Models & transparency
Real models. Their real biases.
Every AI model has a grain it cuts against — a house style, a blind spot, a way of being confidently wrong. Decidi runs your decision across several of them at once, so no single model’s bias decides your outcome. Here’s what each brings, openly.
GPT
OpenAI GPT · OpenAI
A versatile, dependable generalist — the steady centre of the table that other members can push against.
Strengths
- Broad, balanced general reasoning across an unusually wide range of topics
- Strong, reliable instruction-following and structured output (JSON, tables, schemas)
- Mature tool use and function calling, which keeps multi-step tasks on the rails
- Fast, fluent drafting that reads cleanly with little editing
- Deep, well-rounded coding ability across many languages and frameworks
Known biases
- Can be agreeable to a fault — it will often validate a premise rather than contest it
- Tends toward a polished, middle-of-the-road consensus answer that can smooth over real trade-offs
- A recognisable house tone (measured, balanced, lightly hedged) that flattens distinct viewpoints
- Knowledge has a training cutoff, so it can speak confidently about events past that date
- Will sometimes present a plausible-sounding answer with more certainty than the evidence warrants
General-purpose reasoning and synthesisDrafting and structuring documentsTool-driven, multi-step task executionSoftware design and code generation
Full GPT profile Claude
Anthropic Claude · Anthropic
The thorough, self-aware analyst — strongest on depth, context and naming what it is unsure about.
Strengths
- Careful, structured long-form reasoning that holds a complex argument together
- Very large context window, so it can hold whole documents, codebases or datasets at once
- Strong, maintainable coding with clear explanations of the reasoning behind it
- Nuanced handling of ambiguity, ethics and trade-offs without collapsing to a single take
- Tends to surface its own uncertainty and caveats rather than papering over them
Known biases
- Can be over-cautious — adding caveats or declining edge cases that a user genuinely needs answered
- Tendency to verbosity; it explains thoroughly even when a crisp answer would serve better
- A deliberate, careful house style that can read as hedging when a decisive call is wanted
- Knowledge has a training cutoff and no inherent live-web view of recent events
- Its emphasis on balance can under-weight a bold-but-correct minority position
Long-document and large-codebase analysisNuanced reasoning over ethics, risk and trade-offsCareful technical writing and reviewStructured argument and synthesis
Full Claude profile Gemini
Google Gemini · Google DeepMind
The multimodal, data-grounded specialist — strongest where breadth of input and current facts matter.
Strengths
- Genuinely large context and strong recall across very long inputs
- Native multimodal reasoning over text, images, audio and video together
- Strong quantitative, mathematical and structured-data reasoning
- Tight integration with current information through Google search grounding
- Flash tier is fast and cost-efficient for high-volume sub-tasks
Known biases
- Output quality can swing between the Pro and Flash tiers, so consistency varies by configuration
- Can be terse or under-explain its reasoning, leaving the "why" implicit
- Search grounding helps recency but can introduce or over-trust a weak source
- A factual, encyclopaedic register that can read as flat for persuasive or human-toned work
- Occasionally over-confident on quantitative claims that deserve a second check
Long-context and multimodal analysisQuantitative and data-heavy reasoningTasks needing current, grounded informationHigh-volume sub-tasks on the fast tier
Full Gemini profile Grok
xAI Grok · xAI
The current-and-candid contrarian — earns its seat by knowing what is happening now and saying so plainly.
Strengths
- Real-time awareness of current events through live access to public web and social data
- Willing to engage directly with contested or sensitive questions rather than deflect
- Competitive reasoning and coding in its most recent releases
- A more candid, less hedged voice that states a position plainly
- Useful for surfacing the live public conversation around a topic
Known biases
- Live-data exposure can pull in unverified or low-quality sources that need filtering
- Its more opinionated, candid voice can read as confident even when the basis is thin
- Less established track record than longer-lived families, so behaviour is less predictable
- Tone can skew informal or provocative for formal decision contexts
- Real-time signals reflect whatever is loud now, which is not always what is true
Current-events and live-sentiment contextEngaging directly with contested questionsSurfacing the present public conversationA blunt, decisive counter-voice in debate
Full Grok profile Why this matters
When you ask one model, you inherit its biases silently. When a council of independent models debates — each assigned a different live model, round-robin across providers — they cross-check each other’s facts and cancel each other’s slant. You see where they disagreed, and a moderator turns it into one decisive verdict.