Xpersona

Public benchmark record

Scores with dates, definitions, and limits.

This page documents the historical score snapshot already used on Xpersona's homepage. It separates the recorded results from the current Claude Fable 5 product so an older comparison is never mistaken for a fresh run.
Run a live request Check service health
6
benchmarks
May 15
Xpersona snapshot
Historical
not Fable 5

Snapshot

The score record currently shown on Xpersona.

Xpersona scores entered the repository on May 15, 2026. The three comparison columns were recorded on May 25, 2026 and remain anonymized as Reference A, B, and C on the public site.

Coding issue resolution

SWE-Bench Pro

Recorded Xpersona

58.3%

Reference A

64.3%

Reference B

54.2%

Reference C

58.6%

SWE-Bench Pro repository

Agentic web research

BrowseComp

Recorded Xpersona

84.1%

Reference A

79.3%

Reference B

85.9%

Reference C

83.2%

BrowseComp overview

Tool workflows

Toolathlon

Recorded Xpersona

55.3%

Reference A

Not reported

Reference B

48.8%

Reference C

50.0%

Toolathlon repository

MCP orchestration

MCP Atlas

Recorded Xpersona

73.0%

Reference A

79.1%

Reference B

78.2%

Reference C

Not reported

MCP Atlas repository

Desktop computer use

OSWorld-Verified

Recorded Xpersona

78.4%

Reference A

78.0%

Reference B

Not reported

Reference C

73.1%

OSWorld repository

Method and limits

What this record proves, and what it does not.

Public evidence is useful only when the boundary around it is just as visible as the number.

Recorded profile

The original repository record identifies Xpersona Frieren Coder with xhigh reasoning, a 400,000-token context window, and a 128,000-token output limit. It predates Claude Fable 5.

Benchmark definitions

Each row links to its benchmark owner or public repository. Those sources define the task family; they do not independently certify Xpersona's recorded score.

Reproducibility limit

Raw prompts, harness configuration, trajectories, and result artifacts are not currently published. Treat these numbers as a historical product record, not an independently reproducible evaluation.
Current product proof

The First-Run Lab is the current verifiable path: it sends a real Claude Fable 5 request and opens its exact usage receipt.

Open Lab