GraphWalks: what it measures

One row per model and compatible score measure. Mean = equal-weight arithmetic average of imported reports. Distinct configurations reported for the same model in a source table remain separate. Reports can overlap; the range is not a confidence interval. Unbundle to inspect each observation.
		Evidence	What the result shows
Gemini 4 Argon →Google	91.95% 2 reports · range 84.2–99.7%	GraphWalks[1] View 2 reports 99.7% · GraphWalks BFS ≤128k · F1 [1] 84.2% · GraphWalks BFS 256k–1M · F1 [1]	Published observations span 15.5 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. Leads this displayed comparison.
GPT-6 Astra →OpenAI	85.25% 2 reports · range 71.8–98.7%	GraphWalks[1] View 2 reports 98.7% · GraphWalks BFS ≤128k · F1 [1] 71.8% · GraphWalks BFS 256k–1M · F1 [1]	Published observations span 26.9 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #2 of 4 in this displayed comparison; 6.7 pp below the leader.
Claude Opus 5.5 →Anthropic	78.7% 2 reports · range 66.8–90.6%	GraphWalks[1] View 2 reports 90.6% · GraphWalks BFS ≤128k · F1 [1] 66.8% · GraphWalks BFS 256k–1M · F1 [1]	Published observations span 23.8 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #3 of 4 in this displayed comparison; 13.25 pp below the leader.
Claude Fable 5.1 →Anthropic	78.2% 2 reports · range 65–91.4%	GraphWalks[1] View 2 reports 91.4% · GraphWalks BFS ≤128k · F1 [1] 65.0% · GraphWalks BFS 256k–1M · F1 [1]	Published observations span 26.4 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty. #4 of 4 in this displayed comparison; 13.75 pp below the leader.

One row per model and compatible score measure. Mean = equal-weight arithmetic average of imported reports. Distinct configurations reported for the same model in a source table remain separate. Reports can overlap; the range is not a confidence interval. Unbundle to inspect each observation.

Evidence

What the result shows

Gemini 4 Argon →Google

91.95%

2 reports · range 84.2–99.7%

GraphWalks[1]

View 2 reports

99.7% · GraphWalks BFS ≤128k · F1 [1]
84.2% · GraphWalks BFS 256k–1M · F1 [1]

Published observations span 15.5 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.

Leads this displayed comparison.

GPT-6 Astra →OpenAI

85.25%

2 reports · range 71.8–98.7%

GraphWalks[1]

View 2 reports

98.7% · GraphWalks BFS ≤128k · F1 [1]
71.8% · GraphWalks BFS 256k–1M · F1 [1]

Published observations span 26.9 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.

#2 of 4 in this displayed comparison; 6.7 pp below the leader.

Claude Opus 5.5 →Anthropic

78.7%

2 reports · range 66.8–90.6%

GraphWalks[1]

View 2 reports

90.6% · GraphWalks BFS ≤128k · F1 [1]
66.8% · GraphWalks BFS 256k–1M · F1 [1]

Published observations span 23.8 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.

#3 of 4 in this displayed comparison; 13.25 pp below the leader.

Claude Fable 5.1 →Anthropic

78.2%

2 reports · range 65–91.4%

GraphWalks[1]

View 2 reports

91.4% · GraphWalks BFS ≤128k · F1 [1]
65.0% · GraphWalks BFS 256k–1M · F1 [1]

Published observations span 26.4 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.

#4 of 4 in this displayed comparison; 13.75 pp below the leader.

Sources

Gemini 4 Argon Model Card ↗

Google DeepMind · 2026

Dates & evaluation details

GraphWalks BFS ≤128k · F1 · 2026-10-26

A transcription of the October model card, including its external leaderboard snapshots. These are not uniformly matched runs. Effort, tools, hardware, frame budgets and task subsets differ. GraphWalks reports F1; OSWorld uses the offline subset and partial credit. Harvey refers to the archived Vals v1 protocol, not the remediated v1.1 or AA variant. Consult the linked methodology.

GraphWalks BFS 256k–1M · F1 · 2026-10-26

TRACE THE SCORE · Per-model notes use published breakdowns where available.

Reported in recent model launches

These model reports include this benchmark. Their harnesses, task subsets, scoring policies, and resource budgets can differ; follow each report’s evaluation notes before joining the numbers.

Google · Gemini 4 Argon →

What it measures

A long input describes a graph, and the model must retrieve related nodes or follow graph structure to the requested depth.

How to read the result

F1 balances missing correct nodes against adding incorrect ones. It is not a binary percentage of fully solved questions.

Limits and comparison traps

BFS and parent retrieval, short and very long contexts, and dataset corrections must be distinguished. Tool access would change the construct being tested.

Benchmark reference notes

Multi-hop graph traversal and relational reasoning benchmark over large synthetic and real graphs (reported in Gemini 4 Argon).

Epoch data and evaluation setting

Epoch AI publishes this benchmark in its capability collection. Imported rows retain the published score column and model configuration. Original authors and evaluation providers remain credited in the result references.

Epoch attribution and score coverage

Data is attributed to Epoch AI under CC BY 4.0, with original evaluation providers credited in the linked source snapshots. Scores are imported only where a published mapping identifies the field and its scale; missing or incompatible records are omitted rather than assigned zero.

Where this fits in an evaluation suite

A large context window is a capacity specification. Successful retrieval and cross-document reasoning are observed capabilities. Test both at the actual lengths your workflow needs.

Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.

A reproducible comparison

Pin the benchmark release, task split, model configuration, tools, sampling budget, and evaluation harness described by the primary source.

Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.

PRIMARY EVIDENCE & FURTHER READING

Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.

Sources

Reported in recent model launches

What it measures

How to read the result

Limits and comparison traps

Benchmark reference notes

Epoch data and evaluation setting

Epoch attribution and score coverage

Where this fits in an evaluation suite

A reproducible comparison

Continue exploring

GDP.pdf

AA-LCR v1.1

MLCR-AA

RULER