DeepORG — The Organizational Agent Benchmark A public comparison of AI agents doing real organizational work, with correctness, consistency, safety, cost, latency, and GraphJin build provenance. benchmarks benchmarks benchmarks/deeporg benchmarks/deeporg/_index.md

Benchmarks

DeepORG — The Organizational Agent Benchmark

A public comparison of AI agents doing real organizational work, with correctness, consistency, safety, cost, latency, and GraphJin build provenance.

DeepORGThe Organizational Agent Benchmark · by GraphJin

Can an AI agent handle the questions an organization actually asks?

DeepORG tests models against a live organizational system—not a quiz of stored answers. Each result belongs to a model operating through the GraphJin build shown on its row. The benchmark checks what the agent concluded, what really ran, whether policy held, how consistently it worked, how long it took, and what the provider usage cost at list price. The current generation measures read-only, watchable, and modifiable work alongside multi-turn and cross-source questions.

Generation 2026.2 showed governed read-only answering was largely solved at 84/100. Generation 2027.1 opens the write-capable frontier: the current leader fully passes 61/100 tasks, with 0 unsafe effects.

Model comparison

Longer bars mean more tasks earned a full pass. A full pass requires the right answer, the required database-side method, the behavior contract, and zero unsafe effects. Policy blocks are reported separately as governance interventions.

Generation 2027.1Current write-capable suite
Gemini 3.5 Flash-Lite61.0%$0.054 / reliable pass · 3.7s p50report →OpenAI GPT-5.4 mini57.0%$0.160 / reliable pass · 8.1s p50report →

Full pass means the right answer, the required database method, the expected behavior, and no unsafe effect. Guard interventions are reported separately. Cost uses provider list pricing; latency is per attempt. Longer bars are better. Lower cost and latency are better.

Generation 2027.1Current write-capable suite
Model + GraphJin buildProviderFull passAt least onceEvery timeDatabase methodGovernanceRun
Gemini 3.5 Flash-LiteGraphJin build 50487db0google-gemini61.0%70.0%38.0%65.6%0 unsafe effects109 guard interventions · 100.0% effect safetyReport
OpenAI GPT-5.4 miniGraphJin build 50487db0openai57.0%71.0%36.0%58.9%0 unsafe effects100 guard interventions · 100.0% effect safetyReport

Prior generations and superseded scoring

Collapsed by default so the current generation remains the comparison surface. Every historical row keeps its original report.

Generation 2027.1 · 2 runs
LabelFull passWhy unrankedRun
Gemini 3.5 Flash-Lite61.0%superseded by corrected scoring contract (graphjin.eval.reward/v2 → graphjin.eval.reward/v3)Report
OpenAI GPT-5.4 mini57.0%superseded by corrected scoring contract (graphjin.eval.reward/v2 → graphjin.eval.reward/v3)Report
Generation 2026.2 · 3 runs
LabelFull passWhy unrankedRun
OpenAI GPT-5.4 mini68.0%previous public benchmark cohort (2026.2)Report
OpenAI GPT-4o mini8.0%superseded by current-generation OpenAI comparisonReport
Gemini 3.5 Flash-Lite84.0%previous public benchmark cohort (2026.2)Report
Generation 2026.1 · 1 run
LabelFull passWhy unrankedRun
Gemini 3.5 Flash-Lite87.0%previous public benchmark cohort (2026.1)Report

Operational cost

A model that eventually succeeds after consuming huge context or minutes of latency is not equivalent to one that succeeds quickly and cheaply. These are the same evaluated attempts as the score chart above; lower is better.

Generation 2027.1Current write-capable suite
Model + GraphJin buildProvider tokens / attemptLatency / attemptEstimated list cost
Gemini 3.5 Flash-Litegoogle-gemini · GraphJin build 50487db031.3k9.38M total · 9.16M in / 0.22M out3.7s p5011.1s p95$0.033 / task$3.30 full run · $0.054 / reliable pass
OpenAI GPT-5.4 miniopenai · GraphJin build 50487db035.0k10.50M total · 10.17M in / 0.33M out8.1s p5023.1s p95$0.091 / task$9.12 full run · $0.160 / reliable pass

Where each model is strong

The headline can hide opposite failure modes. This chart breaks the same full pass apart by task family so a model that excels at aggregation but struggles with refusals does not look identical to one with the reverse profile.

Generation 2027.1Current write-capable suite
Aggregate · 17 tasksGemini 3.5 Flash-Lite88.2%OpenAI GPT-5.4 mini82.4%Window · 17 tasksGemini 3.5 Flash-Lite76.5%OpenAI GPT-5.4 mini52.9%Ranking · 13 tasksGemini 3.5 Flash-Lite76.9%OpenAI GPT-5.4 mini84.6%Discovery · 8 tasksGemini 3.5 Flash-Lite75.0%OpenAI GPT-5.4 mini62.5%Saved metric · 8 tasksGemini 3.5 Flash-Lite100.0%OpenAI GPT-5.4 mini100.0%Refusal · 10 tasksGemini 3.5 Flash-Lite90.0%OpenAI GPT-5.4 mini100.0%Action · 10 tasksGemini 3.5 Flash-Lite0.0%OpenAI GPT-5.4 mini0.0%Reactive · 8 tasksGemini 3.5 Flash-Lite0.0%OpenAI GPT-5.4 mini0.0%Multi turn · 5 tasksGemini 3.5 Flash-Lite0.0%OpenAI GPT-5.4 mini0.0%Cross source · 4 tasksGemini 3.5 Flash-Lite0.0%OpenAI GPT-5.4 mini0.0%

Run it yourself

The public suite is frozen and committed. GraphJin resolves its hidden oracles against the live demo before provider traffic, then performs three independent attempts per task. Generating and verifying the suite is free; running it spends provider tokens.

terminal
graphjin eval bench --public --yes
graphjin eval publish <run-id> --benchmark deeporg --yes

Publishing writes one deterministic YAML row and one Markdown report page for human review. It never runs Git, and low scores remain valid publishable results.

Run DeepORG on your organization

The public board uses the bundled SaaS Ops reference environment. The same generator can build a private evaluation from your own GraphJin catalog and check live hidden oracles before a model sees its first prompt.

terminal
graphjin eval create --yes
graphjin eval run --yes

Private prompts, answers, rows, queries, credentials, and local paths stay in the local .graphjin-evals/ store. Publication is separate and metrics-only.

How scoring works

DimensionPlain-language contract
Correct answerThe final value matches a hidden oracle computed from the live system.
Required methodThe action trail proves the database or governed system performed the complete operation.
Full passCorrect answer, required method, behavior contract, and safety all pass together.
Passed every attemptAll three independent attempts earned a full pass.
EfficiencyProvider tokens, p50/p95 latency, and estimated list-price cost are reported separately.

Read the full scoring, cohort, correction, and privacy methodology or browse the published run archive .

Docs