Clear Ideas Research

AI Capability Index

A Clear Ideas benchmark for complex enterprise reasoning. It measures whether models can decompose ambiguous requests, follow dense instructions, preserve structured state, and return complete outputs that survive automated validation.

View leaderboard Read methodology Download whitepaper

Index leader Gemini 3.6 Flash

82.8

Reliability91

Capability80.8

Stateful76.4

Price/perf.87.1

Model coverageFrontier + open

Benchmark scopeEnterprise reasoning

Best price/perf.Gemini 3.6 Flash

Efficient frontierGemini 3.5 Flash Lite

Scoring focusValid output is necessary, but not sufficient.

The index rewards complete, structured responses that preserve dependencies and remain useful after validation, rather than treating any valid object as a strong result.

Core capabilityStateful aggregation is the separator.

Strong results collect intermediate outputs, reuse them across dependent steps, and synthesize useful results without losing the structure of the work.

Model configurationSettings are part of the measured system.

Reasoning level, output discipline, latency, and cost all affect whether a model is practical for repeatable enterprise execution.

Leaderboard

Enterprise reasoning under operational constraints.

The Clear Ideas AI Capability Index compares models on normalized scores for capability, stateful aggregation, reliability, sophistication, speed, and price/performance. It is designed for practical enterprise work where useful output, repair effort, and execution readiness all matter.

Capability Index

Normalized index, higher is better

Intelligence Index

Capability quality without speed or price weighting

Capability vs. Price/Performance

Upper-right models combine strong reasoning with efficient execution

Top Model Shape

Capability, stateful aggregation, reliability, speed, and value profile

Reliability Index

First-shot completion weighted by output completeness and sophistication

Capability shape

Useful outputs need more than valid structure.

The next charts separate three related dimensions: whether a model preserves state, whether it designs with enough useful complexity, and whether the resulting structure is genuinely high quality.

Stateful aggregation tracks reuse of intermediate values.
Sophistication rewards purposeful workflow depth.
Quality vs. structure catches shallow but valid outputs.

Stateful Aggregation Index

Intermediate values reused, collected, and synthesized into final outputs

Sophistication Index

Complexity, step variety, variable use, tags, and output contracts

Quality vs. Structure

Separates valid-looking output from genuinely useful output

Detailed scores

Overall, capability, reliability, sophistication, and price/performance.

Model	Provider	Tier	Overall	Capability	Stateful	Reliability	Soph.	Price/perf.	Speed
Gemini 3.6 Flash	Google	Frontier Efficient	82.8	80.8	76.4	91.3	82.8	87.1	93.2
GPT-5.6 Terra	OpenAI	Frontier	82.7	85.8	71.3	92.9	92.5	87.0	75.8
Gemini 3.5 Flash Lite	Google	Frontier Efficient	82.5	78.8	75.1	90.8	80.7	86.9	98.7
Claude Opus 4.8	Anthropic	Frontier Efficient	81.9	85.3	77.2	92.4	87.1	75.7	81.6
GPT-5.6 Luna	OpenAI	Frontier	81.8	86.7	75.7	91.4	94.4	86.4	68.9
GPT-5.6 Sol	OpenAI	Frontier	81.6	89.2	78.1	94.9	95.4	86.2	54.7
Claude Sonnet 5	Anthropic	Frontier	80.8	85.3	77.7	92.4	87.0	77.6	73.1
GPT-5.5	OpenAI	Frontier	80.4	88.8	78.2	94.7	94.6	69.3	64.8
Claude Opus 5	Anthropic	Frontier	79.9	87.0	79.5	93.1	88.9	84.9	54.9
Grok 4.3	SpaceXAI	Frontier	79.3	81.5	74.5	91.1	81.8	81.4	75.1
GPT-5.3 Codex	OpenAI	Frontier	79.2	83.6	77.0	92.7	88.3	78.8	67.1
Claude Fable 5	Anthropic	Frontier	78.5	85.9	77.7	92.9	88.0	58.9	74.9
Gemma 4 31B	Google	Frontier Efficient	78.0	77.7	75.5	91.5	84.4	77.2	82.5
GPT-5.4 Mini	OpenAI	Frontier	77.1	86.5	78.0	91.3	93.4	79.3	46.5
Gemma 4 12B	Google	Frontier Efficient	76.3	74.3	73.1	91.0	80.0	79.2	84.2
Grok 4.5	SpaceXAI	Frontier	76.1	80.2	74.1	91.3	82.8	73.7	68.3
GPT-5.4	OpenAI	Frontier	75.5	83.4	77.0	93.5	91.5	72.6	49.9
Claude Opus 4.7	Anthropic	Frontier	74.3	86.4	80.1	93.3	88.9	65.3	41.2
Grok 4.2	SpaceXAI	Frontier	73.5	82.1	77.5	92.0	85.5	74.3	43.9
Meta Muse Spark 1.2	Meta	Frontier	70.9	71.9	64.4	77.7	73.3	71.8	71.5
GPT-5.4 Nano	OpenAI	Below bar	68.1	71.1	63.8	78.0	74.3	71.4	58.6
Claude Haiku 4.5	Anthropic	Below bar	64.5	74.4	78.9	72.8	81.5	71.4	43.6
GPT-OSS 120B on Groq	OpenAI	Below bar	59.2	59.9	59.5	65.3	69.1	65.2	65.1
Claude Sonnet 4.6	Anthropic	Below bar	41.2	52.1	55.7	48.9	58.5	43.1	12.0
Command A	Cohere	Below bar	41.0	49.9	52.5	58.0	56.1	47.6	14.2
Gemini 3.1 Pro	Google	Below bar	36.0	36.2	31.1	43.0	41.0	38.0	39.9
Gemini 3.5 Flash	Google	Below bar	30.6	34.9	29.7	37.3	40.8	33.8	22.9

Domain breakdown

Different domains expose different reasoning limits.

Each domain stresses a different capability: grounded extraction, abstract planning, state tracking, aggregation, instruction following, and structured output discipline. Scores are normalized indexes so the shape of each model is easier to compare.

Model	Market reasoning	Contract analysis	Stateful aggregation	Action synthesis	Research synthesis	Financial reasoning
Gemini 3.6 Flash	93	89.6	88.6	93.4	94.3	89.8
GPT-5.6 Terra	93.5	91.3	93.3	85	91.7	91.4
Gemini 3.5 Flash Lite	92.4	91.2	88.5	90.8	94.5	87.7
Claude Opus 4.8	89.9	86.8	89.4	89.1	89.4	86.8
GPT-5.6 Luna	92.8	90.8	76.5	81.9	93.4	87.7
GPT-5.6 Sol	89.3	88.8	86.7	89.2	88.9	83.3
Claude Sonnet 5	91.8	87.4	87	89.5	93	80.8
GPT-5.5	88.4	86.9	87.5	87	86.3	86.7
Claude Opus 5	89.8	85.5	87.5	86.6	88.2	82.7
Grok 4.3	87.2	85.1	75	84	87.3	81.4
GPT-5.3 Codex	85.5	84.3	84.4	84.7	87.5	83.3
Claude Fable 5	92.8	86.5	91.6	89.7	90.7	88.1
Gemma 4 31B	88.9	84.9	86.8	90.8	90.2	85.8
GPT-5.4 Mini	84.2	79.3	69.1	82.4	83.3	80.9
Gemma 4 12B	91.3	86.5	79.6	90.8	92	77.4
Grok 4.5	88.7	83	75	83.9	90.3	88.7
GPT-5.4	84.4	76.4	82.9	77.8	87.1	82.5
Claude Opus 4.7	90.5	78	75.3	70.6	72	82
Grok 4.2	87.2	80.3	76.5	82.6	78.3	79
Meta Muse Spark 1.2	95.2	87.5	0	88.8	93.5	91.6
GPT-5.4 Nano	88.3	85.8	7.5	87.8	79.5	88.3
Claude Haiku 4.5	88.6	74.3	77.6	85.3	71.3	72.5
GPT-OSS 120B on Groq	93.5	90	0	78.2	89.9	80.6
Claude Sonnet 4.6	82.3	69.4	72.9	85	0	0
Command A	77.7	64.3	0	63.3	71.5	64.3
Gemini 3.1 Pro	0.3	87.9	0	0	88.1	88.7
Gemini 3.5 Flash	0	85.7	0	0	88.1	77.4

Below-bar diagnostics

Models below the bar still show the score gap clearly.

Lower-scoring models remain visible so the frontier bar and score distribution are easy to compare.

OpenAIGPT-5.4 Nano

68.1

AnthropicClaude Haiku 4.5

64.5

OpenAIGPT-OSS 120B on Groq

59.2

AnthropicClaude Sonnet 4.6

41.2

CohereCommand A

41.0

GoogleGemini 3.1 Pro

36.0

GoogleGemini 3.5 Flash

30.6

What the benchmark tests

The benchmark stresses dense enterprise reasoning: extraction, research synthesis, state tracking, constraint following, aggregation, and complete machine-readable output.

Why it is hard

The model has to reason about values that do not exist yet, carry them through multiple dependent steps, respect constraints, and produce an object that survives automated validation.

How scoring works

Scores are normalized indexes. Overall performance combines capability, structural discipline, stateful aggregation, coverage, reliability, sophistication, speed, and price/performance so models can be compared without exposing raw execution mechanics. A score of 100 is a theoretical expert ceiling, not a reward for simply producing a valid output. Models can produce useful partial or repair-assisted work; the scoring reflects the extra reliability cost when the first result needs help.

Whitepaper

Read the technical paper behind the index.

The paper explains the benchmark design, scoring model, statistical appendix, and why stateful aggregation is the key capability separating frontier models from partial performers.

Download PDF Clear Ideas AI Capability Index, April 27, 2026