Growth Protocol outperformed every system tested

Across an 80-question litigation workflow, Growth Protocol reached a 96.0% composite pass rate while using dramatically less estimated inference spend.

96.0%Composite pass rate

+11.6Points ahead of #2

~$0.03Estimated cost / seed

Reasoning quality vs. inference cost

Composite pass rate plotted against estimated inference cost per seed.

Composite pass rate ↑ Estimated inference cost / seed → 100% 85% 70% $0 $1 $2 $3 Growth Protocol · 96.0% · ~$0.03 Single Agent ChatGPT + Context OpenAI + Context ReAct Analyst / Coder / Critic AutoGen MetaGPT SOP · 84.5% · $2.82 Anthropic + Context

Growth Protocol is the only result in the upper-left frontier Mean across three seeds

The result

No quality-cost trade-off

Growth Protocol led the field on reasoning quality while requiring a fraction of the estimated inference spend of the nearest architecture.

+11.6Percentage points ahead of the nearest architecture

~94×Lower estimated inference spend than the runner-up

TraceableAnswers connected to the evidence and rules that produced them

Method at a glance

Same evidence. Same questions. Nine architectures.

The benchmark combined normalized exact-match and open-response questions. Every system received identical evidence, and reported results are the mean across three seeds.

Questions

48 normalized exact-match and 32 open-response questions.

Architectures

Frontier-model, context-augmented, and multi-agent approaches.

Seeds

Reported results are the mean across three benchmark runs.

LLM judges

Open responses used a two-judge scoring process.

Controlled inputs: identical evidence universe across all systems; no customer data used.

Interpretation: comparative reasoning performance and estimated inference efficiency under the benchmark conditions described here.

11.6 points clear of the field

A common 70–100% scale makes the separation visible without exaggerating it. Every architecture is shown.

01Growth Protocol96.0%

02MetaGPT SOP84.5%

03Single Agent83.8%

04Analyst / Coder / Critic82.4%

05ChatGPT + Context80.0%

06ReAct79.7%

07OpenAI + Context79.5%

08AutoGen78.6%

09Anthropic + Context72.5%

Research disclosure. The forthcoming paper will present expanded methodology and analysis. The underlying question set, evidence corpus, prompts, workflow design, scoring implementation, rules, and ontology remain proprietary.

Full research paper forthcoming

The paper will expand on the evaluation framework, methodology, findings, and implications for enterprise reasoning systems.

Enterprise Reasoning Benchmark
October 2026

See what GP can do in your workflow

Bring a consequential enterprise decision. See how Growth Protocol turns evidence, rules, and context into a traceable answer.

Get Started