.69–$23.59 47 / 48
Claude Opus 5.5 High anthropic $54.57–$55.35 $54.57–$55.35 48 / 48
Claude Sonnet 5.5 High anthropic $21.57 $21.57 48 / 48

Spark 1.3’s normal-price estimate is $9.61–$10.21, versus Wonder’s $22.65. Recorded Spark spend was $0.61, with an accounted upper bound of $0.64, on a discounted data-sharing route. The estimate reprices measured input, output and cache tokens at published Standard rates, retaining retries and unresolved-charge bounds.

Three failed Spark 1.3 requests retain $0.0282 in reservations. Spark Open has one zero-charge receipt; its Standard estimate prices those reported tokens. Astra’s recorded route was Flex. Price estimates do not predict another endpoint’s behavior or speed.

Compare cost, time and token use

Switch tiers and metrics below. The default shows these eight High-effort models from a frozen 25-model snapshot.

Measured request time
Mean provider-call time per case · Includes retries; excludes restart pauses
Model Effort Basic cases Intrusion cases
Muse Spark 1.3 High 82.7 s 89.8 s
Muse Wonder High 61.2 s 96.4 s
Muse Spark Open High 73.8 s 104.0 s
GPT-6.1 Sol High 88.1 s 144.7 s
GPT-6 Astra High 98.9 s 152.1 s
Gemini 4 Argon High 380.2 s 584.4 s
Claude Opus 5.5 High 74.9 s 147.0 s
Claude Sonnet 5.5 High 47.0 s 68.3 s

Times reflect the actual pinned routes, including Astra’s Flex route. A Standard price estimate is not a Standard latency measurement.

Token totals and coverage
Known tokens across 48 cases · Input plus output; thinking included once
Model Effort Input tokens Output tokens Total tokens Requests with usage
Muse Spark 1.3 High ≥14,003,064 ≥490,867 ≥14,493,931 848 / 851
Muse Wonder High 11,317,490 383,361 11,700,851 749 / 749
Muse Spark Open High 12,926,078 486,161 13,412,239 713 / 713
GPT-6.1 Sol High 7,882,927 242,893 8,125,820 711 / 711
GPT-6 Astra High Unknown Unknown Unknown Journals unavailable
Gemini 4 Argon High ≥23,144,581 ≥1,301,359 ≥24,445,940 942 / 945
Claude Opus 5.5 High Unknown Unknown Unknown Journals unavailable
Claude Sonnet 5.5 High 9,258,383 305,639 9,564,022 537 / 537

Spark 1.3 reports usage on 848/851 requests; its totals are lower bounds (≥), and its basic token mean remains unavailable. Spark Open covers 713/713 and Wonder 749/749. Spark Open’s optional thinking counts contradict output totals in 31 receipts, so its thinking total remains unknown while valid input/output totals are retained. Gemini has usage on 942/945 requests. Older Astra and Opus journals are unavailable here. Provider tokenizers differ; tokens do not measure internal compute.

What this comparison supports

Spark 1.3 looks more economical than Wonder on this workload at normal rates, with a slightly higher basic proof count. All three Muse models left a large gap between correct classifications and accepted proof.

The limits: One pass per model, 36 basic cases in 12 related families and 12 intrusions in four. Related cases are not independent incidents. These preliminary results need replication and do not establish statistical superiority or performance on unseen investigations. We measured model endpoints in one harness; consumer apps are outside this comparison.

Technical settings and exact API provenance

All eight models used High effort, pinned providers, no fallback and the same case, prompt, tool and grading bindings. High effort does not imply equal internal computation across providers. The shared response limit was 16,384 output tokens and timeout 600 seconds. Independent replay checked every outcome and billing record. Exact Muse IDs: meta/muse-spark-1.3-contributor, meta/muse-spark-open and meta/muse-wonder. Spark 1.3 used Meta’s discounted route requiring paid-model data sharing; Spark Open used Modal and Wonder used Meta. Public names omit the discount suffix. No normal-route rerun was performed.

Data sources and release details

Original GPT and Opus results · Sonnet 5.5 results · GPT-6.1 Sol results · Gemini 4 Argon results · Muse Spark 1.3 results · Muse Spark Open results · Muse Wonder results · Methodology · Main tables

Published Standard rates rechecked October 16, 2026: Spark 1.3, Spark Open, Wonder, OpenAI, Anthropic and Google. Rates matched existing dated quotes; Gemini uses introductory rates through November 30.

Release hashes

investigations-v06.json: 5e38e2573494a3ce7136d2c680c7dfd55d10d55834c6c8d5883e189b91abcb33

investigations-v06-sonnet55.json: ba4bc2aab466f11fa90a06f073f57401f7f5c67bdc1cbc575d84bad84d6a3042

investigations-v06-gpt61sol.json: 91d579a0c603731ad72df3e99a76b070d856bedcacbf2f4c95b9c9519e127edf

investigations-v06-gemini4argon.json: baec2bbe8f3e38a53b60539c8824f11878a296ae49af5a4c6f109e3029f420fd

investigations-v06-muse-spark13.json: 9040633f01020fa1acbff2a05ea04e72936b78ecbb7c7b97c32fb9291c514c4f

investigations-v06-muse-spark-open.json: e870e2b76fff0ac950093827140138d3bb70e8d8c36452e698790e7884f662c6

investigations-v06-muse-wonder.json: eb6ffd9f730782f058011f735615fd2b1ab9de5b88d2bcfba2939ae438bdcb32

Frozen chart snapshot: a48678383c8ff155a3d874e84f93cf4116f2a60f49254f4ec48d104bfc78634f