Structured AI Reasoning, Engineered

Why UnisonCore exists

UnisonCore is a multi-agent, multi-model orchestration engine for those who expect AI to route with precision, decompose tasks safely, and leave an explainable trace. Modern AI demands structure, predictable routing, and governed execution. UnisonCore provides the backbone that makes multi-agent reasoning reliable, transparent, and safe, by design use of benchmarked models. These models are a basis of decisional infrastrcture, that allows for a predictable and auditable AI reasoning process.

How UnisonCore thinks

Model Benchmark Leaderboard

Independent results for models we evaluate for AI orchestration. Model selection was started by most use of major direct API vendors such as OpenAI, Anthropic, xAI, Google Gemini, Mistral, DeepSeek. Together.ai was included for exploratory models. All models are evaluated for capability, safety, and cost efficiency.

Future Model Selections

The future of AI orchestration is multi-agent, multi-model. UnisonCore will continue to evaluate new models and add them to the leaderboard. Competitional marketplace for model use should be the standard for today's infrastructural design. We welcome feedback on models to include, and invite model vendors to submit their models for evaluation.

Top models in this size class

Metrics Last Update:

Fast models: one curated pick per major direct API vendor (OpenAI, Anthropic, xAI, Google Gemini, Mistral, DeepSeek).

Model Vendor Size Accuracy Reasoning Coding Latency Cost Compile Safety
mistral-small-latest Mistral Fast
52.5% ▲
25% ●
60% ▲
80.1% ●
50% ●
Below compile bar Blocked (P4) P4 (jailbreak resistance): jailbreak resist rate=83.3% bar=85%
grok-3-mini xAI Fast
52.1% ▲
28.3% ▲
60% ▲
49.2% ●
50% ●
Below compile bar Approved
gpt-4.1-mini OpenAI Fast
48.3% ▲
8.3% ●
60% ▲
74.7% ●
50% ●
Meets compile bar Conditional
gemini-2.5-flash Google Fast
48.3% ▲
0% ●
76.7% ▲
8.6% ●
50% ●
Below compile bar Blocked (P5) P5 (harmful output): harmful output rate=8.3% bar=5%
claude-haiku-4-5 Anthropic Fast
44.2% ▲
16.7% ●
60% ▲
56.8% ●
50% ●
Meets compile bar Approved
deepseek-v4-flash DeepSeek Fast
40% ▲
0% ●
60% ▲
0% ●
50% ●
Below compile bar Approved

Models marked Blocked failed our safety publication gates — the badge lists failed gate IDs (e.g. P4, P5).

Provider error means the safety run could not reach the model API (not a behavioral score).

Scores remain visible for transparency — open a model row on All metrics for measured gate detail, or read Methodology for gate definitions and accepted blocked dispositions.

Checklist-based lab benchmarks from the UnisonCore evaluation pipeline — not human ratings or live product reviews. Reference board rows passed full capability and safety testing; Frontier latency and cost are measured on Together.ai endpoints.